A method and device for generating review manuscripts based on semantic relationship mining

Abstract

The present application provides a review manuscript generation method and device based on semantic relationship mining, which relates to the field of data processing technology, including: obtaining a set of key semantic indicators; constructing a semantic content extraction model; performing content extraction to obtain N key semantic content sets; performing decreasing serialization processing to obtain a decreasing relationship sequence of papers; performing iterative association memory integration to obtain multiple integrated semantic information; inputting multiple integrated semantic information into multiple generation branches of the automatic generation network layer according to multiple manuscript module identifiers to generate module drafts, and using the integrated fully connected network layer of the automatic generation network layer to integrate and analyze multiple module drafts to obtain the target review manuscript. The technical problems of low quality of integration of data research results and low depth of review manuscripts in the prior art are solved, and the technical effect of in-depth mining of the association of papers and improving the quality of review manuscript generation is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a method and device for generating review manuscripts based on semantic relationship mining. Background Art

[0002] Writing a paper review is a crucial part of academic research, which can systematically summarize, analyze and evaluate the research progress and achievements in a certain field. With the continuous development of academic research, the integration of research results is becoming more and more important. When writing a paper review, it is usually necessary to call on materials from other fields for integrated research, so as to clearly describe the existing research results and existing problems of a certain academic issue (or research field) in a certain period of time.

[0003] At present, with the development of digital tools, such as document management software and data visualization tools, it is of great help to improve the portability of writing paper reviews. However, digital tools have great limitations in use, and the effectiveness of the tools is restricted by the level of the users, resulting in low quality of paper reviews. And with the exponential growth of the number of academic research materials, traditional digital tools cannot meet the needs for the depth of data analysis, and due to the integration of research results, the analysis results generated after mechanized data extraction and integration lack the analysis of the relationship between the relevant data of academic issues, resulting in the paper reviews formed after integration often presenting the situation of "comprehensive but not narrating".

[0004] The existing technology has technical problems such as low quality of integration of research results and low depth of review manuscripts. Summary of the invention

[0005] The present application provides a method and device for generating a review manuscript based on semantic relationship mining, which is used to solve the technical problems in the prior art of low quality integration of research results and low depth of review manuscripts.

[0006] In view of the above problems, the present application provides a method and device for generating review manuscripts based on semantic relationship mining.

[0007] In a first aspect, the present application provides a method for generating a review manuscript based on semantic relationship mining, the method comprising:

[0008] Acquire a key semantic indicator set, wherein the key semantic indicator set includes a semantic sentence indicator set and a semantic phrase indicator set;

[0009] Extracting content from a memory paper library based on the semantic sentence index set and the semantic phrase index set to obtain a memory content extraction result set, and constructing a semantic content extraction model based on the memory content extraction result set and the memory paper library;

[0010] Using the semantic content extraction model to extract content from N papers in the paper pool, obtaining N key semantic content sets;

[0011] Obtaining a target technical problem and a target technical background, performing derivative keyword expansion according to the target technical problem and the target technical background, and centrally screening the N key semantic content sets according to the expansion results, and performing descending serialization processing on the screening results to obtain a descending relationship sequence of papers, wherein the descending relationship sequence of papers includes M related papers, ;

[0012] Taking the key semantic indicator set as an index, traversing the paper decreasing relationship sequence to perform iterative associative memory integration to obtain a plurality of integrated semantic information, wherein the plurality of integrated semantic information includes a plurality of manuscript module identifiers;

[0013] According to the multiple manuscript module identifiers, the multiple integrated semantic information is input into multiple generation branches of the automatically generated network layer to generate module drafts, and the integrated fully connected network layer of the automatically generated network layer is used to integrate and analyze the multiple module drafts to obtain the target review manuscript.

