Intelligent creation method based on NLP intention recognition

Abstract

The present invention relates to the field of artificial intelligence and natural language processing, and discloses an intelligent creation method based on NLP intention recognition, including: obtaining a number of original materials, performing a first preprocessing on the several original materials to obtain a number of knowledge fragments based on different slicing methods, and constructing a knowledge graph based on the several knowledge fragments; analyzing the written creation catalog and the selected or input creation content sentences, so as to obtain the creation intention; screening and matching the corresponding first entity, first association relationship and attribute from the knowledge graph according to the creation intention, and then matching the corresponding knowledge fragment according to the found first entity, first association relationship and attribute; reorganizing the found several knowledge fragments according to the preset constraints to obtain a complete article. The present invention can automatically identify the user intention and generate content that meets the needs, thereby improving the efficiency and accuracy of intelligent creation.

Description

Technical Field

[0001] The present invention relates to the fields of artificial intelligence and natural language processing, and in particular to an intelligent creation method based on NLP intent recognition. Background Art

[0002] When writing large documents, such as a proposal report or a feasibility study report, the writer sets the main framework of the document (creative directory) and gives each creative directory a central theme in one sentence. The auxiliary creative system can automatically help retrieve appropriate content, supplement the remaining text, and complete the basic content assembly.

[0003] With the continuous development of artificial intelligence and natural language processing technology, more and more applications require intelligent processing of text to help achieve intelligent creation. However, most of the current intelligent creation is designed based on rule-based methods, which lacks flexibility and scalability, or the creative materials recommended by the system have a low match with the user's intention and cannot adapt to the needs of different fields and scenarios. Therefore, an intelligent creation technology based on intent recognition is needed to solve these problems. Summary of the invention

[0004] The present invention provides an intelligent creation method based on NLP intention recognition to solve the above problems.

[0005] The present invention is achieved through the following technical solutions:

[0006] An intelligent creation method based on NLP intent recognition, comprising:

[0007] S1. Acquire a number of original materials, perform a first preprocessing on the original materials to obtain a number of knowledge fragments based on different slicing methods, and construct a knowledge graph based on the knowledge fragments;

[0008] S2. Analyze the written creative catalog and the selected or input creative content sentences to obtain the creative intention, wherein each level of creative catalog corresponds to a creative content sentence;

[0009] S3, screening and matching the corresponding first entity, first association relationship and attribute from the knowledge graph according to the creative intention, and then matching the corresponding knowledge fragment according to the found first entity, first association relationship and attribute;

[0010] S4. Reorganize the found knowledge fragments according to the preset constraints to obtain a complete article.

[0011] As an optimization, the specific process of performing the first preprocessing on the plurality of original materials in S1 to obtain a plurality of knowledge fragments based on different slicing methods is as follows:

[0012] S1.1. Using full-text extraction technology to extract text content of original materials of various file types, and performing word segmentation, part-of-speech tagging and word sense disambiguation processing on the text content of several of the original materials to obtain several first words;

[0013] S1.2, using the BERT model to identify the first entity and the first association relationship between the first attribute of the first entity and the first words processed by S1.1;

[0014] S1.3, according to the identified first association relationship between the first entity and the first attribute, the first entity having the first association relationship and the corresponding first attribute and the first association relationship between the first entity and the corresponding first attribute are extracted using the BERT model to obtain a corresponding extraction result;

[0015] S1.4. Slice and process the text contents corresponding to the extraction results into minimum units according to classification, entity and association relationship, and divide the text contents of the original material into different knowledge fragments.

[0016] As an optimization, the specific process of S1.4 is:

[0017] S1.4.1. Classify the text contents corresponding to the extraction results according to a preset analysis system, and record and store the text contents and corresponding categories as knowledge fragments;

[0018] S1.4.2. For each extraction result, split the first entity in the extraction result, store the text content corresponding to the extraction result as a knowledge fragment according to the key value, and classify the first entity;

[0019] S1.4.3, for each extraction result, split the first association relationship between the first entity of the extraction result and the corresponding first attribute, store the text content corresponding to the extraction result as a knowledge fragment according to the key value, and classify the first association relationship according to a preset classification system;

[0020] S1.4.4. For each extraction result, the first attribute of the extraction result is split, the text content corresponding to the extraction result is stored as a knowledge fragment according to the key value, and the first attribute is classified.

[0021] As an optimization, in S1, the specific process of constructing a knowledge graph based on several of the above-mentioned knowledge fragments is as follows:

[0022] S1.5, storing and representing several knowledge fragments according to knowledge graphs, key values, and vectors;

[0023] S1.6, using the knowledge graph embedding model KGE to embed the first entity and the first association relationship in the knowledge fragment into the vector space, store the graph structure information, and complete the knowledge completion;

[0024] S1.7, performing forward reasoning and reverse reasoning according to the first entity, the first association relationship and the first attribute to obtain a reasoning result, and performing knowledge representation on the reasoning result;

[0025] S1.8, repeat S1.7 to continuously improve the reasoning results, and improve the accuracy of knowledge reasoning with the help of SLM and NTN;

[0026] S1.9. Merge the knowledge fragments of the same first entity, first association relationship and first attribute from several original materials to complete the alignment of the first entity, the first association relationship and the first attribute, and finally complete the construction of the knowledge graph.

[0027] As an optimization, the specific process of S2 is:

[0028] S2.1. Merge the information of the creation directory in order and level, and record the directory information and feature vector information of each level in the creation directory;

[0029] S2.2. Using the BERT model, identify and extract the second entity, the second association relationship, and the second attribute of the merged creation directory information;

[0030] S2.3, analyzing the extracted second entity, second association relationship, and second attribute to obtain key information in the creation catalog, and performing structured storage and representation of the key information;

[0031] S2.4. Using the BERT model, complete contextual semantic analysis of the creative catalog, thereby identifying and calibrating the extracted key information;

[0032] S2.5. Analyze all the key information and all the creative content sentences obtained by merging in order and level, encode and decode them using RNN and Word2Vec, and generate first text information with a similarity to the key information higher than a threshold, the first text information being the overall creative intention;

[0033] S2.6. The text content obtained by merging the directories at each level and the creative content sentences corresponding to the directories at that level in each creative directory is respectively encoded and decoded using RNN and Word2Vec, and second text information having a similarity higher than a threshold with the text content obtained by merging the directories at each level and the corresponding creative content sentences is respectively generated. The second text information is the local creative intention corresponding to the text content obtained by merging the directories at that level and the corresponding creative content sentences;

[0034] S2.7. Identify the feature vector of the overall creative intention and classify the feature vector to finally obtain the classified overall creative intention.

[0035] As an optimization, the specific process of S3 is:

[0036] S3.1. Extracting a feature vector of the overall creative intent using deep learning technology;

[0037] S3.2, performing text familiarity analysis based on the feature vector of the overall creative intention and the knowledge fragments in the knowledge graph, preliminarily extracting knowledge fragments whose similarity is not less than a threshold, thereby recommending several initial knowledge fragments whose similarity with the feature vector of the overall creative intention is higher than a set threshold;

[0038] S3.3. Use deep learning technology to extract feature vectors of the text content in several of the initial knowledge fragments, and perform semantic analysis, timing analysis, and classification analysis on the recommended initial knowledge fragments and local creative intentions. By calculating the text familiarity of the initial knowledge fragments and local creative intentions, extract knowledge fragments with a similarity not less than a threshold as knowledge fragments suitable for the overall creative intent, local creative intent, directory framework, and selected or input creative content. This knowledge fragment is the recommended knowledge fragment.

