Text outline automatic generation method and system based on artificial intelligence

Abstract

The invention relates to the technical field of data processing, in particular to a text outline automatic generation method and system based on artificial intelligence, and the method comprises the steps: calculating the weight of a temperature parameter in a PEGASUS model decoding stage, training a PEGASUS model to obtain a text outline generation model, and generating a text outline; processing a to-be-processed document by using the text outline generation model to generate a text outline; the calculation method of the weight of the temperature parameter comprises the following steps: converting a preprocessed original corpus into a high-dimensional word vector sequence; extracting a hierarchical tree which generates the candidate outline at present, and calculating a hierarchical aggregation coefficient of each hierarchical node in the hierarchical tree; calculating a semantic fidelity index of the hierarchical node; calculating an iterative stability index based on the semantic fidelity index; and taking the normalized iteration stability index as the weight of the temperature parameter. According to the method, the problem of logic chaos when a traditional generative model is used for processing a long professional document is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and in particular to a text outline automatic generation method and system based on artificial intelligence. BACKGROUND

[0002] With the explosive growth of digital information, it has become an urgent need in the fields of scientific research, office work, content creation and academic research to quickly extract the core logic from massive long texts and generate structured outlines. The development of artificial intelligence technology, especially natural language processing, has made it possible to automatically generate text outlines. Not only does the outline generation require the system to accurately capture the central idea of the text, but it also requires the generated content to have a high degree of logical coherence, hierarchy and key information coverage, thereby helping users grasp the article context in a short period of time. In the prior art, the PEGASUS model is a representative model in the field of generative text summarization. This algorithm has a significant advantage in processing text structure and semantic extraction through the missing sentence generation task in large-scale pre-training, and can generate semantically coherent and highly summarized text segments.

[0003] However, due to the unique characteristics of the text outline generation task, such as strong logical hierarchy and high concentration of domain knowledge, directly using the original PEGASUS model can lead to technical problems such as logical level confusion and inability to accurately distinguish major and minor points when facing long professional documents. The reason is that the loss function of the original algorithm mainly focuses on overall semantic similarity, and lacks fine description of structured features such as chapter association and indentation level of the document, making it difficult to directly use the generated outline to guide actual document reconstruction or in-depth reading. SUMMARY

[0004] In order to solve the technical problem that the generated outline has logical level confusion and cannot accurately distinguish major and minor points when facing long professional documents using the PEGASUS model, the present application provides a text outline automatic generation method and system based on artificial intelligence.

[0005] In a first aspect, the present application provides a text outline automatic generation method based on artificial intelligence, which adopts the following technical solution: A text outline automatic generation method based on artificial intelligence, comprising the steps of: calculating the weight of the PEGASUS model decoding stage temperature parameter, training the PEGASUS model to obtain a text outline generation model, and processing the text outline generation model to generate a text outline for the document to be processed; The calculation method of the weight of the temperature parameter is: obtaining original corpus and preprocessing, converting the preprocessed original corpus into a high-dimensional word vector sequence; extracting a hierarchical tree of the current generated candidate outline, calculating a hierarchical aggregation coefficient of each hierarchical node in the hierarchical tree, the hierarchical aggregation coefficient being negatively correlated with a vector standard deviation of a same-level node set to which the corresponding hierarchical node belongs; calculating a semantic fidelity index of the hierarchical node, the semantic fidelity index being negatively correlated with the hierarchical aggregation coefficient; calculating an iteration stability index based on the semantic fidelity index, the iteration stability index being positively correlated with the semantic fidelity index; and normalizing the iteration stability index, taking the normalized iteration stability index as the weight of the temperature parameter.

[0006] By introducing the hierarchical aggregation coefficient, the semantic fidelity index and the iteration stability index to dynamically adjust the weight of the temperature parameter in the decoding stage of the PEGASUS model, the problems of logical level confusion and unclear major and minor points that are prone to occur when the existing PEGASUS model directly processes long professional documents are effectively solved, the logical coherence, hierarchical structure rationality and core semantic retention capability of the generated outline are improved, and it is ensured that the generated outline conforms to the macro context of the document and accurately covers the micro details.

[0007] Preferably, the calculation method of the hierarchical aggregation coefficient is: calculating a cosine similarity between the hierarchical node vector and the direct parent hierarchical node vector, and taking the ratio of the cosine similarity to the vector standard deviation of the same-level node set to which the corresponding hierarchical node belongs as the hierarchical aggregation coefficient.