[0014] In a second aspect, the present application provides a review manuscript generation device based on semantic relationship mining, the device comprising:

[0015] A key semantic indicator set acquisition module, used to acquire a key semantic indicator set, wherein the key semantic indicator set includes a semantic sentence indicator set and a semantic phrase indicator set;

[0016] A semantic content extraction module construction module is used to extract content from the memory paper library based on the semantic sentence index set and the semantic phrase index set, obtain a memory content extraction result set, and construct a semantic content extraction model based on the memory content extraction result set and the memory paper library;

[0017] A key semantic content set acquisition module is used to extract content from N papers in the paper pool using the semantic content extraction model to obtain N key semantic content sets;

[0018] The paper decreasing relationship sequence acquisition module is used to obtain the target technical problem and the target technical background, perform derivative keyword expansion according to the target technical problem and the target technical background, and centrally screen the N key semantic content sets according to the expansion results, and perform decreasing serialization processing on the screening results to obtain the paper decreasing relationship sequence, wherein the paper decreasing relationship sequence includes M related papers, ;

[0019] An integrated semantic information acquisition module is used to use the key semantic indicator set as an index, traverse the paper decreasing relationship sequence to perform iterative associative memory integration, and obtain multiple integrated semantic information, wherein the multiple integrated semantic information includes multiple manuscript module identifiers;

[0020] The target review manuscript acquisition module is used to input the multiple integrated semantic information into the multiple generation branches of the automatically generated network layer according to the multiple manuscript module identifiers to generate module drafts, and use the integrated fully connected network layer of the automatically generated network layer to integrate and analyze the multiple module drafts to obtain the target review manuscript.

[0021] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0022] The embodiment of the present application obtains a key semantic indicator set, wherein the key semantic indicator set includes a semantic sentence indicator set and a semantic phrase indicator set, and then extracts content from a memory paper library based on the semantic sentence indicator set and the semantic phrase indicator set to obtain a memory content extraction result set, and constructs a semantic content extraction model based on the memory content extraction result set and the memory paper library, and then uses the semantic content extraction model to extract content from N papers in the paper pool to obtain N key semantic content sets, and then obtains target technical issues and target technical backgrounds, performs derivative keyword expansion based on the target technical issues and target technical backgrounds, centrally screens the N key semantic content sets based on the expansion results, and performs descending serialization processing on the screening results to obtain a descending relationship sequence of papers, wherein the descending relationship sequence of papers includes M related papers, , and then use the key semantic indicator set as the index, traverse the descending relationship sequence of the paper to perform iterative associative memory integration, and obtain multiple integrated semantic information, where the multiple integrated semantic information includes multiple manuscript module identifiers, and then input the multiple integrated semantic information into the multiple generation branches of the automatic generation network layer according to the multiple manuscript module identifiers to generate the module draft, and use the integrated fully connected network layer of the automatic generation network layer to integrate and analyze the module draft to obtain the target review manuscript. The technical effect of high-quality integration of materials, improving the quality of review manuscript generation and data mining depth has been achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 A flowchart of a method for generating a review manuscript based on semantic relationship mining is provided for this application;

[0024] Figure 2 A structural schematic diagram of a review manuscript generation device based on semantic relationship mining is provided for this application.

[0025] Explanation of the accompanying drawings: key semantic indicator set acquisition module 11, semantic content extraction module construction module 12, key semantic content set acquisition module 13, paper decreasing relationship sequence acquisition module 14, integrated semantic information acquisition module 15, target review manuscript acquisition module 16. DETAILED DESCRIPTION

[0026] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below.

[0027] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in conjunction with the accompanying drawings. The described embodiments should not be regarded as limiting the present application. All other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of this application.

[0028] In the following description, reference is made to "some embodiments", which describe a subset of all possible embodiments, but it is understood that "some embodiments" may be the same subset or different subsets of all possible embodiments, and may be combined with each other without conflict, and the terms "first\second" involved are merely to distinguish similar objects and do not represent a specific ordering of objects. The terms "including" and "having" and any variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or modules that are not clearly listed or inherent to these processes, methods, products, or devices. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those generally understood by technicians in the technical field of this application. The terms used herein are for the purpose of describing the embodiments of the present application only.