[0039] As an optimization, the specific steps of S4 are:

[0040] S4.1. Perform word segmentation, entity recognition and word vector processing on the recommended knowledge fragments and local creative intentions, respectively, to obtain a number of first word vectors of the recommended knowledge fragments and a number of second word vectors of the local creative intentions;

[0041] S4.2, respectively input the first word vector and the second word vector into the encoder of the Transformer model for feature extraction and conversion, calculate the weight between each first word vector, the second word vector and all word vectors except the word vector, and add the weights between the first word vector and all word vectors except the word vector as the output of the word vector;

[0042] S4.3. According to the weights of the first word vector and the second word vector, RNN is used to analyze the dependency relationship, sequence, subject-verb-object semantics, and active-passive semantics of the second word vector of the local creative intention and the first word vector of the recommended knowledge fragment, thereby completing the syntactic rationality analysis and sorting of the local creative intention and the recommended knowledge fragment, and forming a new local creative intention and recommended knowledge fragment with reasonable syntax;

[0043] S4.4. Reorganize and sort the selected or input creative content sentences and recommended knowledge fragments, and perform multi-paragraph semantic analysis: combine the selected or input creative content sentences and the recommended and screened knowledge fragments, sort them according to the creative catalog frame hypothesis, reorganize and sort the paragraphs in order, perform context semantic analysis, and conduct consistency analysis on the overall theme. When the framework of the creative catalog, the selected or input creative content sentences and the recommended knowledge fragment theme are consistent and the context semantics are smooth, determine the content and order of the recommended knowledge fragments, and finally form a creative article.

[0044] As an optimization, the specific process of S4.4 is:

[0045] S4.4.1. Set the hierarchical order of the creative directory content to the first order and select knowledge fragments for each creative directory content according to the first order, and each creative directory content is matched with a corresponding creative content sentence;

[0046] S4.4.2. For one of the creative catalog contents, rank the at least one recommended knowledge fragment corresponding to the creative catalog content, and let i=1;

[0047] S4.4.3, select the i-th recommended knowledge fragment;

[0048] S4.4.4, determine whether the knowledge fragment meets the semantic requirements of the creation catalog and whether the theme is consistent. If so, temporarily retain the i-th recommended knowledge fragment and jump to S4.4.5. Otherwise, discard the i-th recommended knowledge fragment and jump to S4.4.7;

[0049] S4.4.5, determine whether there is a knowledge segment adjacent to the previous segment of the knowledge segment, if so, jump to S4.4.6, otherwise, retain the i-th recommended knowledge segment as the knowledge segment of the creation catalog content, and then jump to S4.4.7;

[0050] S4.4.6, calculate the similarity between the first word vector of the knowledge fragment and the third word vector of the knowledge fragment adjacent to the previous segment of the knowledge fragment, and determine whether the similarity between the first word vector of the knowledge fragment and the third word vector of the knowledge fragment adjacent to the previous segment of the knowledge fragment is not less than the set threshold. If so, jump to S4.4.9, otherwise, jump to S4.4.7;

[0051] S4.4.7, determine whether i is greater than N, where N is the number of recommended knowledge fragments. If so, jump to S4.4.8, otherwise, set i=i+1, and then jump to S4.4.3;

[0052] S4.4.8. Determine whether the first sequence has been traversed. If so, end. Otherwise, jump to the next creation directory content according to the first sequence, and then jump to S4.4.2, so as to match the knowledge fragment of the next creation directory content;

[0053] S4.4.9. Determine whether the knowledge fragment has any adjacent M knowledge fragments, where M is a positive integer not less than 2. If so, perform semantic analysis on the knowledge fragment and the knowledge fragments adjacent to the M knowledge fragments before it. If the knowledge fragment and the knowledge fragments adjacent to the M knowledge fragment belong to the same semantics, retain the recommended knowledge fragment and jump to S4.4.7. Otherwise, discard the recommended knowledge fragment and jump to S4.4.7.

[0054] As an optimization, if new creation directory content and corresponding creation content statements are added to the original creation directory content, the following judgment is made:

[0055] A1. For the new creation catalog content, sort at least one recommended knowledge fragment corresponding to the new creation catalog content, and let i=1;

[0056] A2, select the i-th recommended knowledge fragment;

[0057] A3, judging whether the knowledge fragment meets the semantic requirements of the creation catalog and whether the theme is consistent. If so, temporarily retain the i-th recommended knowledge fragment and then jump to A4. Otherwise, discard the i-th recommended knowledge fragment and then jump to A5.

[0058] A4, calculate the similarity between the first word vector of the knowledge segment and the third word vector of the knowledge segment adjacent to the previous segment of the knowledge segment and the fourth word vector of the knowledge segment adjacent to the next segment of the knowledge segment. If the similarity between the first word vector and the third word vector and the similarity between the first word vector and the fourth word vector are both greater than the set threshold, jump to A6, otherwise, jump to A5;

[0059] A5, determine whether i is greater than N, where N is the number of recommended knowledge fragments. If so, end. Otherwise, set i=i+1 and jump to A2;

[0060] A6, determine whether the knowledge fragment has the previous M knowledge fragments, where M is a positive integer not less than 2. If so, jump to A7, otherwise, jump to A8;

[0061] A7, perform semantic analysis on the knowledge fragment and the previous M knowledge fragments of the knowledge fragment respectively. If the knowledge fragment and the previous M knowledge fragments of the knowledge fragment have the same semantics, jump to A8; otherwise, discard the recommended knowledge fragment, set i=i+1, and then jump to A2;

[0062] A8, determine whether the knowledge fragment has Q knowledge fragments after it, where Q is a positive integer not less than 2. If so, jump to A9, otherwise, keep the recommended knowledge fragment;

[0063] A9. Perform semantic analysis on the knowledge fragment and the knowledge fragments in the Q segments after the knowledge fragment. If the knowledge fragment and the knowledge fragments in the Q segments after the knowledge fragment have the same semantics, retain the recommended knowledge fragment. Otherwise, discard the recommended knowledge fragment, set i=i+1, and then jump to A2.

[0064] As an optimization, in A4, if the new creative directory content does not have the previous level of creative directory content, the similarity between the first word vector and the fourth word vector is greater than the set threshold, that is, jump to A6, otherwise, jump to A5; if the new creative directory content does not have the next level of creative directory content, the similarity between the first word vector and the third word vector is greater than the set threshold, that is, jump to A6, otherwise, jump to A5.

[0065] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0066] The present invention is based on deep learning and natural language processing technology. Through the study and analysis of a large number of corpora, it can realize the accurate recognition and generation of user intentions, extract knowledge from texts, analyze creative intentions, automatically recommend creative materials, and complete copywriting.