[0008] By calculating the semantic similarity between the hierarchical node and its parent hierarchical node and the dispersion degree between the same-level nodes, the hierarchical aggregation coefficient is constructed, which can quantitatively evaluate the structural rationality of the generated outline, effectively prevent the logical misplacement problem in the outline generation process, ensure that the sub-items are closely related to the parent item theme and the same-level sub-items maintain appropriate semantic differences to reduce redundancy, and thus improve the structural quality of the outline.

[0009] Preferably, before calculating the semantic fidelity index of the hierarchical node, it further includes: vectorizing the core semantics of the full text by using an encoder, calculating the modulus of each core semantic vectorization, and segmenting the modulus of the core semantic vectorization to obtain a theme keyword sequence.

[0010] By vectorizing the core semantics of the full text by using the encoder and extracting the theme keyword sequence based on the modulus, the central idea and key information of the document can be accurately captured, providing a reliable benchmark reference for subsequent calculation of semantic fidelity, preventing the model from deviating from the main theme of the document when generating a deep outline, and enhancing the coverage of the generated outline to the core content of the original text.

[0011] Preferably, the modulus of the core semantics is segmented by using the Otsu threshold segmentation method, and a sequence of core semantics greater than a segmentation threshold is taken as a sequence of topic keywords of the document.

[0012] The segmentation threshold of the core semantics is adaptively determined by using the Otsu threshold segmentation method, and compared with fixed threshold cutting, the method can more objectively and dynamically filter out the most representative topic keywords, avoids missing of keywords or introduction of noise caused by manual setting of the threshold, and further improves the accuracy and robustness of extraction of the key information.

[0013] Preferably, the semantic fidelity index of the hierarchical node is calculated by: calculating a mutual information matching score of the hierarchical node outline item and the sequence of topic keywords; calculating a product of a logical depth where the hierarchical node is located and a hierarchical aggregation coefficient, and taking the product as an exponential term of an exponential function; and taking a ratio of the mutual information matching score and the exponential function as the semantic fidelity index.

[0014] By combining the mutual information matching score and the hierarchical depth to construct the semantic fidelity index, the semantic drift phenomenon in the deep outline generation process can be effectively monitored and inhibited, the problem of running off topic or generating irrelevant details in the existing algorithm in multi-level generation is solved, the outline is ensured to be consistent with the core topic of the document while being deep into details, and the unification of the global logic and the local semantics is realized.

[0015] Preferably, the calculation method of the iteration stability index is: a sequence of the semantic fidelity indexes generated in the last multiple iterations is taken as an outline index sequence, and an expression of the iteration stability index is:

[0016] wherein, indicates the iteration stability index of the i th iteration; indicates the iteration stability index of the i th iteration; indicates the average value of the semantic fidelity indexes of all hierarchical nodes in the i th generation result; indicates the average value of the semantic fidelity indexes of all hierarchical nodes in the i th generation result; indicates the average value of the semantic fidelity indexes of all hierarchical nodes in the i th generation result; indicates the average value of the semantic fidelity indexes of all hierarchical nodes in the i th generation result; indicates the average value of the semantic fidelity indexes of all hierarchical nodes in the i th generation result; indicates the average value of the semantic fidelity indexes of all hierarchical nodes in the i th generation result; indicates the number of data in the outline index sequence; indicates a summation count variable; indicates a preset parameter.

[0017] The semantic fidelity variation in the generation process is quantified by calculating the iterative stability index, the dynamic stability of the generation system can be captured in real time, the quality of the outline is prevented from fluctuating sharply due to parameter randomness, accurate quantitative feedback signals are provided for dynamic optimization of the model parameters, and it is ensured that the final output outline is a result that is fully converged and stable in quality.

[0018] Preferably, the method for processing a to-be-processed document to generate a text outline by using a text outline generation model comprises: processing the to-be-processed document to generate a first-level outline by using the text outline generation model, taking the generated first-level outline as context, and generating deep sub-items in combination with a temperature parameter adjusted by the iterative stability index; calculating the iterative stability index in real time after each round of iteration, and stopping iteration and outputting a final text outline with clear levels and coherent semantics when a convergence requirement is reached.

