Large model-based automated text generation and content creation method and system

Abstract

The application provides a large model-based automatic text generation and content creation method and system, wherein the method comprises: obtaining material text data in a multi-turn dialogue scene, creation theme information, creation history state and current environment data; jointly processing the material text data and the creation theme information, and constructing text structure information according to a joint result; determining a core description element based on the text structure information; adopting a memory-enhanced attention mechanism to fuse the environment data, the creation history state and the core description element, and generating a hidden state sequence; and generating an automatic text conforming to the current dialogue scene based on the hidden state sequence. The application improves the context coherence and theme fitting degree of an intelligent dialogue system in a complex environment.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of natural language processing technology in the field of artificial intelligence, in particular to a large model-based automatic text generation and content creation method and system. BACKGROUND

[0002] In multi-turn dialogue scenarios such as smart home and smart customer service, automatic text generation and content creation need to consider context coherence, theme fit and environmental adaptability. Existing technologies usually rely on pre-trained language models to generate text, but when faced with complex interaction scenarios, the system needs to dynamically integrate environmental data, dialogue history and creation theme and other multi-dimensional information to ensure that the generated content meets the user's intention and adapts to real-time environmental changes.

[0003] At present, a typical scheme directly integrates dialogue text and environmental sensor data using an attention mechanism, and generates response text through weighted calculation. This scheme converts environmental data into a feature vector and simply concatenates it with dialogue text features, uses a standard attention layer to assign weights, and finally outputs the generation result.

[0004] However, this scheme is prone to theme drift in long-term dialogue, and environmental data is only used as an auxiliary feature in the calculation, failing to form a deep synergy with the dialogue history, resulting in a lack of logical consistency in the generated content in complex scenarios, and difficulty in maintaining the long-term stability of the creation theme. SUMMARY

[0005] The present application provides a large model-based automatic text generation and content creation method and system to solve the problem of poor context coherence and low theme fit of intelligent dialogue systems in complex environments in the prior art.

[0006] In a first aspect, the present application provides a large model-based automatic text generation and content creation method, comprising:

[0007] Obtaining material text data, creation theme information, creation history state and current environmental data in a multi-turn dialogue scenario;

[0008] Jointly processing the material text data and the creation theme information, and constructing text structure information according to the joint result;

[0009] Determining a core description element based on the text structure information;

[0010] Using a memory-enhanced attention mechanism to integrate the environmental data, the creation history state and the core description element, and generating a hidden state sequence;

[0011] Generating an automatic text that meets the current dialogue scenario based on the hidden state sequence.

[0012] Optionally, the memory-enhanced attention mechanism is adopted to fuse the environment data, the creation history state and the core description element to generate a hidden state sequence, including:

[0013] The environment data is converted into an environment feature vector;

[0014] The creation history state is subjected to memory-enhanced processing;

[0015] The environment feature vector, the memory-enhanced creation history state and the core description element are input into a memory-enhanced unit of the memory-enhanced attention mechanism to output a hidden state sequence through multiple rounds of state iteration.

[0016] Optionally, the memory-enhanced attention mechanism is adopted to fuse the environment data, the creation history state and the core description element to generate a hidden state sequence, including:

[0017] The core description element is taken as a query reference;

[0018] The matching degrees between the query reference and each history state vector in the memory-enhanced creation history state are calculated, and a preset number of target history states with the highest matching degrees are filtered out from the memory-enhanced creation history state;

[0019] The target history states and the environment feature vector are merged, and the merged result is subjected to three consecutive calculations to generate three hidden states;

[0020] The three hidden states are combined according to the generation order of the three hidden states to form a hidden state sequence.

[0021] Optionally, the core description element is determined based on the text structure information, including:

[0022] Based on the text structure information, an association metric value between each pair of entity objects is calculated;

[0023] According to the association metric value, the plurality of entity objects are sorted;

[0024] The first N entity objects are selected from the sorting result as the core description element, where N is a preset positive integer.

[0025] Optionally, the core description element is determined based on the text structure information, including:

[0026] Based on the text structure information, a theme weight value of each entity object is obtained;

[0027] Calculate the interaction frequency and relationship distance between the entity objects in the text structure information.

[0028] According to the theme weight value, adjust the interaction frequency between the entity objects;

[0029] Weighted calculation is performed on the adjusted interaction frequency and the relationship distance to generate a correlation measure value.

[0030] Optionally, the jointing of the material text data and the creation theme information includes:

[0031] Identify multiple entity objects and connection relationships between entity objects from the material text data;

[0032] Parse the creation theme information to obtain theme keywords;

[0033] Establish the association relationship between the theme keywords and the entity objects;

[0034] According to the connection relationship between the entity objects and the association relationship, form text structure information.

[0035] Optionally, the generation of the automated text in accordance with the current dialogue scenario based on the hidden state sequence includes:

[0036] Input the hidden state sequence into a text decoder of a pre-trained large model;

[0037] According to the creation theme information, set a constraint condition that prohibits the generation of text units conflicting with the creation theme;

[0038] According to the constraint condition and the time sequence of the hidden state sequence, generate a text unit corresponding to the current hidden state;

[0039] Environmentally adapt and adjust the generated text unit;

[0040] Combine the adaptively adjusted text units in accordance with a preset dialogue logic sequence to output an automated text.

[0041] In a second aspect, the present application provides a large model-based automated text generation and content creation system, comprising:

[0042] An acquisition module is configured to acquire material text data, creation theme information, creation history state, and current environment data in a multi-round dialogue scenario;

[0043] A construction module is configured to joint the material text data and the creation theme information, and construct text structure information according to the jointing result;

[0044] determining a core description element based on the text structure information;

[0045] a first generating module configured to fuse the environment data, the creative history state and the core description element by using a memory-enhanced attention mechanism to generate a hidden state sequence;

[0046] a second generating module configured to generate an automated text conforming to a current dialogue scene based on the hidden state sequence.

[0047] In a third aspect, the present application provides a computing device comprising a processor and a memory, wherein the memory stores a computer program, and the processor is configured to execute the computer program to implement the method for generating an automated text and creating content based on a large model according to any one of the first aspect.

[0048] In a fourth aspect, the present application provides a computer storage medium having computer program instructions stored thereon, wherein the computer program instructions are executed by a processor to implement the method for generating an automated text and creating content based on a large model according to any one of the first aspect.

[0049] In the present application, a method for generating an automated text and creating content based on a large model is provided, which comprises: obtaining material text data, creative theme information, creative history state and current environment data in a multi-turn dialogue scene; jointly processing the material text data and the creative theme information, and constructing text structure information according to a joint result; determining a core description element based on the text structure information; fusing the environment data, the creative history state and the core description element by using a memory-enhanced attention mechanism to generate a hidden state sequence; and generating an automated text conforming to a current dialogue scene based on the hidden state sequence.