[0067] The present invention can automatically identify user intentions and generate content that meets the needs, improving the efficiency and accuracy of intelligent creation. At the same time, based on deep learning and natural language processing technology, it has strong flexibility and scalability and can adapt to the needs of different fields and scenarios. In addition, the present invention can effectively solve the problems and shortcomings of the prior art and improve the level and quality of intelligent creation technology.

[0068] The present invention can automatically identify user intentions and generate content that meets the needs, thereby improving the efficiency and accuracy of intelligent creation.

[0069] The present invention is based on deep learning and natural language processing technology, has strong flexibility and scalability, and can adapt to the needs of different fields and scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] The drawings described herein are used to provide a further understanding of the embodiments of the present invention, constitute a part of this application, and do not constitute a limitation of the embodiments of the present invention. In the drawings:

[0071] Figure 1 A process of an intelligent creation method based on NLP intention recognition of the present invention;

[0072] Figure 2 This is an interface diagram of a system made by an intelligent creation method based on NLP intention recognition of the present invention;

[0073] Figure 3 for Figure 2 The enlarged diagram of the bottom interface is the interface diagram of the creation directory;

[0074] Figure 4 for Figure 2 The enlarged diagram of the interface in the middle is a diagram of the paragraph interface formed after selecting a knowledge fragment;

[0075] Figure 5 for Figure 2 The enlarged diagram of the top interface is a diagram of the interface of the recommended knowledge fragments. DETAILED DESCRIPTION

[0076] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments and drawings. The exemplary embodiments of the present invention and their description are only used to explain the present invention and are not intended to limit the present invention.

[0077] The main process steps of the present invention are as follows Figure 1 As shown, including:

[0078] Parse, extract and slice original data (raw materials) to form knowledge materials (knowledge fragments);

[0079] Analyze the creative catalog and input statements to identify creative intent;

[0080] Knowledge fragment screening and matching;

[0081] Reorganize and sort knowledge fragments and article paragraphs to form chapters;

[0082] Next, each step is described in detail.

[0083] This embodiment 1 provides an intelligent creation method based on NLP intention recognition, such as Figure 1 As shown, including:

[0084] S1. Acquire a number of original materials, perform a first preprocessing on the original materials to obtain a number of knowledge fragments based on different slicing methods, and construct a knowledge graph based on the knowledge fragments;

[0085] In S1, the first preprocessing of the raw materials to obtain the knowledge fragments based on different slicing methods is specifically to use multiple models to extract knowledge, split knowledge, and slice knowledge from the raw materials collected, uploaded or imported into the system to form knowledge fragments. The specific steps are as follows:

[0086] S1.1. Using full-text extraction technology to extract text content of original materials of various file types, and performing word segmentation, part-of-speech tagging and word sense disambiguation processing on the text content of several of the original materials to obtain several first words;

[0087] That is, this step is to perform full-text extraction: use full-text extraction technology to extract the file content of various file types, and then perform word segmentation, part-of-speech tagging, word sense disambiguation and other processing to provide a basis for subsequent knowledge processing and processing.

[0088] S1.2. Use the BERT model to identify the first entity and the first association relationship between the first attribute of the first entity for the first words processed by S1.1; here, it is mainly to identify specific entities and attributes in the text content, such as names of people, places, organizations, dates, etc.

[0089] S1.3, according to the identified first association relationship between the first entity and the first attribute, the first entity having the first association relationship and the corresponding first attribute and the first association relationship between the first entity and the corresponding first attribute are extracted using the BERT model to obtain a corresponding extraction result;

[0090] The main purpose here is to identify and extract the associations between entities. The BERT model can be used to identify and extract the relationships.

[0091] S1.4, the text content corresponding to the extraction results is sliced ​​into the smallest unit according to the classification, entity, and association relationship, and the text content of the original material is divided into different knowledge fragments. Through this step, the subsequent matching and recommendation of the corresponding material can be more accurate.

[0092] The knowledge fragment here is a piece of knowledge text content. It can be a sentence or a paragraph that describes an object, a person, a scene, or expresses an idea, etc.

[0093] The specific process of S1.4 is as follows:

[0094] S1.4.1. Classify the text contents corresponding to the extraction results according to a preset analysis system, and record and store the text contents and corresponding categories as knowledge fragments;

[0095] The preset analysis system here specifically refers to analyzing the content of knowledge fragments according to different dimensions, such as:

[0096] According to the type of description content, it can be divided into: writing about people, writing about events, writing about objects, writing about moods, etc.

[0097] According to the subject: disease, physical phenomenon, chemical phenomenon, biology, natural phenomenon, etc.;

[0098] According to the industry: transportation, education, energy, health care, etc.

[0099] S1.4.2. For each extraction result, split the first entity in the extraction result (split by entity extraction, which can be extracted by BERT model), store the text content corresponding to the extraction result as a knowledge fragment according to the key value (the key value here is the first entity), and classify the first entity;

[0100] The programming language obtains the key value information of the text content;

[0101] This is a way of storing knowledge. The key is a unique identifier and the value is the knowledge content associated with the key. This storage method is simple and direct, and facilitates fast retrieval. The value in the key-value pair can be text, number, list, or other data structure, depending on the content and representation requirements of the knowledge fragment.

[0102] Similarly, the classification of the first entity here also refers to different dimensions, such as:

[0103] According to gender, it is divided into: male and female;

[0104] According to the type of animal, it is divided into: humans, cats, dogs, pigs, cows, etc.;

[0105] According to the name classification, it is divided into: Zhang San, Li Si, Wang Wu, etc.

[0106] For example, if an extraction result has two first entities, namely name and specific date of birth, and their association is "date of birth", for example, if the extraction result is ("Zhang San", "date of birth", "April 26, 1988"), then "Zhang San" is the first entity, "date of birth" is the first association, and "April 26, 1988" is the first attribute.

[0107] S1.4.3. For each extraction result, the first association relationship between the first entity of the extraction result and the corresponding first attribute is split (the entities and relationships are extracted using the BERT model and stored separately), and the text content corresponding to the extraction result is stored as a knowledge fragment according to the key value (the key value at this time is the first association relationship), and the first association relationship is classified according to a certain classification system.

[0108] Similarly, the classification of the first association relationship of the first entity here also refers to classification according to different dimensions, such as:

[0109] Classify by date: date of birth, elementary school graduation date, high school graduation date, marriage date, etc.

[0110] Classified by location: place of birth, place of graduation, place of work, etc.

[0111] S1.4.4. For each extraction result, the first attribute of the extraction result is split, and the text content corresponding to the extraction result is stored as a knowledge fragment according to the key value (the key value in this case is the first attribute), and the first attribute is classified.

[0112] By slicing the knowledge fragments in the above dimensions, the text content can be sliced ​​more finely, so that the same knowledge fragment has different classifications, which can make the subsequent matching and recommendation of corresponding materials more accurate. Next, build the knowledge graph: for the knowledge fragments, complete the knowledge graph processing through rule reasoning, knowledge reasoning, and semantic annotation. Through multi-step path reasoning, node relationship, edge relationship rule definition, attribute calculation rules, and path condition rules, the reasoning of entity attributes, relationships, and labels is realized to complete the construction of the knowledge graph and provide support for subsequent knowledge fragment matching. The specific process is as follows:

[0113] The knowledge fragments are stored and represented according to knowledge graphs, key values, and vectors.