[0019] The temperature parameter generated is adjusted in real time by using the iterative stability index, randomness is automatically inhibited when the generation process is unstable, and appropriate creativity is allowed when the generation process is stable, the dynamic feedback mechanism makes the generated deep sub-items have rich details and will not go out of control, and through multiple rounds of iteration and refinement, a high-quality text outline with clear hierarchical structure and coherent semantics is finally output.

[0020] Preferably, the method for preprocessing the original corpus is: removing web noise by using a regular matching method; structurally splitting long texts by using a segmentation function, and identifying and retaining original chapter numbers and hierarchical titles; calling an NLP word segmentation tool to perform Chinese word segmentation, part-of-speech tagging, and stop word filtering.

[0021] Preferably, the method for obtaining the original corpus is: collecting to-be-processed long text data through an enterprise internal knowledge base API, and obtaining the original corpus through keyword search and URL deep traversal.

[0022] Through the enterprise internal knowledge base API in combination with the keyword search and deep traversal technology, efficient and automated collection of to-be-processed long text data is realized, an original corpus library containing rich format information can be quickly constructed, and the data input needs of a large-scale document outline generation task are met.

[0023] In a second aspect, the present application provides a text outline automatic generation system based on artificial intelligence, and adopts the following technical scheme: A text outline automatic generation system based on artificial intelligence comprises a processor and a memory, and the memory stores computer program instructions, which, when executed by the processor, implement the text outline automatic generation method based on artificial intelligence.

[0024] The computer program generated by the above-mentioned artificial intelligence-based text outline automatic generation method is stored in the memory to be loaded and executed by the processor, so that the system is made according to the memory and the processor, and the use is facilitated.

[0025] The present application has the following technical effects: By constructing the hierarchical aggregation coefficient, the semantic fidelity index and the iterative stability index, the structure logic, the core semantic reservation and the stability of the generation process of the outline are quantitatively evaluated respectively, the temperature parameter of the model decoding stage is adaptively adjusted by using the iterative stability index, the problems of logical confusion, semantic drift and detail redundancy of the traditional generation model in processing long professional documents are effectively solved, and the automatic generation of high-quality, hierarchical clear and semantic coherent outline is realized. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 is a flowchart of the artificial intelligence-based text outline automatic generation method of the present application.

[0027] Figure 2 is a contrast schematic view of the semantic fidelity changing with the hierarchical depth.

[0028] Figure 3 is a contrast schematic view of the iterative process convergence stability.

[0029] Figure 4 is a schematic view of the comprehensive dimension evaluation of the outline generation quality. DETAILED DESCRIPTION

[0030] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the 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.

[0031] The embodiments of the present application disclose an artificial intelligence-based text outline automatic generation method, referring to Figure 1 , comprising the following steps: S1: data acquisition and preprocessing.

[0032] The long text data to be processed is collected through the enterprise internal knowledge base API, and the collection equipment includes a cloud server cluster with high-performance processing capacity and a distributed storage array. During the collection process, the Requests library and the Scrapy framework of Python are used as software tools to obtain original corpus containing HTML tags or PDF format through keyword retrieval and URL depth traversal.

[0033] The original corpus is fine-tuned. First, regular matching techniques are used to remove web noise, such as ads, illegal characters, and extra blank lines. Then, a segmentation function is used to structure the long text, identify and retain the original chapter numbers and hierarchical titles. Second, the NLP segmentation tool is used for Chinese segmentation, part-of-speech tagging and stop word filtering. Finally, the preprocessed original corpus is converted into a high-dimensional word vector sequence using the BERT encoder.

[0034] S2: Constructing hierarchical aggregation coefficients.

[0035] In the text outline generation process, due to the tree topology structure of the document, there are semantic overlap and cross-level logical dependence features. According to the information increment theory, the outline items at the same level should maintain a moderate degree of dispersion in the semantic space to reduce redundancy, while the child items and parent items should exhibit a high degree of implicit semantic correlation. This feature causes the data to exhibit a clear cluster distribution and centripetal force offset rule in the vector space. Directly using the original algorithm often ignores this inter-level constraint, resulting in the generated outline items being semantically correct but misclassified in the hierarchy. Therefore, hierarchical aggregation coefficients need to be constructed to reflect the hierarchical logical properties of the th generated node.