[0050] The technical scheme provided by the present application has the following beneficial effects:

[0051] The present application integrates dialogue text, creative theme, history state and environment data to provide complete input for multi-dimensional information fusion. Through collaborative analysis of the theme and the text, an entity relationship network with weight constraint is established to improve content relevance. High-correlation entity objects are selected to focus on key description elements to avoid deviation of the generated content from the theme. The influence of environmental interference on dialogue history is dynamically compensated to maintain long-term context consistency. The text is output based on the state sequence to ensure that the generated result has both scene adaptability and logical coherence.

[0052] Further, the memory-enhanced attention mechanism is used to convert the environment data into a feature vector, and the creative history state is enhanced, and the three are input into a memory-enhanced unit for multi-round iterative calculation, and finally a hidden state sequence is output.

[0053] And, the application compensates for historical memory bias by environmental characteristics, realizes state progressive optimization through multi-round iteration, and makes the generated hidden state sequence accurately reflect the change of conversation context under environmental interference and maintain the stability of creative theme, thereby improving the adaptability and consistency of text generation in complex scenarios.

[0054] These aspects or other aspects of the application will be more apparent in the following description of the embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0055] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed to be used in the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0056] Figure 1 A flowchart of a large model-based automatic text generation and content creation method provided by an embodiment of the application;

[0057] Figure 2 A structural schematic diagram of a large model-based automatic text generation and content creation system provided by an embodiment of the application;

[0058] Figure 3 A structural schematic diagram of a computing device provided by an embodiment of the application. DETAILED DESCRIPTION

[0059] In order to make those skilled in the art better understand the technical solutions of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application.

[0060] In some processes described in the specification and claims of the present application and the above-mentioned drawings, a plurality of operations appearing in a specific order are included, but it should be clearly understood that these operations can be executed or in parallel without the order in which they appear in this text. The serial numbers of the operations, such as 101, 102, etc., are only used to distinguish different operations, and the serial numbers themselves do not represent any execution order. In addition, these processes can include more or fewer operations, and these operations can be executed in sequence or in parallel. It should be noted that the "first", "second", etc. in this text are used to distinguish different messages, devices, modules, etc., and do not represent the order of sequence, nor do "first" and "second" represent different types.

[0061] The existing attention mechanism-based text generation scheme can integrate environmental data and dialogue text, but the feature splicing and static weighting method used has significant limitations: environmental data is only used as auxiliary features for calculation and cannot form deep coordination with dialogue history, leading to the generated content deviating from the theme in complex multi-round dialogue due to environmental interference; at the same time, the standard attention mechanism lacks dynamic compensation ability in memory processing of historical state, and key information gradually decays in long-term dialogue, causing context break and theme drift. This defect is due to the isolated processing mode of environmental features and text features, and the passive response characteristics of the memory mechanism to environmental interference.

[0062] To solve the above problems, the present application provides an automatic text generation and content creation method based on a large model, which obtains material text data, creation theme information, creation historical state and current environmental data in a multi-round dialogue scene, constructs text structure information with theme weight by combining the material text data and the creation theme information, and determines the core description elements; then, a memory-enhanced attention mechanism is used to dynamically integrate the environmental data, the creation historical state and the core description elements, to generate a hidden state sequence with environmental adaptability, and finally output an automatic text that meets the scene requirements. The method ensures that the content generation is always anchored to the creation goal through the text structure construction guided by the theme information; the memory-enhanced mechanism is used to realize deep coordination between environmental data and dialogue history, and dynamically compensate for memory deviation caused by interference; finally, the state sequence is generated to ensure the long-term coherence and theme consistency of the text. This solution essentially solves the theme drift and context break caused by the isolated processing of environmental features and the passive decay of memory in the prior art, and improves the robustness and creation intention compliance of text generation in complex scenarios.

[0063] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0064] Figure 1 A flowchart of an automatic text generation and content creation method based on a large model provided by an embodiment of the present application is shown in Figure 1 The method comprises the following steps:

[0065] Step 101: obtaining material text data, creation theme information, creation historical state and current environmental data in a multi-round dialogue scene.

[0066] In step 101, the material text data represents the text content input by the user in the current dialogue scenario, including sentences, keywords, etc., for subsequent text analysis and generation. The creation theme information represents the pre-set text generation target or theme, used to constrain the direction and style of the generated content. The creation history state represents the historical dialogue state recorded by the system, including the text of the previous rounds of dialogue, user intent, etc., used to maintain contextual coherence. The environment data represents real-time collected environmental parameters, such as temperature and humidity data, illumination intensity data, and background noise data, etc., used to adjust the generated content to adapt to the actual scene.

[0067] In the embodiments of the present application, the system first obtains the text content of the current dialogue from the user input as the material text data, and reads the pre-set creation theme information to clarify the generation direction. Then, the system retrieves the dialogue records of the last few rounds as the creation history state, ensuring that the newly generated content is logically connected with the historical dialogue. In addition, the system collects temperature, humidity, and light data in real time through environmental sensors as the current environment data. These data together constitute the input basis for subsequent processing.

[0068] For example, in the smart home scenario, the user inputs "the living room is too hot" as the material text data, the system reads the pre-set theme "energy saving and cooling" as the creation theme information, and retrieves the previous two rounds of dialogue "turn on the air conditioner" and "set to 26 degrees" as the creation history state. At the same time, the environmental sensor collects the current living room temperature of 30 degrees and humidity of 60% as the environment data.

[0069] Step 102: Joint the material text data and the creation theme information, and construct text structure information according to the joint result.

[0070] In step 102, the joint result is a weighted entity relationship network that combines the text syntax structure and the theme constraints. The source is to superimpose the entity relationship of the material text and the association weight of the theme keywords, and the final weighted text structure information is obtained by multiplying the original relationship strength between entities by the corresponding theme weight coefficient. The text structure information represents a structured representation containing entity objects and their relationships, where the entity objects represent key nouns in the text, and the relationships describe the logical connections between entities.

[0071] In the embodiments of the present application, the system extracts nominal units from the material text data as entity objects, and analyzes the syntax relationships between them. At the same time, the keywords in the creation theme information are parsed, matched with the entity objects, and assigned weights. For example, if the theme is "energy saving", it has a higher correlation with entities such as "air conditioner" and "temperature". Finally, the system integrates the syntax relationships between entities and the theme weights to generate weighted text structure information.

[0072] For example, the entity objects "living room" and "temperature" are extracted from the user input "the living room is too hot", and their modification relationship is analyzed. The keyword "energy saving" of the theme "energy saving and temperature reduction" is associated with "temperature" with a weight of 0.9, and with "living room" with a weight of 0.7. Finally, the text structure information is generated, including entities and their weighted relationships.

[0073] Step 103: determining the core description element based on the text structure information.

[0074] In step 103, the core description element represents the entity object with the highest weight in the text structure information, representing the core focus of the generated content, such as "temperature", which is the most important entity object selected after sorting.