[0114] KGE is used to embed entities and relationships into vector space, store graph structure information, and complete knowledge completion.

[0115] According to the entity, object, and relationship knowledge units, forward reasoning, reverse reasoning, and knowledge representation are performed. This process is continuously circulated and revised to continuously improve the reasoning results. The accuracy of knowledge reasoning can be improved with the help of SLM and NTN.

[0116] Finally, it is necessary to merge the knowledge units of the same entity or object and relationship from multiple sources to align entities, relationships, and attributes and complete the construction of the graph.

[0117] That is, in S1, the specific process of constructing a knowledge graph based on several of the above-mentioned knowledge fragments is:

[0118] S1.5, storing and representing several knowledge fragments according to knowledge graphs, key values, and vectors;

[0119] There are several different approaches to knowledge management and representation:

[0120] Knowledge graph: A knowledge graph is a graphical structure that uses nodes (representing entities or concepts) and edges (representing relationships) to represent knowledge. It is often used to display complex relationships between entities, making it easier for people to query, reason, and visualize.

[0121] Key-value: A key-value pair is a simple data structure consisting of a key and a value. In knowledge representation, the key can be an identifier of a concept or entity, while the value is the corresponding attribute or description.

[0122] Vector: Vectors are often used to represent semantic information of entities or concepts. Vectors are usually represented in the form of numerical arrays, where each numerical value corresponds to a dimension in the vector space. By calculating the similarity between vectors (such as cosine similarity, Euclidean distance, etc.), the semantic relationship between entities or concepts can be evaluated.

[0123] The most basic method is to store and represent data in vector form. To increase the accuracy of the analysis, you can also use all three methods together.

[0124] S1.6, using the knowledge graph embedding model KGE to embed the first entity and the first association relationship in the knowledge fragment into the vector space, store the graph structure information, and complete the knowledge completion;

[0125] S1.7, performing forward reasoning and reverse reasoning according to the first entity, the first association relationship, and the first attribute to obtain a reasoning result, and performing knowledge representation on the reasoning result;

[0126] That is, knowledge is represented according to entities, relationships, and attributes.

[0127] For example, Zhang San's occupation is teacher

[0128] Zhang San: Entity

[0129] Occupation: Relationship

[0130] Teacher: Attributes.

[0131] Forward reasoning usually starts from one or more known facts and derives new conclusions by applying a set of rules or models. In the context of KGE, forward reasoning may involve the following steps:

[0132] Known facts: given one or more entities, relations and their embedding representations.

[0133] Rules or models: Use pre-trained KGE models (such as TransE, RotatE, etc.) to capture the semantic information between entities and relations.

[0134] Reasoning: Based on known entities and relations, use the KGE model to predict new triples (i.e., predict the possible relationship between two given entities).

[0135] Backward reasoning (also called abductive reasoning) starts from the target or hypothesized conclusion and searches for known facts or rules that support the conclusion. In the context of KGE, backward reasoning may involve the following steps:

[0136] Objective Conclusion: Given a triple (i.e., a hypothesized entity-relation-entity combination) that you want to verify.

[0137] Query: Use the KGE model to evaluate the plausibility or likelihood of the triple. This usually involves calculating the similarity or distance between the embedding representations of the entities and relations in the triple.

[0138] Explanation: If the probability of the target triple is high, then further exploration of known facts or rules that support this conclusion can be done. This may require querying other triples in the KGE or leveraging other forms of explanatory techniques.

[0139] S1.8, repeat S1.7 to continuously improve the reasoning results, and use SLM and NTN to improve the accuracy of knowledge reasoning until all text contents are analyzed and knowledge fragments are formed respectively;

[0140] The specific process of improving the accuracy of knowledge reasoning with the help of SLM (pre-trained language models, such as BERT, GPT, etc.) and NTN (neural tensor network model) is as follows:

[0141] First, we use the powerful capabilities of SLM for semantic representation and contextual understanding. SLM has been pre-trained on a large amount of text data and has learned a wealth of language knowledge and contextual information. Through SLM, we can convert entities and relations in the knowledge graph into vector representations. These vector representations are rich in semantic information and help the subsequent reasoning process. Next, we use NTN for relational reasoning. NTN is a model based on neural tensor networks, which is suitable for relational reasoning between two entities. By introducing tensors as parameters, NTN can capture high-order relationships between entities and relations. Based on the vector representation provided by SLM, NTN can further explore the potential relationships between entities, thereby improving the accuracy of reasoning.

[0142] S1.9. Merge the knowledge units of the same first entity, first association relationship and first attribute from several original materials to complete the alignment of the first entity, the first association relationship and the first attribute, and finally complete the construction of the knowledge graph.

[0143] S2. Analyze the written content of the creative directory and the selected or input creative content sentences to obtain the creative intent, where each level of directory corresponds to a creative content sentence; mainly use the "input content sentence" + creative directory content to combine and identify the creative intent. The input content is the main focus, and the creative directory content is supplementary.

[0144] The specific process of S2 is:

[0145] S2.1. Merge the information of the creation directory in order and level, and record the directory information and feature vector information of each level in the creation directory;

[0146] Specifically, they are combined in the order of the creation directories from front to back (including sub-levels).

[0147] For example, the directory is:

[0148] 1. Construction Background

[0149] 1.1 Background of the Development of Digital China

[0150] 1.2 Current Status of Digital Transformation in Units

[0151] 2. Construction Objectives

[0152] 2.1 Build a knowledge management system for the entire unit

[0153] 2.2 Construction Department Knowledge Base

[0154] Combined together:

[0155] 1. Construction background, 1.1 Background of the development of Digital China, 1.2 Current status of the unit’s digital transformation; 2 Construction goals, 2.1 Construction of a knowledge management system for the entire unit, 2.2 Construction of a departmental knowledge base.

[0156] The directory hierarchy information is the directory name added by the system, and the system can automatically identify the content of the hierarchy. The feature vector information is the information converted from the directory information content, which is generally converted using the TF-IDF method.

[0157] The purpose of recording the directory information and feature vector information of each level in the creation directory is to analyze the content of the creation directory + the input text content together when identifying the creation intention.

[0158] Record the directory information and feature vector information of each level in the creation directory, that is, record the feature information such as word frequency, syntactic features, semantic features, directory hierarchical relationship, etc. Record it in order to analyze the directory features and semantics, and analyze the writing intention in combination with the input text content.

[0159] S2.2. Using the BERT model, identify and extract the second entity, the second association relationship, and the second attribute of the merged creation directory information;

[0160] S2.3, analyzing the extracted second entity, second association relationship, and second attribute to obtain key information in the creation directory (obtained through directory content and hierarchical relationships between directories), and performing structured storage and representation of the key information;

[0161] Taking the above case as an example, the key information can be:

[0162] The construction background mainly describes the development background of Digital China and the digital transformation background of the unit;

[0163] The construction goal is mainly to build a knowledge management system for construction units and a departmental knowledge base.