[0036] First, the hierarchical tree of the current generated candidate outline is extracted, each hierarchical node is represented by a vector, and the cosine similarity between the vector of each hierarchical node and the vector of its direct parent hierarchical node is calculated. The expression of the hierarchical aggregation coefficient is:

[0037] wherein, represents the hierarchical aggregation coefficient of the th hierarchical node; represents the semantic feature vector of the th hierarchical node outline item; represents the semantic feature vector of the direct parent hierarchical node corresponding to the th hierarchical node; represents the vector dot product operation, used to calculate the semantic correlation; represents the vector standard deviation of the same level node set to which the th hierarchical node belongs, used to measure the semantic dispersion of the same level; is a preset parameter, taking a fixed reference value to avoid the case where the denominator is zero. Among them and are obtained by pre-training deep learning models.

[0038] The cosine similarity between the semantic vector of the current level node and the semantic vector of its direct parent level node is calculated to depict the inheritance degree of the child node to the parent node topic. The standard deviation of the semantic vector of the sibling node set is calculated to measure the semantic dispersion level between different nodes in the same level. When the semantic similarity of the sibling nodes is high, the vector standard deviation is small, indicating that there is a risk of information redundancy in the level.

[0039] Therefore, by normalizing the combination of semantic correlation and sibling dispersion, the level aggregation coefficient constructed can comprehensively reflect the aggregation degree of the current node in the level structure. When the level aggregation coefficient is large, it indicates that the node is highly consistent with the parent node in semantics and has low differentiation degree with the sibling nodes, which has the risk of level belonging ambiguity or structural redundancy. When the level aggregation coefficient is small, it indicates that the level node inherits the superior level topic while maintaining reasonable differentiation degree, and the level structure is more reasonable.

[0040] S3: Construct a semantic fidelity index.

[0041] On the basis of the level feature, due to the outline generation involving multiple rounds of iteration, the semantic drift phenomenon may occur, i.e. the child items are prone to deviate from the core topic with the deepening of the level. According to the attention distribution dilution law, the model pays less attention to the initial global feature when processing deep details, resulting in irrelevant details at the end of the outline. To reflect this feature, a semantic fidelity index reflecting the semantic retention degree of the first node is constructed . The index uses the level aggregation coefficient as a weight gain to evaluate the retention degree of the deep outline to the core semantics.

[0042] Before constructing the semantic fidelity index, the BERT encoder is used to vectorize the core semantics of the full text, the modulus of each core semantic vectorization is calculated, all core semantics are taken as the input of the Otsu threshold segmentation method, the segmentation threshold of the core semantic modulus is obtained, the sequence composed of the core semantics greater than the segmentation threshold is taken as the sequence of the document topic keywords, and the matching score of the current node and the core sequence is calculated. The expression of the semantic fidelity index is:

[0043] Wherein, represents the semantic fidelity index of the first level node; represents the level aggregation coefficient of the first level node; represents the mutual information matching score of the first level node outline item and the topic keyword sequence, and the value range is ; represents the semantic fidelity index of the first the logical depth of the hierarchical node (e.g., the first-level title ); exp() represents the exponential function with the natural constant as the base; the calculation method of the mutual information matching score is prior art, and the specific calculation steps are not described here.

[0044] When the outline item can still accurately anchor the theme at a deep level and the logic is clear, the semantic fidelity index is greater. This index solves the problem that existing algorithms are prone to digressions when generating at multiple levels, ensuring the unity of the global and local of the outline.

[0045] S4: Construct an iteration stability index.

[0046] During the iterative refinement process, gradient oscillation and generated semantic collapse phenomena may occur. Since the outline generation is a discrete search process, parameter changes may cause the generated outline to fluctuate dramatically between different iteration rounds. According to the dynamic balance theory, a high-quality outline generation process should tend to be stable. Therefore, an iteration stability index reflecting the stability of the th iteration round needs to be constructed. This index reflects the robustness of the generation system through the changes in the semantic fidelity index .

[0047] First, the sequence of semantic fidelity indexes generated by the last iteration is denoted as the outline index sequence, and the absolute value of the change in the semantic fidelity index at time step t is calculated. The expression of the iteration stability index is:

[0048] wherein denotes the iteration stability index of the th iteration round; is the average of the semantic fidelity indexes of all hierarchical nodes in the result of the th iteration round; is the average of the semantic fidelity indexes of all hierarchical nodes in the result of the th iteration round; is the average of the semantic fidelity indexes of all hierarchical nodes in the result of the th iteration round; is the number of data in the outline index sequence, which takes a preset reference value ; is the cumulative operation within the window; is a preset parameter, which takes a fixed reference value to avoid a zero denominator.