[0075] In the embodiments of the present application, the system calculates the comprehensive weight of each entity object in the text structure information (combining the theme weight and the interaction frequency), and selects the top N entities as the core description element after sorting by weight. For example, if "temperature" has the highest weight, it is determined as the core element.

[0076] For example, in the text structure information, "temperature" has a weight of 0.9 and "living room" has a weight of 0.7, so "temperature" is selected as the core description element.

[0077] Step 104: using a memory-enhanced attention mechanism to fuse the environment data, the creation history state, and the core description element to generate a hidden state sequence.

[0078] In step 104, the memory-enhanced attention mechanism is an improved attention calculation method that adds a memory enhancement function to the historical state based on the standard attention mechanism, dynamically adjusts the weight of the historical state to compensate for information decay caused by environmental interference, and enables the system to better maintain long-term dialogue consistency. The hidden state sequence refers to a set of intermediate states reflecting the dialogue context and environmental influence, which is used to guide text generation.

[0079] In the embodiments of the present application, the system converts the environment data into a feature vector, and simultaneously enhances the key memory by strengthening the creation history state. Then, the environment feature vector, the strengthened history state, and the core description element are input into the memory-enhanced attention mechanism to generate a hidden state sequence through multiple rounds of calculation. Each calculation takes the state of the previous round as input and gradually optimizes the output.

[0080] For example, the environment data 30 degrees is converted into a feature vector 0.3, and the historical state "set 26 degrees" is strengthened to increase the weight. The memory-enhanced mechanism fuses these data to generate an initial hidden state 0.6 in the first round, 0.8 in the second round, and 0.9 in the final round, forming the sequence [0.6, 0.8, 0.9].

[0081] Step 105: generating an automated text that conforms to the current dialogue scene based on the hidden state sequence.

[0082] In step 105, the current dialogue scene refers to the dialogue context containing the latest user input and real-time environmental state. The judgment basis is that the system records the dialogue round timestamp, marks the latest user input and the environmental data collected at the same time as "current", and continuously updates the scene state to ensure that the generated text is always based on the latest dialogue and environmental information. The automated text represents the final generated natural language reply that conforms to the theme, environment, and context, maintaining semantic coherence with the historical dialogue and environmental adaptability.

[0083] In the embodiments of the present application, the system inputs the hidden state sequence into the text decoder one by one to generate text units step by step. Each time, check if the text unit conforms to the theme constraint (such as prohibiting "force mode" and other conflicting words), and make environmental adaptation adjustment (such as inserting the current temperature value). Finally, combine the units in logical order to form a complete text.

[0084] For example, based on the hidden state sequence [0.6, 0.8, 0.9], the decoder generates "suggest setting the air conditioner" in the first round, "temperature to 27 degrees" in the second round, and "and turn on the energy-saving mode" in the last round. After combination, the output is "suggest setting the air conditioner temperature to 27 degrees and turn on the energy-saving mode".

[0085] This method dynamically fuses environmental data, historical dialogue, and theme information to generate text content that conforms to user intent and adapts to the actual scene. It solves the theme drift and context break caused by environmental interference in traditional solutions, improving the coherence and scene adaptability of the generated text.

[0086] To solve the context break problem caused by environmental interference in multi-round dialogue scenes and further improve the theme consistency and environmental adaptability of text generation, in some embodiments, step 104: the memory-enhanced attention mechanism is used to fuse the environmental data, the creative history state, and the core description element to generate a hidden state sequence, including:

[0087] Step 201: convert the environmental data into an environmental feature vector.

[0088] In step 201, the environmental feature vector is a numerical representation obtained by normalizing the original data collected by environmental sensors (such as temperature, humidity, etc.). The source is obtained by subtracting the reference value from each environmental parameter and dividing by the range, for example, temperature feature value = (current temperature - 25℃) / 10℃, and finally combining the parameter feature values to form a vector.

[0089] In the embodiments of the present application, the system first reads the original data collected by the current environment sensor, including temperature, humidity and other parameters. Then these parameters are converted into standardized values through linear transformation, the temperature value is divided by the temperature range after subtracting the reference temperature, and the humidity value is directly divided by the maximum humidity value. The converted value combination forms an environmental feature vector, which is one of the inputs for subsequent fusion.

[0090] Step 202: Memory reinforcement processing is performed on the creation history state.

[0091] In step 202, the memory reinforcement processing process refers to the operation process of importance weighting and time sequence decay compensation for the historical dialogue state.

[0092] In the embodiments of the present application, the system obtains the text feature representation of the last three rounds of dialogue, and assigns different weight coefficients according to the dialogue round and content correlation. The latest round of dialogue has the highest weight, and the previous two rounds decrease in turn. Then all the historical state features are weighted and summed to highlight the influence of key historical information, and the reinforced creation history state representation is obtained.

[0093] Step 203: Input the environmental feature vector, the reinforced creation history state and the core description element into the memory enhancement unit of the memory enhancement attention mechanism, and output the hidden state sequence through multiple rounds of state iteration.

[0094] In step 203, the memory enhancement unit refers to the calculation module that implements the memory enhancement attention mechanism, and the memory enhancement unit receives the environmental feature vector, the reinforced creation history state and the core description element as input. The multiple rounds of state iteration process refers to the process of gradually optimizing the hidden state through loop calculation.

[0095] In the embodiments of the present application, the system takes the core description element as the query condition, the environmental feature vector as the environmental compensation signal, and the reinforced creation history state as the memory bank. First, calculate the matching degree of the query condition and the memory bank, and filter out the most relevant historical information. Then, the filtering results and the environmental compensation signal are weighted and fused to generate an initial hidden state. Then, the initial state is taken as the input for secondary calculation to generate an intermediate state. Finally, the final hidden state is generated based on the intermediate state, and the three states are sequentially combined to form the hidden state sequence.

[0096] The following is a specific example:

[0097] In an embodiment of the smart home scenario, the system obtains the user's current dialogue "the bedroom needs ventilation" as the material text data, combines the preset theme "healthy air conditioning", calls the previous two dialogues "close the window" and "turn on the air purifier" as the creative history state, and simultaneously collects the current bedroom environment data including the temperature 28 degrees, the humidity 65%, and the air quality index 120. The temperature data is calculated by the formula temperature feature value = (current temperature - comfortable temperature benchmark) / temperature variation range, where the current temperature is 28, the comfortable temperature benchmark is 25, and the temperature variation range is 10, to obtain (28-25) / 10 = 0.3; the humidity data is calculated by the formula humidity feature value = current humidity value / 100 to obtain 65 / 100 = 0.65; the air quality index is calculated by the formula air quality feature value = current index / standard value, where the current index is 120 and the standard value is 100, to obtain 120 / 100 = 1.2. When the history dialogue state is reinforced, the memory strength of the last round "turn on the air purifier" is set to 1.0, and the memory strength of the previous round "close the window" is set to 0.8. The entity objects "bedroom" and "ventilation" are extracted from the current dialogue, the core description element "ventilation" is determined in combination with the theme "healthy air conditioning", and the association weight between the core description element "ventilation" and the theme keyword "health" is set to 0.9. The environment feature vector [0.3, 0.65, 1.2] is input into the memory enhancement unit together with the reinforced history state [1.0, 0.8] and the core description element "ventilation", and when the matching degree of "ventilation" and the history state is calculated, the semantic similarity algorithm is adopted, and the matching degree = semantic similarity x history state memory strength, where the semantic similarity between "ventilation" and "turn on the air purifier" is 0.7, the memory strength is 1.0, and the calculation result is 0.7 x 1.0 = 0.7; the semantic similarity between "ventilation" and "close the window" is 0.5, the memory strength is 0.8, and the calculation result is 0.5 x 0.8 = 0.4. The history state "turn on the air purifier" with a matching degree greater than 0.5 is selected. Then the selection result is combined with the environment feature vector according to the weight, and the weight distribution is temperature 0.3, humidity 0.2, and air quality 0.5, and hidden states 0.7, 0.8, and 0.9 are respectively generated through three iterations.