[0164] S2.4. Use the BERT model to complete contextual semantic analysis of the creative catalog (i.e., the content of the previous and next chapters of the creative catalog. In the above case, “1. Construction background, 1.1 Digital China development background, 1.2 Current status of digital transformation of the unit; 2 Construction goals, 2.1 Construction of a knowledge management system for the entire unit, 2.2 Construction of a department knowledge base”), so as to identify and calibrate the extracted key information; the BERT model outputs knowledge triples of the creative catalog content, which can perform semantic analysis of the catalog content. Knowledge triples are generally entities (subject objects), relationships, and attributes (object objects).

[0165] S2.5. Analyze all the key information and all the creative content sentences obtained by merging in order and level, encode and decode them using RNN and Word2Vec, and generate first text information with a similarity to the key information higher than a threshold, the first text information being the overall creative intention;

[0166] That is, all the directories (outlines) are analyzed together in order and hierarchical combination + the text content corresponding to the input of the current directory to obtain the creative intention of the entire article.

[0167] S2.6. The text content obtained by merging the directories at each level and the creative content sentences corresponding to the directories at that level in each creative directory is respectively encoded and decoded using RNN and Word2Vec, and second text information having a similarity higher than a threshold with the text content obtained by merging the directories at each level and the corresponding creative content sentences is respectively generated. The second text information is the local creative intention corresponding to the text content obtained by merging the directories at that level and the corresponding creative content sentences;

[0168] The current chapter directory (i.e. the creation directory at the current level) + the entered text content (creation content statements) are analyzed together to obtain the local creation intention of the current chapter.

[0169] Analyze the text content in the creative catalog and the selected or input creative content sentences, use RNN and Word2Vec to capture the word order information and semantic information of the feature vector in the sentence (creative catalog + input sentence, combined analysis), so as to understand the meaning of the entire sentence, and encode and decode the input creative catalog or an already entered paragraph using Word2Vec and RNN to generate text information with a similarity higher than a threshold with the creative catalog combined text (the content of all creative catalogs combined in sequence, which is equivalent to key information) and the selected or input creative content sentences, which is the overall creative intention; use RNN and Word2Vec to analyze the text semantics, and use simHash and minHash models to generate text information with a similarity threshold of more than 85% with the catalog + input sentence content, which is the local creative theme.

[0170] S2.7, identify the feature vector of the overall creative intent (obtained by the Word2Vec model in S2.5 above) and classify the feature vector (using the RNN model for classification), and finally obtain the classified creative intent. The classification here refers to the subject category, and RNN can be used to classify text data by subject. For example, given a news article, the model can determine which subject category it belongs to, such as politics, economy, sports, entertainment, etc.

[0171] When screening knowledge fragments later, first screen on a large scale (screen the entire article's creative intent), and then filter a second time to screen more appropriate knowledge fragments for the current chapter. This can greatly reduce the screening time.

[0172] In summary, by capturing the temporal dependencies and semantic information in the sentences, as well as the keywords and phrases extracted from the text, we can derive the user's intention and central idea of ​​the work to be created. We can also identify and classify the feature vectors of the creative intention to form a classification of the creative intention.

[0173] S3, screening and matching the corresponding first entity, first association relationship and first attribute from the knowledge graph according to the creative intention, and then matching the corresponding knowledge fragment according to the found first entity, first association relationship and first attribute;

[0174] The specific process of S3 is:

[0175] S3.1. Extracting a feature vector of the overall creative intent using deep learning technology;

[0176] S3.2. Perform text familiarity analysis based on the feature vector of the overall creative intent and the knowledge fragments in the knowledge graph using the simhash and minhash models, and preliminarily extract knowledge fragments whose similarity is not less than a threshold, so as to recommend several initial knowledge fragments whose similarity with the feature vector of the creative intent is higher than the set threshold; specifically, it mainly extracts features such as the theme, keywords, and first association relationship of the text content, and recommends relevant content to users based on the similarity of these features.

[0177] The above process is content recommendation: by analyzing the characteristics of the knowledge fragment text content and the content characteristics of the creative intention, relevant content is recommended to users. Specifically, it mainly extracts the text's theme, keywords, associations and other features, and recommends relevant content to users based on the similarity of these features.

[0178] S3.3. Use deep learning technology to extract feature vectors of the text content in several of the initial knowledge fragments, and perform semantic analysis, timing analysis, and classification analysis on the recommended initial knowledge fragments and local creative intentions. By calculating the text familiarity of the initial knowledge fragments and local creative intentions, extract knowledge fragments with a similarity not less than a threshold as knowledge fragments suitable for the overall creative intent, local creative intent, directory framework, and selected or input creative content. This knowledge fragment is the recommended knowledge fragment.

[0179] Specifically, Word2Vec and RNN are used to analyze the semantics, timing, and classification of the recommended initial knowledge fragments and local creative intentions:

[0180] Word2Vec is a technique for generating word vectors that converts each word into a fixed-length vector so that semantically similar words are close in position in the vector space. By comparing the cosine similarity or Euclidean distance between word vectors of different words, the semantic similarity between them can be evaluated.

[0181] RNN is a special neural network structure that can process sequence data and save previous state information during the processing, making it suitable for time series analysis and classification analysis.

[0182] The above process is knowledge fragment matching: using deep learning technology to extract the features of the creative intent and the text in the knowledge fragment, and to analyze the semantics, time series analysis, and classification analysis of the recommended knowledge fragment and the creative content. By calculating the text familiarity, the knowledge fragments with higher familiarity are extracted as the knowledge fragments suitable for the creative intent, catalog framework, and the input creative content.

[0183] S4. Reorganize the found knowledge fragments according to the preset constraints to obtain a complete article.

[0184] The creative catalog framework, input content, and recommended knowledge fragments are jointly analyzed, arranged, and combined to form a complete chapter.

[0185] The specific steps of S4 are:

[0186] S4.1. Perform word segmentation, entity recognition and word vector processing on the recommended knowledge fragments and local creative intentions (creation catalogs and selected or input creative content), and obtain a number of first word vectors of the recommended knowledge fragments and a number of second word vectors of the local creative intentions respectively;

[0187] S4.2, respectively input the first word vector and the second word vector into the encoder of the Transformer model for feature extraction and conversion, calculate the weight between each word vector (the first word vector and the second word vector) and all word vectors except the word vector, and add the weights between the first word vector and all word vectors except the word vector as the output of the word vector; judge the semantics and order of the words in the text according to the output result (the output word vector), and whether they are appropriate, and discard them if they are not appropriate;

[0188] The above process is lexical analysis. S4.3. According to the weight of the word vector, RNN is used to analyze the dependency relationship, front-back order, subject-predicate-object semantics, and subject-passive semantics of the first word vector of the local creative intention and the second word vector of the recommended knowledge fragment, so as to complete the syntactic rationality analysis and sorting of the local creative intention and the recommended knowledge fragment, and form a new local creative intention and recommended knowledge fragment with reasonable syntax (this is to judge whether the local creative intention and the recommended knowledge fragment are smooth and whether the sentence meaning is reasonable); it can be analyzed by part of speech and relationship (subject-predicate-object). The subject is usually a noun or a pronoun, the predicate is a verb, and the object can be a noun, a pronoun, or a verb phrase. By identifying the grammatical roles and dependencies of these words, it can be determined whether they constitute a subject-predicate-object structure.