[0049] When the iteration process gradually tends to be stable and the quality continuously improves, the denominator becomes smaller, and the iteration stability index significantly increases. The index can capture abnormal jumps in the generation process in real time, providing accurate feedback for parameter tuning.

[0050] S5: improving the PEGASUS model and generating a text outline.

[0051] The PEGASUS model is trained to obtain a text outline generation model. During the training process, the iterative stability index is normalized, and the normalized iterative stability index is used as the weight of the temperature parameter in the model decoding stage. When the normalized iterative stability index is lower than a preset stability threshold, it indicates that the current generation process is unstable, and the system automatically reduces the temperature parameter to suppress randomness. In this embodiment, the stability threshold is 0.7.

[0052] The text outline generation model is used to process the to-be-processed document to generate a primary outline, and then enters an iterative refinement cycle: the text outline generation model takes the generated outline as the context, and generates deep sub-items in combination with the temperature parameter adjusted by the iterative stability index . After each round of generation, the hierarchical aggregation coefficient , the semantic fidelity index and the iterative stability index are calculated in real time. When the convergence requirement is met, the iteration is stopped and the final text outline with clear hierarchy and semantic coherence is outputted, wherein the convergence requirement is that the number of iterations reaches a preset number, and in this embodiment, the preset number is 50.

[0053] The technical effects of the present application can also be demonstrated in combination with the drawings, Figure 2 which are a comparison diagram of the semantic fidelity changes of the present application and the prior art at different outline logical depths. The horizontal coordinate represents the logical depth level of the outline, and the vertical coordinate represents the semantic fidelity index value. The dotted curve represents the semantic fidelity change trend of the prior art, and the solid curve represents the semantic fidelity change trend of the present application.

[0054] As can be seen from the figure, as the logical depth of the outline increases, the semantic fidelity index of the prior art presents a significant downward trend, while the present application method can still maintain a high and stable semantic fidelity level under a multi-level structure, indicating that the present application effectively suppresses the semantic drift phenomenon in the multi-level generation process by introducing the hierarchical aggregation coefficient and the semantic fidelity constraint mechanism, and improves the consistency between the deep outline content and the core theme.

[0055] Figure 3For the convergence stability comparison between the method of the present application and the existing method in the iterative generation process, a schematic diagram is shown in Figure 1, wherein the horizontal axis represents the number of iterations, and the vertical axis represents the value of the iterative stability index. Figure 3 It can be seen that in the initial stage of iteration, the iterative stability index of the present application rises rapidly and tends to be stable within a few rounds, while the existing method has obvious fluctuation phenomenon in the process of multiple iterations, and the stability improves slowly. It shows that by constructing the iterative stability index based on the change of semantic fidelity and introducing it into the temperature parameter adjustment process in the decoding stage, the generation model can adaptively suppress random disturbance in the training and inference process, and improve the convergence speed and robustness of the iterative process.

[0056] Figure 4 For the comprehensive comparison between the method of the present application and the existing method under the multi-dimensional evaluation index of outline generation quality, the diagram is a radar chart structure, which is used to compare the comprehensive performance of the method of the present application and the existing method under multiple evaluation dimensions such as hierarchical rigor, semantic fidelity, generation stability, key information coverage and field accuracy, from Figure 4 It can be seen from Figure 2 that the index value of the method of the present application in each evaluation dimension is significantly higher than that of the existing method, especially in the rationality of hierarchical structure, semantic consistency and generation stability, which is more significant. It shows that the present application realizes all-round optimization of the quality of text outline generation by multi-index joint constraint and dynamic adjustment mechanism, and improves the overall reliability and practical value of the generated results.

[0057] The embodiment of the present application also discloses a text outline automatic generation system based on artificial intelligence, which comprises a processor and a memory, and the memory stores computer program instructions.

[0058] The above system also includes a communication bus and a communication interface and other components familiar to those skilled in the art, the setting and function of which are known in the art, so they will not be repeated here.

[0059] The above are the preferred embodiments of the present application, but not limit the protection scope of the present application, therefore: any equivalent changes made on the structure, shape, principle of the present application should be covered within the protection scope of the present application.