[0098] In the embodiments of the present application, the memory enhancement attention mechanism is used to deeply integrate the environment data, the history dialogue, and the core element, and the generated hidden state sequence can accurately reflect the influence of environment changes on the dialogue while maintaining the theme consistency. This method effectively solves the context break problem caused by environmental interference, so that the generated text not only meets the actual scene requirements, but also maintains the coherence of the dialogue.

[0099] To solve the problem of context break caused by environmental interference in multi-round dialogue scenarios and further improve the consistency of topic and adaptability of environment in text generation, in some embodiments, step 203: the memory enhancement unit inputs the environmental feature vector, the reinforced creative history state, and the core description element into the memory enhancement attention mechanism, and outputs a hidden state sequence through multi-round state iteration, including:

[0100] Step 301: taking the core description element as a query reference.

[0101] In step 301, the query reference refers to the core basis for retrieving relevant historical dialogues, which is the most critical content element extracted from the current dialogue and topic, and is used to guide the system to select the most relevant historical dialogue fragments.

[0102] In the embodiments of the present application, the system converts the core description element into a feature representation form as a query condition input into the memory enhancement unit. This query reference will be used for subsequent matching calculation with the historical dialogue state to ensure that the selected historical information is highly related to the current dialogue topic.

[0103] Step 302: calculating the matching degree between the query reference and each historical state vector in the reinforced creative history state, and selecting a preset number of target historical states with the highest matching degree from the reinforced creative history state.

[0104] In step 302, the historical state vector is a feature representation of the past dialogue content after reinforcement processing. The matching degree refers to the relevance score between the query reference and the historical state vector. The preset number refers to the number of target historical states to be selected and retained, which is determined according to the dialogue complexity and hardware computing power, and is usually set to 3-5. The value is selected through experimental verification to save the most computing resources under the premise of ensuring the effect, and is set as a fixed parameter during system initialization. The target historical state refers to the top-ranked historical dialogue state in the matching degree calculation result.

[0105] In the embodiments of the present application, the system first calculates the similarity between the query reference feature and each historical state vector feature, and uses dot product operation to measure the matching degree of the two. Then, the matching degrees of all historical states are sorted, and the top-ranked states are selected as the target historical states, which will participate in the subsequent fusion calculation.

[0106] Step 303: merging the target historical state and the environmental feature vector, and performing three consecutive calculations on the merging result to generate three hidden states.

[0107] In step 303, the merging process refers to the operation process of feature splicing and weighted fusion of the target historical state and the environment feature vector. The merging result refers to the fusion feature formed after the selected target historical state and the environment feature vector are weighted and spliced. The source is obtained by linearly combining the target historical state feature and the environment feature vector according to the preset weight ratio (such as a historical state weight of 0.6 and an environment feature weight of 0.4), to form a new feature representation containing historical information and environment features. The three consecutive calculation processes refer to three state update calculations by the recurrent neural network, with the output of the previous calculation as the input each time, to gradually optimize the generation of more accurate hidden states. The three hidden states represent the context representation generated by the system at different calculation stages: the initial hidden state captures the basic environment and historical features, the intermediate hidden state incorporates more dialogue context information, and the final hidden state integrates all elements to form a complete representation, which together constitute the time evolution sequence of the dialogue state.

[0108] In the embodiments of the present application, the system first splices the selected target historical state and the environment feature vector, and then inputs the recurrent neural network for state update. The first calculation generates an initial hidden state with the initial state as the input, the second calculation generates an intermediate hidden state with the initial hidden state as the input, and the third calculation generates a final hidden state with the intermediate hidden state as the input. For example, the system selects the target historical state "turn on the humidifier" (memory strength 0.9) and the environment feature vector [0.4, 0.7] (temperature 0.4 calculated by (26-22) / 10, humidity 0.7 calculated by 70 / 100) for merging processing: first, splice 0.9 of the target historical state with the two environment feature values to form a merging vector [0.9, 0.4, 0.7]; the first calculation inputs the vector into the recurrent neural network to generate an initial hidden state 0.6 in combination with the current user instruction "air is dry"; the second calculation generates an intermediate hidden state 0.75 with 0.6 as the input in combination with the device state data, which strengthens the humidification demand; the third calculation generates a final hidden state 0.85 by comprehensively considering all information, which determines the need to adjust the humidity and temperature at the same time. The final hidden state sequence [0.6, 0.75, 0.85] completely records the gradual reasoning process from environment perception to decision making, and provides a state basis for generating the reply "the current humidity is low, it is suggested to turn on the humidifier and set the air conditioner to 26 degrees comfort mode".

[0109] Step 304: combining the three hidden states in the order of generation of the three hidden states to form a hidden state sequence.

[0110] In step 304, the combination process is to concatenate the three hidden states in time sequence to form a sequence, first taking the initial hidden state as the first element of the sequence, then taking the intermediate state as the second element, and finally taking the final state as the last element, forming an ordered sequence of [initial, intermediate, final], which records the state evolution process for the decoder to use.

[0111] In the embodiments of the present application, the system arranges and combines the three calculated hidden states in time sequence to form the final hidden state sequence. This sequence will be used by the decoder to generate text, ensuring that the output content remains coherent with the conversation history and the current environment.

[0112] The following is a specific example:

[0113] In the embodiment of the smart home scene, the system takes "ventilation" as the query benchmark and calculates its matching degree with the reinforced historical states "turn on air purifier" and "close the window". The matching degree calculation uses a semantic similarity algorithm, in which the semantic correlation score of "ventilation" with "turn on air purifier" is 0.7 and with "close the window" is 0.4, and the preset screening threshold is 0.5, so "turn on air purifier" is selected as the target historical state. The target historical state 1.0 is weighted and merged with the environmental feature vector 0.3, 0.65, 1.2, with weight distribution of temperature 0.3, humidity 0.2, and air quality 0.5, and the merging formula is that the merging result is equal to the target historical state multiplied by 0.4 plus the temperature feature multiplied by 0.3 plus the humidity feature multiplied by 0.2 plus the air quality feature multiplied by 0.1, where 0.4, 0.3, 0.2, and 0.1 are the weight coefficients of each feature. The merging result is calculated three times in succession: the first calculation takes the merging result as input to generate the initial hidden state 0.7 through a neural network; the second takes the initial hidden state 0.7 as input to generate the intermediate hidden state 0.8 in combination with the environmental features; the third calculation generates the final hidden state 0.9 by synthesizing all features. These three states are combined in sequence as 0.7, 0.8, and 0.9.