[0189] This process is syntactic analysis. The content of the analysis here is the same as that of the creative intent analysis, but the goal and method are different. The previous one is to analyze the creative intent, and this one is to analyze whether the intent and the recommended fragment sequence are reasonable and whether the semantics are smooth.

[0190] S4.4. Reorganize and sort the selected or input creative content sentences and recommended knowledge fragments, and conduct multi-paragraph semantic analysis: combine the selected or input creative content sentences and recommended and screened knowledge fragments, sort them according to the creative catalog frame hypothesis, reorganize and sort the paragraphs in order, and conduct consistency analysis of context semantics and the theme of the whole article (the recommended knowledge fragments to be inserted must be compared with the context paragraphs respectively, and the similarity of each comparison must be greater than 85% to be used as knowledge fragments inserted into the article). When the framework of the creative catalog, the selected or input creative content sentences and the theme of the recommended knowledge fragments are consistent (semantically reasonable, high similarity (greater than 85%) are considered consistent), and the context semantics are smooth (semantically reasonable), the content and order of the recommended knowledge fragments are determined, and finally the creative article is formed. This process is paragraph analysis.

[0191] The specific process of S4.4 is as follows (i.e. the process of reorganization according to the preset constraints):

[0192] S4.4.1. Set the hierarchical order of the creative directory content to the first order and select knowledge fragments for each creative directory content according to the first order, and each creative directory content is matched with a corresponding creative content sentence;

[0193] like Figure 2 , 3 As shown, the creation directory has two first-level directories, namely project background and construction target. Therefore, the first order is project background → construction target.

[0194] If the directory is:

[0195] 1. Construction Background

[0196] 1.1 Background of the Development of Digital China

[0197] 1.2 Current Status of Digital Transformation in Units

[0198] 2. Construction Objectives

[0199] 2.1 Build a knowledge management system for the entire unit

[0200] 2.2 Construction Department Knowledge Base

[0201] So, the first order is: construction background → Digital China development background → current status of unit’s digital transformation → construction goals → construction of unit-wide knowledge management system → construction of department knowledge base.

[0202] S4.4.2. For one of the creative catalog contents, rank the at least one recommended knowledge fragment corresponding to the creative catalog content, and let i=1;

[0203] like Figure 4 As shown, it is the knowledge snippet recommended for the background item.

[0204] like Figure 5 As shown, there are at least 2 recommended knowledge fragments.

[0205] S4.4.3, select the i-th recommended knowledge segment as the knowledge segment of the creation catalog content temporarily;

[0206] S4.4.4.1. It is manually determined whether the knowledge fragment meets the semantic requirements of the creation catalog and whether the theme is consistent. If so, the i-th recommended knowledge fragment is temporarily retained and the process jumps to S4.4.5. Otherwise, the i-th recommended knowledge fragment is discarded and the process jumps to S4.4.7.

[0207] Starting from the first knowledge fragment, determine whether it meets the semantic requirements of the creation catalog. If so, temporarily retain the knowledge fragment, otherwise, directly discard the knowledge fragment. After retaining the knowledge fragment, determine the similarity between the knowledge fragment and the previous knowledge fragment (if there is a previous knowledge fragment) and whether the semantics of the previous M knowledge fragments (if there are previous M knowledge fragments) are consistent. Therefore, the following process is carried out.

[0208] S4.4.5, determine whether there is a knowledge segment adjacent to the previous segment of the knowledge segment, if so, jump to S4.4.6, otherwise, retain the i-th recommended knowledge segment as the knowledge segment of the creation catalog content, and then jump to S4.4.7;

[0209] S4.4.6, determine whether the similarity between the first word vector of the knowledge segment and the third word vector of the knowledge segment immediately before the knowledge segment is not less than the set threshold. If so, jump to S4.4.9; otherwise, jump to S4.4.7;

[0210] The third word vector here can be obtained through existing means (refer to the means of obtaining the first word vector and the second word vector). The third word vector refers to the word vector of the knowledge fragment. It is only to distinguish the word vector of this knowledge fragment from the previous knowledge fragment, so the first and third annotations are made.

[0211] S4.4.7, determine whether i is greater than N, where N is the number of recommended knowledge fragments. If so, jump to S4.4.8, otherwise, set i=i+1, and then jump to S4.4.3;

[0212] S4.4.8. Determine whether the first sequence has been traversed. If so, end. Otherwise, jump to the next creation directory content according to the first sequence, and then jump to S4.4.2, so as to match the knowledge fragment of the next creation directory content;

[0213] S4.4.9. Determine whether the knowledge fragment has any adjacent M knowledge fragments, where M is a positive integer not less than 2. If so, perform semantic analysis on the knowledge fragment and the knowledge fragments adjacent to the M knowledge fragments before it. If the knowledge fragment and the knowledge fragments adjacent to the M knowledge fragment belong to the same semantics, retain the recommended knowledge fragment and jump to S4.4.7. Otherwise, discard the recommended knowledge fragment and jump to S4.4.7.

[0214] The above process can be simply described as matching paragraphs starting from the first creative directory, and then selecting recommended knowledge fragments. If the recommended knowledge fragment conforms to the semantics of the creative directory (which can be judged manually), the knowledge fragment is retained, and then the next knowledge fragment is selected in reverse order for re-judgment.

[0215] When judging a certain knowledge fragment, the similarity between the knowledge fragment of the previous paragraph and the knowledge fragment will be compared, and then the semantic analysis will be performed on the knowledge fragment and the M knowledge fragments before the knowledge fragment, where M is R-1 and R is the number of paragraphs of the knowledge fragment. That is, if the knowledge fragment is the 10th paragraph, the similarity between the knowledge fragment of the 10th paragraph and the knowledge fragment of the 9th paragraph will be compared, and then the semantic analysis will be performed on the knowledge fragment of the 10th paragraph and the knowledge fragments of the 1st and 8th paragraphs.

[0216] Further introduction:

[0217] In S4.4.9, semantic analysis is performed on the knowledge fragment and the M knowledge fragments adjacent to it respectively. If the knowledge fragment and the M knowledge fragments adjacent to it belong to the same semantic essence, it is to judge whether the context semantics are smooth, which can be done through the Transformer model. If the recommended knowledge fragment is not suitable, then the recommended knowledge fragment is discarded and the next recommended knowledge fragment is looked at.

[0218] Since the above steps are performed from front to back according to the first order, there is no need to consider the semantics backward, only the context semantics forward. However, it is also possible to add a new creation directory and the corresponding creation content sentence to the creation directory that has already formed a good paragraph. Therefore, if a new creation directory is added, the following process is performed.