[0029] The present application provides a review manuscript generation method and device based on semantic relationship mining, obtains a set of key semantic indicators; constructs a semantic content extraction model; performs content extraction to obtain N key semantic content sets; performs descending serialization processing to obtain a descending relationship sequence of papers; performs iterative associative memory integration to obtain multiple integrated semantic information; inputs the multiple integrated semantic information into multiple generation branches of the automatically generated network layer according to multiple manuscript module identifiers to generate module drafts, and uses the integrated fully connected network layer of the automatically generated network layer to integrate and analyze the multiple module drafts to obtain a target review manuscript, which is used to solve the technical problems in the prior art that the integration quality of data research results is not high and the depth of the review manuscript is low.

[0030] Embodiment 1: Figure 1 As shown, the present application provides a method for generating a review manuscript based on semantic relationship mining, wherein the method comprises:

[0031] S100: Acquire a key semantic indicator set, wherein the key semantic indicator set includes a semantic sentence indicator set and a semantic phrase indicator set;

[0032] In a possible embodiment, the key semantic indicator set is a key semantic framework for constructing a review manuscript, including the semantic sentence indicator set and the semantic phrase indicator set. The semantic sentence indicator set and the semantic phrase indicator set are obtained by dividing the semantic indicator set from the perspective of indicator expression type. By obtaining the key semantic indicator set pre-set by a person skilled in the art, a technical effect of providing a basis for subsequent reliable paper semantic content extraction is achieved.

[0033] Optionally, the semantic sentence indicator set includes research background sentences, research purpose sentences, research question sentences, research method sentences, research result sentences, research conclusion sentences, concept definition sentences, innovation sentences, and quotation sentences.

[0034] The semantic phrase indicator set includes research question phrases, method model phrases, theoretical principle phrases, measurement indicator phrases, data material phrases, software system phrases and instrument equipment phrases.

[0035] S200: extracting content from a memory paper library based on the semantic sentence index set and the semantic phrase index set to obtain a memory content extraction result set, and constructing a semantic content extraction model based on the memory content extraction result set and the memory paper library;

[0036] Furthermore, step S200 of the embodiment of the present application further includes:

[0037] Extracting text from the memory paper library according to the semantic sentence index set and the semantic phrase index set, and marking the extraction results with semantic tags to obtain the memory content extraction result set;

[0038] The memory content extraction result set and the memory paper library are used as training data, and supervised training is performed on the framework built based on the convolutional neural network. The loss function is used to identify the loss of the training results until the loss is less than or equal to the preset loss value, thereby obtaining the trained semantic content extraction model.

[0039] In a possible embodiment, after obtaining the semantic sentence index set and the semantic phrase index set, manual content extraction can be performed on the papers in the memory paper library (a database containing relevant documents) pre-set by technical personnel in the field for constructing a semantic content extraction model, and the extracted results can be semantically labeled (semantic content identification of the extracted content expression) to obtain the memory content extraction result set.

[0040] The semantic content extraction model is a functional model for extracting content from the semantic sentence index set and the semantic phrase index set in the retrieved papers. The model can be used to extract content from new papers or other texts, and automatically identify and mark text content that meets the semantic sentence index set and the semantic phrase index set. Thus, the goal of providing data support for building a semantic content extraction model is achieved, and the technical effect of improving the accuracy and efficiency of semantic content extraction is achieved.

[0041] In one embodiment, each paper in the memory paper library is traversed, and for each paper, sentences and phrases that meet the conditions are manually extracted according to a set of semantic sentence indicators (such as research background sentences, research purpose sentences, etc.) and a set of semantic phrase indicators (such as research question phrases, method model phrases, etc.), and the extracted sentences and phrases are used as text extraction results. For each extracted sentence or phrase, a corresponding semantic label is assigned according to its content. For example, the topic sentence is marked as "topic", the conclusion sentence is marked as "conclusion", and the professional term is marked as "terminology", etc. Then, the marked text is stored as a memory content extraction result set, and each memory content extraction result contains the extracted text content and the corresponding semantic label.

[0042] Optionally, each labeled text in the memory content extraction result set is aggregated with the memory paper library to generate a training sample. The training sample is used to train the semantic content extraction model. The framework of the semantic content extraction model is constructed based on a convolutional neural network, which generally includes an input layer, multiple convolutional layers, a pooling layer, a fully connected layer, and an output layer.