[0114] In the embodiments of the present application, the hidden state sequence generated by multiple rounds of state iteration can accurately capture the dynamic changes of the conversation context and environmental features, enabling the system to generate text that not only conforms to the current conversation topic but also adapts to environmental changes, effectively solving the context breakage problem caused by environmental interference in traditional methods.

[0115] To solve the problem of inaccurate key information extraction in multi-round conversation scenarios and further improve the topic focus of text generation, in some embodiments, step 103: determining the core description element based on the text structure information, includes:

[0116] Step 401: Calculate the correlation measure value between each pair of entity objects based on the text structure information.

[0117] In step 401, entity objects refer to key noun elements extracted from dialogue text. Correlation measure value refers to a quantitative indicator for measuring the degree of correlation between entity objects, which considers the strength of syntactic relationship and theme relevance weight between entities. The value is obtained by weighting the adjusted interaction frequency and relationship distance.

[0118] In the embodiments of the present application, the system first analyzes the syntactic connection relationship between each pair of entity objects in the text structure information, including subject-predicate relationship, modification relationship, etc. Then, combined with the theme weight given by the creative theme information, the correlation degree of each pair of entities is weighted and calculated, and finally the correlation measure value of each pair of entity objects is obtained.

[0119] Step 402: Sort multiple entity objects according to the correlation measure value.

[0120] In step 402, the sorting process refers to the process of sorting the importance of all entity objects according to the correlation measure value.

[0121] In the embodiments of the present application, the system calculates the sum of the correlation measure values of each entity object with all other entities as the comprehensive correlation degree of the entity. Then, all entities are sorted in descending order of comprehensive correlation degree to generate an entity importance ranking list.

[0122] Step 403: Select the top N entity objects from the sorting result as core description elements, where N is a predetermined positive integer.

[0123] In step 403, the sorting result reflects the importance of each entity under the constraints of dialogue context and theme.

[0124] In the embodiments of the present application, the system selects the top N entity objects from the top of the sorting result as core description elements. These core elements will be the focus of subsequent text generation, ensuring that the generated content focuses on the core topic of the dialogue.

[0125] The following is a specific example:

[0126] In an embodiment of the smart home scenario, the system extracts the entity objects "living room", "air conditioner", "fan" and "power consumption" from the user input "the living room air conditioner and fan are both too power consuming". The relationships between the entities are analyzed: "living room-air conditioner" and "living room-fan" are part-of relationships with a relationship strength of 0.8; "air conditioner-power consumption" and "fan-power consumption" are attribute relationships with a relationship strength of 1.0. In combination with the "energy saving optimization" theme, weights are assigned to each entity: "air conditioner" 1.0, "fan" 0.9, "power consumption" 1.2, and "living room" 0.7. The correlation measure value of each pair of entities is calculated: the correlation measure value of "air conditioner-fan" is 0.8 multiplied by the air conditioner weight 1.0 plus the fan weight 0.9 divided by 2, equaling 0.85; the correlation measure value of "air conditioner-power consumption" is 1.0 multiplied by 1.0 plus 1.2 divided by 2, equaling 1.1; the correlation measure value of "fan-power consumption" is 1.0 multiplied by 0.9 plus 1.2 divided by 2, equaling 1.05; and other combinations are calculated in the same way. The comprehensive correlation of each entity is calculated: air conditioner 0.85+1.1+0.56=2.51, fan 0.85+1.05+0.56=2.46, power consumption 1.1+1.05+0.66=2.81, and living room 0.56+0.56+0.66=1.78. The comprehensive correlation is ranked as power consumption, air conditioner, fan, and living room. The preset N=3, and "power consumption", "air conditioner" and "fan" are selected as the core description elements. Among them, 0.56 in the correlation measure value calculation formula is the correlation measure value of "air conditioner-living room" and "fan-living room", and 0.66 is the correlation measure value of "power consumption-living room", which are calculated according to the foregoing method.

[0127] In the embodiments of the present application, by quantitatively calculating the correlation degree between entities and screening core elements, it is ensured that the system can accurately grasp the key content of the dialogue, and the generated text always revolves around the core topic that the user is most concerned about, effectively avoiding the theme deviation problem caused by interference of secondary information in the traditional method.

[0128] In order to solve the problem of inaccurate entity correlation degree calculation in multi-round dialogue scenarios and further improve the accuracy of core element extraction, in some embodiments, step 401: based on the text structure information, the correlation measure value between each pair of entity objects is calculated, including:

[0129] Step 501: based on the text structure information, obtaining a theme weight value of a plurality of entity objects.

[0130] In step 501, the theme weight value is a quantitative representation of the correlation degree of the entity object with the current creation theme, which is determined according to the matching degree of the entity and the theme keyword, and the higher the matching degree, the greater the weight value.

[0131] In the embodiments of the present application, the system first parses the keywords in the creation theme information, then calculates the semantic similarity of each entity object and the keywords, and takes the normalized similarity value as the theme weight value of the entity.

[0132] Step 502: Calculate the interaction frequency and relationship distance between the entity objects in the text structure information.

[0133] In step 502, the interaction frequency refers to the number of times that the entity objects appear together in multiple rounds of dialogue. The relationship distance refers to the interval degree of the entity objects in the syntax structure. The entities directly connected are close in distance, and the entities indirectly connected are far in distance.

[0134] In the embodiments of the present application, the system traverses the last N rounds of dialogue records, records the cumulative number of times that each pair of entities appears together in the same sentence or adjacent sentences as the interaction frequency, and calculates the relationship distance by analyzing the syntax dependency tree of the current sentence, measuring the shortest path length between the entity nodes. The entities directly connected have a distance of 1, and the distance increases by 1 for each intermediate node.

[0135] Step 503: Adjust the interaction frequency between the entity objects according to the theme weight value.

[0136] In step 503, the process of adjusting the interaction frequency is the process of fusion calculation of the original interaction frequency and the theme weight of the related entity. The adjusted interaction frequency refers to the frequency value considering the theme correlation. By multiplying the original interaction frequency and the theme weight value, the entity pair with higher theme correlation obtains a higher adjusted frequency.

[0137] In the embodiments of the present application, the system takes the average of the theme weight values of each pair of entities, then multiplies the average value with the original interaction frequency, and the formula is: adjusted frequency = original frequency x (weight A + weight B) / 2. This makes the entity pair with high theme correlation obtain a higher adjusted frequency value. In this way, both the historical interaction information and the importance of the theme-related entities are highlighted.