[0219] A1. For the new creation catalog content, sort at least one recommended knowledge fragment corresponding to the new creation catalog content, and let i=1;

[0220] A2, select the i-th recommended knowledge segment as the knowledge segment of the new creation catalog content temporarily;

[0221] A3: manually judge whether the knowledge fragment meets the semantic requirements of the creation catalog and whether the theme is consistent. If so, temporarily retain the i-th recommended knowledge fragment and jump to A4; otherwise, discard the i-th recommended knowledge fragment and jump to A5;

[0222] A4, calculate the similarity between the first word vector of the knowledge segment and the third word vector of the knowledge segment adjacent to the previous segment of the knowledge segment and the fourth word vector of the knowledge segment adjacent to the next segment of the knowledge segment. If the similarity between the first word vector and the third word vector and the similarity between the first word vector and the fourth word vector are both greater than the set threshold, jump to A6, otherwise, jump to A5;

[0223] A5, determine whether i is greater than N, where N is the number of recommended knowledge fragments. If so, end. Otherwise, set i=i+1 and jump to A2;

[0224] A6, determine whether the knowledge fragment has the previous M knowledge fragments, where M is a positive integer not less than 2. If so, jump to A7, otherwise, jump to A8;

[0225] A7, perform semantic analysis on the knowledge fragment and the previous M knowledge fragments of the knowledge fragment respectively. If the knowledge fragment and the previous M knowledge fragments of the knowledge fragment have the same semantics, jump to A8; otherwise, discard the recommended knowledge fragment, set i=i+1, and then jump to A2;

[0226] A8, determine whether the knowledge fragment has Q knowledge fragments after it, where Q is a positive integer not less than 2. If so, jump to A9, otherwise, keep the recommended knowledge fragment;

[0227] A9. Perform semantic analysis on the knowledge fragment and the knowledge fragments in the Q segments after the knowledge fragment. If the knowledge fragment and the knowledge fragments in the Q segments after the knowledge fragment have the same semantics, retain the recommended knowledge fragment. Otherwise, discard the recommended knowledge fragment, set i=i+1, and then jump to A2.

[0228] In the above process, a new creative directory and corresponding creative content sentences are added to the creative directory of the already formed article. The newly added part may be in the middle of the original creative directory, or it may be at the very beginning or the very end. Therefore, it is necessary to consider whether the context semantics of the whole text is normal. Therefore, in addition to forward semantic considerations, backward semantic considerations are also required.

[0229] Similarly, this can be done through the Transformer model. If the recommended knowledge fragment is inappropriate, then ignore the recommended knowledge fragment and look at the next recommended knowledge fragment.

[0230] The specific implementation methods described above further describe the purpose, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention. The specific implementation methods described above further describe the purpose, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An intelligent creation method based on NLP intention recognition, characterized in that: include: S1. Acquire a number of original materials, perform a first preprocessing on the original materials to obtain a number of knowledge fragments based on different slicing methods, and construct a knowledge graph based on the knowledge fragments; S2. Analyze the written creative catalog and the selected or input creative content sentences to obtain the creative intention, wherein each level of creative catalog corresponds to a creative content sentence; The specific process is: S2.

1. Merge the information of the creation directory in order and level, and record the directory information and feature vector information of each level in the creation directory; S2.

2. Using the BERT model, identify and extract the second entity, the second association relationship, and the second attribute of the merged creation directory information; S2.3, analyzing the extracted second entity, second association relationship, and second attribute to obtain key information in the creation catalog, and performing structured storage and representation of the key information; S2.

4. Using the BERT model, complete contextual semantic analysis of the creative catalog, thereby identifying and calibrating the extracted key information; S2.

5. Analyze all the key information and all the creative content sentences obtained by merging in order and level, encode and decode them using RNN and Word2Vec, and generate first text information with a similarity to the key information higher than a threshold, the first text information being the overall creative intention; S2.

6. The text content obtained by merging the directories at each level and the creative content sentences corresponding to the directories at that level in each creative directory is respectively encoded and decoded using RNN and Word2Vec, and second text information having a similarity higher than a threshold with the text content obtained by merging the directories at each level and the corresponding creative content sentences is respectively generated. The second text information is the local creative intention corresponding to the text content obtained by merging the directories at that level and the corresponding creative content sentences; S2.7, identifying the feature vector of the overall creative intention and classifying the feature vector, and finally obtaining the classified overall creative intention; S3, screening and matching the corresponding first entity, first association relationship and attribute from the knowledge graph according to the creative intention, and then matching the corresponding knowledge fragment according to the found first entity, first association relationship and attribute; S4, reorganize the found knowledge fragments according to the preset constraints to obtain a complete article; The specific steps are: S4.

1. Perform word segmentation, entity recognition and word vector processing on the recommended knowledge fragments and local creative intentions, respectively, to obtain a number of first word vectors of the recommended knowledge fragments and a number of second word vectors of the local creative intentions; S4.2, respectively input the first word vector and the second word vector into the encoder of the Transformer model for feature extraction and conversion, calculate the weight between each first word vector, the second word vector and all word vectors except the word vector, and add the weights between the first word vector and all word vectors except the word vector as the output of the word vector; S4.

3. According to the weights of the first word vector and the second word vector, RNN is used to analyze the dependency relationship, sequence, subject-verb-object semantics, and active-passive semantics of the second word vector of the local creative intention and the first word vector of the recommended knowledge fragment, thereby completing the syntactic rationality analysis and sorting of the local creative intention and the recommended knowledge fragment, and forming a new local creative intention and recommended knowledge fragment with reasonable syntax; S4.

4. Reorganize and sort the selected or input creative content sentences and recommended knowledge fragments, and perform multi-paragraph semantic analysis: combine the selected or input creative content sentences and the recommended and screened knowledge fragments, sort them according to the creative catalog frame hypothesis, reorganize and sort the paragraphs in order, perform context semantic analysis, and conduct consistency analysis on the overall theme. When the framework of the creative catalog, the selected or input creative content sentences and the recommended knowledge fragment theme are consistent and the context semantics are smooth, determine the content and order of the recommended knowledge fragments, and finally form a creative article.

2. According to the intelligent creation method based on NLP intention recognition according to claim 1, it is characterized in that: The specific process of performing the first preprocessing on the plurality of original materials in S1 to obtain a plurality of knowledge fragments based on different slicing methods is as follows: S1.

1. Using full-text extraction technology to extract text content of original materials of various file types, and performing word segmentation, part-of-speech tagging and word sense disambiguation processing on the text content of several of the original materials to obtain several first words; S1.2, using the BERT model to identify the first entity and the first association relationship between the first attribute of the first entity and the first words processed by S1.1; S1.3, according to the identified first association relationship between the first entity and the first attribute, the first entity having the first association relationship and the corresponding first attribute and the first association relationship between the first entity and the corresponding first attribute are extracted using the BERT model to obtain a corresponding extraction result; S1.

4. Slice and process the text contents corresponding to the extraction results into minimum units according to classification, entity and association relationship, and divide the text contents of the original material into different knowledge fragments.

3. According to claim 2, the intelligent creation method based on NLP intention recognition is characterized in that: The specific process of S1.4 is as follows: S1.4.

1. Classify the text contents corresponding to the extraction results according to a preset analysis system, and record and store the text contents and corresponding categories as knowledge fragments; S1.4.

2. For each extraction result, split the first entity in the extraction result, store the text content corresponding to the extraction result as a knowledge fragment according to the key value, and classify the first entity; S1.4.3, for each extraction result, split the first association relationship between the first entity of the extraction result and the corresponding first attribute, store the text content corresponding to the extraction result as a knowledge fragment according to the key value, and classify the first association relationship according to a preset classification system; S1.4.