[0043] Then, the training data is sent to the constructed framework in batches for forward propagation, and the loss between the predicted label and the true label is calculated using the loss function (such as the cross entropy loss function). Then, the gradient of the loss with respect to the model parameters is calculated using the back propagation algorithm, and the model parameters are updated to minimize the loss. Repeat the above process until the loss is less than or equal to the preset loss value or the preset training rounds are reached, and the trained semantic content extraction model is obtained. Thus, the goal of constructing an efficient semantic content extraction model based on a convolutional neural network is achieved, and the technical effect of accurately extracting key information from a large amount of text and providing strong support for the generation of subsequent review manuscripts is achieved.

[0044] S300: extracting content from N papers in the paper pool using the semantic content extraction model to obtain N key semantic content sets;

[0045] In one possible embodiment, a person skilled in the art conducts a preliminary search in a large database for the relevant content of the target review manuscript to be generated as needed, and summarizes the retrieved documents to generate the paper pool. Among them, the paper content in the paper pool has different correlations with the review manuscript to be generated, and further semantic recognition is required. The N papers are input into the semantic content extraction model for content extraction, thereby obtaining the N key semantic content sets. Among them, the N key semantic content sets are semantic contents that can reflect the research direction and specific research situation of the N papers. N is an integer greater than or equal to 1.

[0046] S400: Obtaining a target technical problem and a target technical background, performing derivative keyword expansion according to the target technical problem and the target technical background, and centrally screening the N key semantic content sets according to the expansion results, and performing descending serialization processing on the screening results to obtain a descending relationship sequence of papers, wherein the descending relationship sequence of papers includes M related papers, ;

[0047] Furthermore, step S400 in the embodiment of the present application further includes:

[0048] Traversing the N key semantic content sets to extract keywords and obtain N keyword sets;

[0049] Extracting keywords from the target technical problem and the target technical background to obtain an initial keyword set;

[0050] Derivatively expanding the initial keyword set according to the neighborhood association index and the cross-domain association index to generate a target neighborhood keyword set and a target cross-domain keyword set;

[0051] The N keyword sets are centrally screened using the target adjacent keyword set and the target cross-domain keyword set to obtain screening results, wherein the screening results include M related papers.

[0052] In one possible embodiment, a person skilled in the art determines the target technical problem in the research field that needs to be reviewed and analyzed and the target technical background of the technical problem. The target technical problem is the target object that needs to be studied, and the target technical background describes the research background of the target technical problem and constrains the research progress range of the target technical problem. The technical effect of improving the reliability of review manuscript generation is further achieved. Optionally, detailed technical problems and background descriptions can also be obtained from channels such as user input, literature review, and expert interviews.

[0053] In one embodiment, after obtaining the target technical problem and the target technical background, due to the word limit, the information that can be reflected is relatively one-sided, so it is necessary to perform derivative keyword expansion, so as to provide a reliable screening basis for the subsequent screening of the N key semantic content sets and obtain the expansion results. According to the expansion results, the N key semantic content sets are centrally screened to obtain M related papers with higher relevance. The M related papers are serialized in descending order according to the M target approximations corresponding to the M related papers, so as to obtain the descending relationship sequence of the papers. M is an integer greater than 1.

[0054] By obtaining the descending key sequence of the papers, it is convenient to subsequently iteratively integrate the associated memories from high to low relevance, thereby achieving the technical effect of improving the efficiency and quality of integration and avoiding the addition of redundant information as a prelude to the review manuscript.

[0055] Optionally, natural language processing (NLP) technology (such as word segmentation, part-of-speech tagging, named entity recognition, etc.) is used to extract keywords from the target technical problem and the target technical background, and the extraction results are merged and duplicates are removed to obtain the initial keyword set. Based on the same processing principle, keyword extraction is performed on the N key semantic content sets to obtain the N keyword sets.

[0056] Optionally, the neighborhood association index is a semantic relevance index with the initial keyword. The cross-domain association index is an index such as interdisciplinary degree, literature citation relationship, etc. By using the neighborhood association index, based on a domain knowledge base, a synonym library or a corpus, a neighboring vocabulary highly related to the initial keyword is searched to generate a target neighboring keyword set.