[0138] Step 504: Perform weighted calculation on the adjusted interaction frequency and the relationship distance to generate the correlation measure value.

[0139] In step 504, the weighted calculation process refers to first converting the relationship distance into a closeness coefficient by taking the reciprocal, then multiplying the adjusted interaction frequency with the closeness coefficient, and the formula is: correlation measure value = adjusted frequency / (relationship distance^2). The distance square operation amplifies the advantage of close distance entities, and the final correlation measure value considers the co-occurrence frequency, theme correlation and syntax closeness in three dimensions.

[0140] In the embodiments of the present application, the system divides the adjusted interaction frequency by the square of the relationship distance to obtain the final correlation measure value. The closer the relationship distance of the entity pair, the greater the correlation measure value.

[0141] The following is a specific example:

[0142] In an embodiment of the smart home scene, the system obtains the user's current dialogue "the lights and air conditioner in the bedroom have been on for too long" as input, extracts the entity objects "bedroom", "light", "air conditioner" and "on for too long". Count the co-occurrence of entities in the last five rounds of dialogue: "light-bedroom" appears 3 times, "air conditioner-bedroom" appears 4 times, "light-on for too long" appears 1 time, and "air conditioner-on for too long" appears 2 times. These times are the original interaction frequency. Analyze the text structure, "light-bedroom" and "air conditioner-bedroom" are ownership relationships, and the relationship distance is 1; "light-on for too long" and "air conditioner-on for too long" are state description relationships, and the relationship distance is 1. Combined with the theme of "energy saving reminder", give the entity weights: "light" 0.9, "air conditioner" 1.0, "on for too long" 1.2, "bedroom" 0.7. When adjusting the interaction frequency, "light-bedroom" is adjusted to 3 times the light weight 0.9 plus the bedroom weight 0.7 divided by 2, which is equal to 2.4, "air conditioner-bedroom" is adjusted to 4 times 1.0 plus 0.7 divided by 2, which is equal to 3.4, "light-on for too long" is adjusted to 1 times 0.9 plus 1.2 divided by 2, which is equal to 1.05, and "air conditioner-on for too long" is adjusted to 2 times 1.0 plus 1.2 divided by 2, which is equal to 2.2. When calculating the correlation measure value, divide the adjusted frequency by the square of the relationship distance, "light-bedroom" gets 2.4 divided by 1, which is equal to 2.4, "air conditioner-bedroom" gets 3.4, "light-on for too long" gets 1.05, and "air conditioner-on for too long" gets 2.2. Calculate the comprehensive correlation: "light" is 2.4 plus 1.05 equal to 3.45, "air conditioner" is 3.4 plus 2.2 equal to 5.6, "on for too long" is 1.05 plus 2.2 equal to 3.25, "bedroom" is 2.4 plus 3.4 equal to 5.8. Among them, the calculation formula of the adjusted frequency is divided by 2 to take the average of the weights, and the square of the distance is calculated using the square of the relationship distance 1, which is still 1.

[0143] In the embodiments of the present application, by comprehensively considering the theme relevance, historical interaction and grammatical relationship, the correlation measure value calculated can more comprehensively and accurately reflect the actual correlation degree between entities, providing a reliable basis for the determination of core description elements, and effectively improving the theme focus and context coherence of subsequent text generation.

[0144] In order to solve the problem of incomplete text structure construction in multi-round dialogue scene, and further improve the fit degree of generated content and theme, in some embodiments, step 102: the joint of the material text data and the creation theme information, the text structure information is constructed according to the joint result, including:

[0145] Step 601: Identify a plurality of entity objects and connection relationships between the entity objects from the material text data.

[0146] In step 601, the connection relationship refers to the grammatical association between the entity objects, including subject-predicate, verb-object, modification, and other relationship types.

[0147] In the embodiments of the present application, the system first performs word segmentation and part-of-speech tagging on the material text data to identify the nominal components as candidate entity objects. Then the sentence syntax structure is analyzed to determine the dependency relationship type and strength between the entity objects, and a preliminary entity relationship network is constructed.

[0148] Step 602: Analyze the creation theme information to obtain theme keywords.

[0149] In step 602, the process of analyzing the creation theme information is first to perform word segmentation and part-of-speech tagging on the theme text, and to filter out stop words and retain content words such as nouns and verbs. Then the TF-IDF values of the remaining words are calculated, and the top K words with the highest weights are selected as theme keywords. At the same time, the semantic association between these keywords is analyzed to construct a theme keyword network, and finally the weighted theme keyword set and its association relationship are output. The theme keywords refer to the core concepts extracted from the creation theme information. These keywords represent the core content and direction that needs to be highlighted in text generation.

[0150] In the embodiments of the present application, the system performs word segmentation and semantic analysis on the creation theme information, removes stop words, and selects meaningful content words as theme keywords. At the same time, the importance weight of each keyword is calculated according to the word frequency and position information.

[0151] Step 603: Establish the association relationship between the theme keywords and the entity objects.

[0152] In step 603, the association relationship refers to the semantic connection between the theme keywords and the entity objects, which reflects the relevance of the entity objects in the theme context.

[0153] In the embodiments of the present application, the system first calculates the word vector cosine similarity between each theme keyword and the entity object. Then for the entity-keyword pairs with similarity exceeding a certain threshold, the association strength value is calculated according to the similarity size and keyword weight. Finally, all the theme keywords associated with each entity object and the corresponding association strength are labeled to form an entity-theme association matrix. The association strength is determined by the semantic similarity and the keyword weight.

[0154] Step 604: Form text structure information according to the connection relationship between the entity objects and the association relationship.

[0155] In the embodiments of the present application, the system constructs an initial relationship graph by taking entity objects as nodes and the connection relationship between entities as edges; then adds the associated theme keywords and intensity information to each node in the graph; then adjusts the weight of the edges in the graph by multiplying the original connection relationship strength and the theme weight of the associated node; and finally generates a weighted text structure graph containing both syntactic connection relationship and theme semantic association. Then the association strength of the theme keywords and entities is marked in the graph to form a text structure information with theme weight. This structure not only retains the syntactic features of the original text, but also highlights the core content related to the theme.

[0156] The following is a specific example:

[0157] In the embodiments of the smart home scene, the system obtains the user's current dialogue "The air conditioner temperature in the bedroom is too low" as material text data, identifies the entity objects "bedroom", "air conditioner" and "temperature" from it, and determines that "bedroom-air conditioner" is the belonging relationship and "air conditioner-temperature" is the attribute relationship. At the same time, the system reads the preset theme "comfort and energy saving" as the creation theme information, and parses to obtain the theme keywords "comfort" and "energy saving". The semantic similarity of the keywords and the entities is calculated, the similarity of "comfort" and "temperature" is 0.8, and the similarity of "energy saving" and "air conditioner" is 0.9, the association relationship is established and the weight is assigned. The association weight of "temperature" and "comfort" is 0.8, the association weight of "air conditioner" and "energy saving" is 0.9, and the association weight of "bedroom" and the two keywords is 0.5. Combine these association weights with the connection relationship between entities to form text structure information.