Claims

1. An artificial intelligence-based automatic text outline generation method, characterized by, The method comprises the steps of: calculating the weight of the temperature parameter of the PEGASUS model decoding stage, training the PEGASUS model to obtain a text outline generation model, and processing the to-be-processed document by using the text outline generation model to generate a text outline; The calculation method of the weight of the temperature parameter is: obtaining and preprocessing the original corpus, converting the preprocessed original corpus into a high-dimensional word vector sequence; extracting the hierarchical tree of the current generated candidate outline, calculating the hierarchical aggregation coefficient of each hierarchical node in the hierarchical tree, the hierarchical aggregation coefficient being negatively correlated with the vector standard deviation of the same level node set to which the corresponding hierarchical node belongs; calculating the semantic fidelity index of the hierarchical node, the semantic fidelity index being negatively correlated with the hierarchical aggregation coefficient; calculating the iterative stability index based on the semantic fidelity index, the iterative stability index being positively correlated with the semantic fidelity index; and normalizing the iterative stability index, and taking the normalized iterative stability index as the weight of the temperature parameter. 2.The method of claim 1, wherein, The calculation method of the hierarchical aggregation coefficient is: calculating the cosine similarity of the hierarchical node vector and the direct parent hierarchical node vector, and taking the ratio of the cosine similarity to the vector standard deviation of the same level node set to which the corresponding hierarchical node belongs as the hierarchical aggregation coefficient. 3.The method of claim 1, wherein, Before calculating the semantic fidelity index of the hierarchical node, the method further comprises: vectorizing the full-text core semantics by using an encoder, calculating the modulus of each core semantic vectorization, and segmenting the modulus of the core semantic vectorization to obtain a topic keyword sequence.

4. The method of claim 3, wherein the method further comprises: The modulus of the core semantics is segmented by using the Otsu threshold segmentation method, and the sequence of the core semantics greater than the segmentation threshold is taken as the topic keyword sequence of the document.

5. The method of claim 1, wherein the method is based on artificial intelligence. The calculation method of the semantic fidelity index of the hierarchical node is: calculating the mutual information matching score of the hierarchical node outline item and the topic keyword sequence; calculating the product of the logical depth at which the hierarchical node is located and the hierarchical aggregation coefficient, taking the product as the exponential term of an exponential function, and taking the ratio of the mutual information matching score to the exponential function as the semantic fidelity index.

6. The method of claim 1, wherein the method further comprises: The calculation method of the iterative stability index is: taking the sequence of the semantic fidelity indexes generated in the last multiple iterations as an outline index sequence, and the expression of the iterative stability index is: wherein, represents the iteration stability index of the i-th iteration; represents the iteration stability index of the i-th iteration; represents the semantic fidelity index of the i-th iteration; represents the average semantic fidelity index of all hierarchical nodes in the i-th iteration; represents the average semantic fidelity index of all hierarchical nodes in the i-th iteration; represents the average semantic fidelity index of all hierarchical nodes in the i-th iteration; represents the average semantic fidelity index of all hierarchical nodes in the i-th iteration; represents the average semantic fidelity index of all hierarchical nodes in the i-th iteration; represents the number of data in the outline index sequence; represents the summation count variable; represents a preset parameter.

7. The method of claim 1, wherein the method further comprises: The method for processing the to-be-processed document by using the text outline generation model comprises: processing the to-be-processed document by using the text outline generation model to generate a first-level outline, taking the generated first-level outline as a context, and generating deep subdivision items in combination with the temperature parameter adjusted by the iterative stability index; calculating the iterative stability index in real time after each iteration generation, stopping the iteration when the convergence requirement is reached, and outputting the final text outline with clear levels and semantic coherence.

8. The method of claim 1, wherein the method is based on artificial intelligence. The preprocessing method of the original corpus is: removing web noise by using regular matching method; structurally splitting long texts by using a segmentation function, identifying and retaining original chapter numbers and hierarchical titles; and calling an NLP word segmentation tool for Chinese word segmentation, part-of-speech tagging, and stop word filtering.

9. The method of claim 1, wherein the method further comprises: The method for obtaining the original corpus is: collecting to-be-processed long text data through an enterprise internal knowledge base API, and obtaining the original corpus through keyword retrieval and URL deep traversal.

10. An artificial intelligence-based automatic text outline generation system, characterized by, The method comprises: A processor and a memory, the memory storing computer program instructions which, when executed by the processor, implement an artificial intelligence-based text outline automatic generation method according to any one of claims 1-9.

Images