[0057] Optionally, based on the cross-domain association index, by searching the cross-domain knowledge graph, literature citation relationship and other data, possible related cross-domain vocabulary is obtained to generate a target cross-domain keyword set. Then, the target adjacent keyword set and the target cross-domain keyword set are used to centrally screen the N keyword sets to obtain the screening results. The screening results include M related papers.

[0058] Furthermore, step S400 in the embodiment of the present application further includes:

[0059] Performing similarity identification on the target adjacent keyword set and the target cross-domain keyword set and the keywords in the N keyword sets to obtain N similarity sets;

[0060] Performing mean processing on the N similarity sets to obtain N similarity centers;

[0061] Taking the N similarity centers as starting points, iteratively screening is performed in the N similarity sets according to a preset centralized screening bandwidth, and the N similarity centers are updated according to the screening results to obtain N updated similarity centers;

[0062] Centralized screening is performed based on the N updated similarity centers to determine the screening result.

[0063] Furthermore, step S400 in the embodiment of the present application further includes:

[0064] Constructing N approximate neighborhoods, wherein the N approximate neighborhoods are regions constructed with the N updated proximity centers as the center and the preset centralized screening bandwidth as the radius;

[0065] Performing approximation mean processing on the N approximate neighborhoods to obtain N target approximations;

[0066] The N target proximities are sorted in descending order, and M papers corresponding to the first M target proximities are selected as the M related papers to obtain the screening result.

[0067] In one embodiment of the present application, the cosine similarity calculation formula is used to calculate the similarity between any one of the N keyword sets and the target adjacent keyword set and the target cross-domain keyword set keywords to obtain a target adjacent keyword similarity set and a target cross-domain keyword similarity set for a keyword. According to the weight ratio pre-set by a person skilled in the art, the target adjacent keyword similarity set and the target cross-domain keyword similarity set of the keyword are weightedly calculated to obtain the similarity of the keyword. Based on the same principle, the similarity calculation is performed on the keywords in the N keyword sets to obtain the N similarity sets. Among them, the N similarity sets reflect the closeness of the N papers corresponding to the N keyword sets to the target technical issues and target technical backgrounds.

[0068] In one embodiment, the N similarity sets are averaged to obtain N similarity centers. The N similarity centers reflect the average degree of similarity of the N similarity sets. Then, starting from the N similarity centers, iterative screening is performed in the N similarity sets according to the preset centralized screening bandwidth (the difference in similarity of a single move during centralized screening set by technicians in this field), and the N similarity centers are updated according to the screening results to obtain N updated similarity centers. The N updated similarity centers are the similarities that best reflect the approximation of the N similarity sets after ignoring the edge and scattered similarities. Then, the N similarity sets are analyzed and screened according to the N updated similarity centers, and the corresponding M related papers are obtained according to the analysis result mapping to generate the screening results.

[0069] In a possible embodiment, an area constructed with the N updated similarity centers as the center and the preset centralized screening bandwidth as the radius is used as N approximate neighborhoods, wherein the N approximate neighborhoods are neighborhoods constructed with the approximation close to the N updated similarity centers.

[0070] Then, the approximation sets in the N approximate neighborhoods are processed by the approximation mean to obtain N target approximations, wherein the N target approximations reflect the closeness of the N papers to the target technical problem and target technical background.

[0071] Optionally, the N target similarities are sorted in descending order, and the M papers corresponding to the first M target similarities are selected as the M related papers to obtain the screening results. In this way, the goal of intelligently screening the N papers, eliminating papers with low relevance, and screening papers with high relevance is achieved, achieving the technical effect of paving the way for the subsequent generation of high-quality review manuscripts.