[0158] In the embodiments of the present application, the text structure information constructed by jointly analyzing the text data and the theme information accurately reflects the semantic content of the original dialogue and closely meets the requirements of the creation theme, providing a structured representation with both grammatical correctness and theme relevance for subsequent processing, effectively improving the quality and applicability of the generated content.

[0159] In order to solve the problem of incoordination between text generation and environment adaptation in multi-turn dialogue scene, and further improve the scene applicability and theme consistency of the generated content, in some embodiments, step 105: generating an automated text that meets the current dialogue scene based on the hidden state sequence, includes:

[0160] Step 701: input the hidden state sequence into the text decoder of the pre-trained large model.

[0161] In step 701, the pre-trained large model refers to a language generation model pre-trained based on massive text data through self-supervised learning. The source is a model pre-trained through mask language modeling tasks on large-scale general corpus (such as web text, books, etc.) by using a Transformer architecture, and then fine-tuned on specific domain data, which has strong language understanding and generation capabilities. The text decoder refers to a component in the large model responsible for converting hidden states into text.

[0162] In the embodiments of the present application, the system inputs the hidden state sequence into the text decoder in chronological order. The decoder generates corresponding text segments based on the feature representation of each hidden state.

[0163] Step 702: According to the creation theme information, set the constraint condition of prohibiting the generation of text units conflicting with the creation theme.

[0164] In step 702, the constraint condition refers to the generation restriction rules set according to the creation theme, which are used to ensure that the generated text does not deviate from the preset theme direction.

[0165] In the embodiments of the present application, the system analyzes the core requirements in the creation theme information and converts them into specific generation constraints, such as prohibiting the use of words or expressions conflicting with the theme. These constraints will be enforced in the subsequent generation process.

[0166] Step 703: According to the constraint condition and the time sequence of the hidden state sequence, generate a text unit corresponding to the current hidden state.

[0167] In step 703, the time sequence refers to the order of the states in the hidden state sequence. The current hidden state refers to the state vector at a certain time step being processed by the decoder. The judgment is based on the sequence order, and the one being processed is the current state. The system tracks the current processing position by maintaining a state pointer. The corresponding text unit refers to the text segment generated by the decoder based on the current hidden state. The correspondence here refers to that each hidden state generates a text unit, and there is a one-to-one time sequence correspondence between the state and the unit.

[0168] In the embodiments of the present application, the decoder processes each hidden state in turn. When generating each text unit, it first checks whether the candidate vocabulary meets the constraint condition, filters out the options that do not meet the requirements, and then selects the text unit that best matches the current hidden state from the remaining candidates.

[0169] Step 704: Perform environment adaptation adjustment on the generated text unit.

[0170] In step 704, the environmental adaptation adjustment process is to identify the adjustable parameters in the text unit (such as temperature value, gear, etc.), then calculate the optimal parameter value according to the real-time environmental data, and finally replace the corresponding content in the text and insert the environmental related descriptive sentences, so that the generated content is more consistent with the actual environment.

[0171] In the embodiments of the present application, the system compares the environmental characteristics with the generated text unit, inserts environmental data related descriptions in the text, or adjusts certain parameter values to make them more consistent with the actual environment.

[0172] Step 705: Combine the text units after adaptation and adjustment according to the preset dialogue logic sequence, and output the automated text.

[0173] In step 705, the preset dialogue logic sequence refers to the text organization method that conforms to human dialogue habits. The preset is to analyze a large number of real dialogue corpus, summarize the general dialogue mode such as "problem description-solution-supplementary explanation", and solidify these modes in the system as the default text combination rule.

[0174] In the embodiments of the present application, the system arranges and combines each adjusted text unit according to the logic sequence of "problem description-suggested solution-supplementary explanation", adds necessary conjunctions, and forms a complete final reply.

[0175] The following is a specific example:

[0176] In an embodiment of the smart home scenario, the system inputs the hidden state sequence [0.6, 0.8, 0.9] into the text decoder of the pre-trained large model. According to the "energy-saving cooling" creative theme, the constraint condition is set to prohibit the generation of expressions such as "powerful cooling" and "minimum temperature" that conflict with energy saving. The decoder first generates the text unit "detected living room" based on the initial hidden state 0.6, and then adjusts it to "detected living room temperature 30 degrees" by combining the environmental temperature of 30 degrees; generates "suggest adjusting to" based on the intermediate hidden state 0.8, and calculates the temperature suggestion value as the base temperature of 26 degrees plus the humidity compensation value of 1 degree (the humidity compensation value is equal to the current humidity divided by 100 multiplied by the compensation coefficient 2, where the compensation coefficient 2 is a preset value) according to the current humidity of 60% and the energy saving requirement, resulting in 27 degrees, and generates "suggest adjusting to 27 degrees"; generates "and use energy-saving operation" based on the final hidden state 0.9, and adjusts the expression to "and automatically enable energy-saving mode" according to the difference between the set value of 26 degrees in the historical state "set to 26 degrees" and the current suggestion value of 27 degrees. Finally, the adjusted text units are combined in the logical order of "current situation description - suggestion scheme - supplementary explanation", and the complete reply "detected living room temperature 30 degrees, suggest adjusting to 27 degrees and automatically enable energy-saving mode" is output, which not only meets the energy-saving theme requirement, but also takes into account the actual environmental data, while maintaining a natural connection with the historical operation. The calculation formula of the temperature suggestion value is: suggested temperature = base temperature + (current humidity / 100) x compensation coefficient, the base temperature is 26 degrees, and the compensation coefficient is 2, the calculation process is 26 + (60 / 100) x 2 = 27.2, rounded to 27 degrees.

[0177] In the embodiments of the present application, the theme consistency is ensured by the constraint condition, and the scene applicability is enhanced by the environmental adaptation, and finally the generated text accurately conveys the key information, maintains good readability and practicality, and effectively improves the user experience and service quality of the intelligent dialogue system.

[0178] Figure 2 The structure diagram of an automatic text generation and content creation system based on a large model provided in an embodiment of the present application is shown in Figure 2 as shown, the system comprises:

[0179] The acquisition module 21 is configured to acquire material text data, creative theme information, creative historical state and current environmental data in a multi-round dialogue scenario.

[0180] The construction module 22 is configured to jointly the material text data and the creative theme information, and construct text structure information according to the joint result.

[0181] The determination module 23 is configured to determine the core description element based on the text structure information.

[0182] The first generation module 24 is configured to adopt a memory-enhanced attention mechanism to fuse the environment data, the creation history state, and the core description element to generate a hidden state sequence.

[0183] The second generation module 25 is configured to generate an automated text conforming to a current dialogue scene based on the hidden state sequence.