4. For each extraction result, the first attribute of the extraction result is split, the text content corresponding to the extraction result is stored as a knowledge fragment according to the key value, and the first attribute is classified.

4. According to claim 3, the intelligent creation method based on NLP intention recognition is characterized in that: In S1, the specific process of constructing a knowledge graph based on several of the above-mentioned knowledge fragments is as follows: S1.5, storing and representing several knowledge fragments according to knowledge graphs, key values, and vectors; S1.6, using the knowledge graph embedding model KGE to embed the first entity and the first association relationship in the knowledge fragment into the vector space, store the graph structure information, and complete the knowledge completion; S1.7, performing forward reasoning and reverse reasoning according to the first entity, the first association relationship and the first attribute to obtain a reasoning result, and performing knowledge representation on the reasoning result; S1.8, repeat S1.7 to continuously improve the reasoning results, and improve the accuracy of knowledge reasoning with the help of SLM and NTN; S1.

9. Merge the knowledge fragments of the same first entity, first association relationship and first attribute from several original materials to complete the alignment of the first entity, the first association relationship and the first attribute, and finally complete the construction of the knowledge graph.

5. According to claim 1, the intelligent creation method based on NLP intention recognition is characterized in that: The specific process of S3 is: S3.

1. Extracting a feature vector of the overall creative intent using deep learning technology; S3.2, performing text familiarity analysis based on the feature vector of the overall creative intention and the knowledge fragments in the knowledge graph, preliminarily extracting knowledge fragments whose similarity is not less than a threshold, thereby recommending several initial knowledge fragments whose similarity with the feature vector of the overall creative intention is higher than a set threshold; S3.

3. Use deep learning technology to extract feature vectors of the text content in several of the initial knowledge fragments, and perform semantic analysis, timing analysis, and classification analysis on the recommended initial knowledge fragments and local creative intentions. By calculating the text familiarity of the initial knowledge fragments and local creative intentions, extract knowledge fragments with a similarity not less than a threshold as knowledge fragments suitable for the overall creative intent, local creative intent, directory framework, and selected or input creative content. This knowledge fragment is the recommended knowledge fragment.

6. According to claim 1, the intelligent creation method based on NLP intention recognition is characterized in that: The specific process of S4.4 is as follows: S4.4.

1. Set the hierarchical order of the creative directory content to the first order and select knowledge fragments for each creative directory content according to the first order, and each creative directory content is matched with a corresponding creative content sentence; S4.4.

2. For one of the creative catalog contents, rank the at least one recommended knowledge fragment corresponding to the creative catalog content, and let i=1; S4.4.3, select the i-th recommended knowledge fragment; S4.4.4, determine whether the knowledge fragment meets the semantic requirements of the creation catalog and whether the theme is consistent. If so, temporarily retain the i-th recommended knowledge fragment and jump to S4.4.

5. Otherwise, discard the i-th recommended knowledge fragment and jump to S4.4.7; S4.4.5, determine whether there is a knowledge segment adjacent to the previous segment of the knowledge segment, if so, jump to S4.4.6, otherwise, retain the i-th recommended knowledge segment as the knowledge segment of the creation catalog content, and then jump to S4.4.7; S4.4.6, calculate the similarity between the first word vector of the knowledge fragment and the third word vector of the knowledge fragment adjacent to the previous segment of the knowledge fragment, and determine whether the similarity between the first word vector of the knowledge fragment and the third word vector of the knowledge fragment adjacent to the previous segment of the knowledge fragment is not less than the set threshold. If so, jump to S4.4.9, otherwise, jump to S4.4.7; S4.4.7, determine whether i is greater than N, where N is the number of recommended knowledge fragments. If so, jump to S4.4.8, otherwise, set i=i+1, and then jump to S4.4.3; S4.4.

8. Determine whether the first sequence has been traversed. If so, end. Otherwise, jump to the next creation directory content according to the first sequence, and then jump to S4.4.2, so as to match the knowledge fragment of the next creation directory content; S4.4.

9. Determine whether the knowledge fragment has any adjacent M knowledge fragments, where M is a positive integer not less than 2. If so, perform semantic analysis on the knowledge fragment and the knowledge fragments adjacent to the M knowledge fragments before it. If the knowledge fragment and the knowledge fragments adjacent to the M knowledge fragment belong to the same semantics, retain the recommended knowledge fragment and jump to S4.4.

7. Otherwise, discard the recommended knowledge fragment and jump to S4.4.

7.

7. According to claim 6, the intelligent creation method based on NLP intention recognition is characterized in that: If a new creation directory content and corresponding creation content statement are added to the original creation directory content, the following judgment is made: A1. For the new creation catalog content, sort at least one recommended knowledge fragment corresponding to the new creation catalog content, and let i=1; A2, select the i-th recommended knowledge fragment; A3, judging whether the knowledge fragment meets the semantic requirements of the creation catalog and whether the theme is consistent. If so, temporarily retain the i-th recommended knowledge fragment and then jump to A4. Otherwise, discard the i-th recommended knowledge fragment and then jump to A5. A4, calculate the similarity between the first word vector of the knowledge segment and the third word vector of the knowledge segment adjacent to the previous segment of the knowledge segment and the fourth word vector of the knowledge segment adjacent to the next segment of the knowledge segment. If the similarity between the first word vector and the third word vector and the similarity between the first word vector and the fourth word vector are both greater than the set threshold, jump to A6, otherwise, jump to A5; A5, determine whether i is greater than N, where N is the number of recommended knowledge fragments. If so, end the process. Otherwise, set i=i+1 and jump to A2. A6, determine whether the knowledge fragment has the previous M knowledge fragments, where M is a positive integer not less than 2. If so, jump to A7, otherwise, jump to A8; A7, perform semantic analysis on the knowledge fragment and the previous M knowledge fragments of the knowledge fragment respectively. If the knowledge fragment and the previous M knowledge fragments of the knowledge fragment have the same semantics, jump to A8; otherwise, discard the recommended knowledge fragment, set i=i+1, and then jump to A2; A8, determine whether the knowledge fragment has Q knowledge fragments after it, where Q is a positive integer not less than 2. If so, jump to A9, otherwise, keep the recommended knowledge fragment; A9. Perform semantic analysis on the knowledge fragment and the knowledge fragments in the Q segments after the knowledge fragment. If the knowledge fragment and the knowledge fragments in the Q segments after the knowledge fragment have the same semantics, retain the recommended knowledge fragment. Otherwise, discard the recommended knowledge fragment, set i=i+1, and then jump to A2.

8. According to claim 7, the intelligent creation method based on NLP intention recognition is characterized in that: In A4, if the new creative directory content does not have the previous level of creative directory content, the similarity between the first word vector and the fourth word vector is greater than the set threshold, that is, jump to A6, otherwise, jump to A5; if the new creative directory content does not have the next level of creative directory content, the similarity between the first word vector and the third word vector is greater than the set threshold, that is, jump to A6, otherwise, jump to A5.

Images