[0184] Figure 2 The large model-based automated text generation and content creation system can perform Figure 1 The large model-based automated text generation and content creation method of the embodiments described above has the implementation principle and technical effects, which will not be repeated here. The specific operation of each module and unit of the large model-based automated text generation and content creation system described above has been described in detail in the embodiments related to the method, and will not be described in detail here.

[0185] In one possible design, Figure 2 The large model-based automated text generation and content creation system of the embodiments described above can be implemented as a computing device, such as a server. Figure 3 As shown in the figure, the computing device can include a storage component 31 and a processing component 32.

[0186] The storage component 31 stores one or more computer instructions, wherein the one or more computer instructions are called and executed by the processing component 32.

[0187] The processing component 32 is configured to execute the above Figure 1 The large model-based automated text generation and content creation method of the embodiments described above.

[0188] The processing component 32 can include one or more processors to execute computer instructions to complete all or part of the steps in the above method. Of course, the processing component can also be one or more application-specific integrated circuits (ASIC), digital signal processors (DSP), digital signal processing devices (DSPD), programmable logic devices (PLD), field programmable gate arrays (FPGA), controllers, microcontrollers, microprocessors, or other electronic components, for executing the above method.

[0189] The storage component 31 is configured to store various types of data to support the operation of the terminal. The storage component can be implemented by any type of volatile or nonvolatile storage devices or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read only memory (EEPROM), erasable programmable read only memory (EPROM), programmable read only memory (PROM), read only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0190] Of course, the computing device can also include other components, such as an input / output interface, a display component, a communication component, etc.

[0191] The input / output interface provides an interface between the processing component and peripheral interface modules, which can be output devices, input devices, etc.

[0192] The communication component is configured to facilitate wired or wireless communication between the computing device and other devices, etc.

[0193] Among them, the computing device can be a physical device or an elastic computing host provided by a cloud computing platform, at which time the computing device can refer to a cloud server, and the processing component, the storage component, etc. can be basic server resources rented or purchased from the cloud computing platform.

[0194] The embodiments of the present application also provide a computer storage medium storing a computer program, and the computer program can implement the above-mentioned Figure 1 The embodiment shown in the figure is a large model-based automatic text generation and content creation method.

[0195] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-described system, device and unit can refer to the corresponding process in the foregoing method embodiments, which will not be described here.

[0196] The device embodiments described above are merely illustrative, wherein the units described as separate components can or can not be physically separate, and the components displayed as units can or can not be physical units, i.e., can be located in one place, or can be distributed to multiple network units. Part or all of the modules can be selected to achieve the purposes of the embodiments according to actual needs. Those skilled in the art can understand and implement without creative labor.

[0197] Through the description of the above embodiments, those skilled in the art can clearly understand that the embodiments can be realized by means of software and the necessary general hardware platform, and of course can also be realized by hardware. Based on such understanding, the above technical solutions can be embodied in the form of software products, and the computer software products can be stored in a computer readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and include a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods described in each embodiment or some parts of the embodiments.

[0198] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A large model-based automated text generation and content creation method, characterized in that, The method comprises the following steps: Obtaining material text data, creation theme information, creation history state and current environment data in a multi-round dialogue scene; Jointing the material text data and the creation theme information, and constructing text structure information according to the jointing result; Determining a core description element based on the text structure information; Fusing the environment data, the creation history state and the core description element by using a memory-enhanced attention mechanism, and generating a hidden state sequence; Generating an automatic text conforming to the current dialogue scene based on the hidden state sequence; The method comprises the following steps: Converting the environment data into an environment feature vector; Memory-enhancing the creation history state; Inputting the environment feature vector, the enhanced creation history state and the core description element into a memory-enhanced unit of the memory-enhanced attention mechanism, and outputting a hidden state sequence through multi-round state iteration.

2. The method of claim 1, wherein, The method comprises the following steps: Taking the core description element as a query reference; Calculating the matching degree between the query reference and each historical state vector in the enhanced creation history state, and selecting a preset number of target historical states with the highest matching degree from the enhanced creation history state; Merging the target historical states and the environment feature vector, and performing three consecutive calculations on the merging result to generate three hidden states; Combining the three hidden states according to the generation order of the three hidden states to form a hidden state sequence.

3. The method of claim 1, wherein, The method comprises the following steps: Based on the text structure information, calculating the correlation measure value between each pair of entity objects; According to the correlation measure value, sorting a plurality of entity objects; Selecting the top N entity objects from the sorting result as the core description element, wherein N is a preset positive integer.

4. The method of claim 3, wherein, The method comprises the following steps: Based on the text structure information, obtaining the theme weight value of a plurality of entity objects; Calculating the interaction frequency and relationship distance between the entity objects in the text structure information; Adjusting the interaction frequency between the entity objects according to the theme weight value; Weighted calculation of the adjusted interaction frequency and the relationship distance generates a correlation measure value.

5. The method of claim 1, wherein, The method comprises the following steps: Identifying a plurality of entity objects and the connection relationship between the entity objects from the material text data; Analyzing the creation theme information to obtain theme keywords; Establishing the association relationship between the theme keywords and the entity objects; According to the connection relationship between the entity objects and the association relationship, forming a text structure information.

6. The method of claim 1, wherein, The generating of the automatic text in accordance with the current dialogue scene based on the hidden state sequence comprises: inputting the hidden state sequence into a text decoder of a pre-trained large model; setting a constraint condition of prohibiting generation of a text unit conflicting with the creation theme according to the creation theme information; generating a text unit corresponding to the current hidden state according to the constraint condition and the time sequence of the hidden state sequence; performing environment adaptation adjustment on the generated text unit; combining the text units after the adaptation adjustment in a preset dialogue logic sequence to output the automatic text.

7. A large model based automated text generation and content creation system, characterized in that, comprise: an acquisition module configured to acquire material text data, creation theme information, creation history state, and current environment data in a multi-round dialogue scene; a construction module configured to jointly process the material text data and the creation theme information, and construct text structure information according to a joint result; a determination module configured to determine a core description element based on the text structure information; a first generation module configured to adopt a memory-enhanced attention mechanism to fuse the environment data, the creation history state, and the core description element, and generate a hidden state sequence; a second generation module configured to generate an automatic text in accordance with the current dialogue scene based on the hidden state sequence; the adopting of the memory-enhanced attention mechanism to fuse the environment data, the creation history state, and the core description element, and generate a hidden state sequence comprises: converting the environment data into an environment feature vector; performing memory enhancement processing on the creation history state; inputting the environment feature vector, the enhanced creation history state, and the core description element into a memory-enhanced unit of the memory-enhanced attention mechanism, and outputting a hidden state sequence through multi-round state iteration.

8. A computing device, comprising: comprise a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement the method of claim 1-6.

9. A computer storage medium, characterized in that a computer program is stored, and the computer program is executed by a computer to implement the method of claim 1-6.

Images