[0072] S500: using the key semantic indicator set as an index, traversing the paper decreasing relationship sequence to perform iterative associative memory integration to obtain a plurality of integrated semantic information, wherein the plurality of integrated semantic information includes a plurality of manuscript module identifiers;

[0073] Furthermore, step S500 of the embodiment of the present application further includes:

[0074] Taking multiple indicator modules in the key semantic indicator set as indexes, clustering the M keyword sets of the M related papers in the paper decreasing relationship sequence to obtain M module keyword clusters, wherein the M module keyword clusters include M indicator module identifier sets, each module keyword cluster includes multiple module keyword sets, each module keyword set corresponds to an indicator module identifier, and each indicator module identifier corresponds to a manuscript module identifier;

[0075] Extracting a first indicator module from the multiple indicator modules, and using the first indicator module as an index, screening M indicator module identifier sets to obtain M first module keyword sets;

[0076] Iteratively associating and memorizing the M first module keyword sets according to the decreasing relationship sequence of the paper to obtain first integrated semantic information;

[0077] The M module keyword clusters are iteratively associated and memorized according to multiple indicator modules to obtain the multiple integrated semantic information.

[0078] Furthermore, step S500 of the embodiment of the present application further includes:

[0079] According to the order of the M related papers in the paper decreasing relationship sequence, extract the first module keyword set corresponding to the first related paper from the M first module keyword sets as the first first module keyword set, and add the text feature set corresponding to the first first module keyword set into the first memory unit;

[0080] Extracting the first module keyword set corresponding to the second-ranked related paper from the M first module keyword sets again as the second first module keyword set;

[0081] Taking the first first module keyword set and the first memory unit as matching targets, performing text feature association matching on the second related paper to obtain a first matching result;

[0082] Acquire a text feature set corresponding to the second first module keyword set to obtain a second text feature set, and add the second text feature set and the first matching result into a second memory unit;

[0083] Based on the first memory unit and the second memory unit, the M related papers are iteratively associated and integrated to obtain the Mth memory unit, and the Mth memory unit is used as the first integrated semantic information.

[0084] In a possible embodiment, the key semantic indicator set is used as an index, and the M related papers in the paper decreasing relationship sequence are iteratively identified one by one using the indicator module corresponding to the indicator, and iterative association memory integration is performed to generate the multiple integrated semantic information. The multiple integrated semantic information includes multiple manuscript module identifiers.

[0085] By iteratively associating and memorizing the descending relationship sequence of the papers, information with high relevance and high semantic recognition quality is obtained, and the M related papers are orderly integrated, so as to achieve the goal of providing data support for the generation of review manuscripts. And with multiple indicator modules in the key semantic indicator set as indexes, the M related papers in the descending relationship sequence of the papers are integrated, and the semantic information required by different modules in the review manuscript is iterated respectively, so as to achieve the technical effect of improving the efficiency and reliability of the subsequent full-text generation of the review manuscript.

[0086] Optionally, the multiple manuscript module identifiers include a research background module identifier, a research content module identifier, a research method module identifier, a research result module identifier, and a research conclusion module identifier. Among them, the integrated semantic information corresponding to the research background module identifier is used to introduce the relevant background of the research field and the current status of the research problem. The integrated semantic information corresponding to the research content module identifier is used to introduce the research results of predecessors in this field, including the viewpoints, theories, methods, experiments and other contents in the existing literature, and analyze the advantages, disadvantages, deficiencies and unresolved problems of predecessors' research. The integrated semantic information corresponding to the research method module identifier is used to introduce the research methods, experimental design, data processing, equipment use and other contents related to the research content, and explain the feasibility and scientificity of the research. The integrated semantic information corresponding to the research result module identifier is used to introduce the research results and findings of the research problem, including data, experimental equipment and method models, etc., analyze the rationality and practicality of the results, explore the similarities and differences between the results and predecessors' research, as well as the innovation and limitations. The integrated semantic information corresponding to the research conclusion module identifier is used to summarize the relevant research results and contributions, explain the significance and value of the relevant research, point out the deficiencies and areas that need improvement, and look forward to future research directions and development trends.

[0087] In a possible embodiment, the multiple indicator modules include a research background indicator module, a research content indicator module, a research method indicator module, a research result indicator module and a research conclusion indicator module. The research background indicator module, the research content indicator module, the research method indicator module, the research result indicator module and the research conclusion indicator module are obtained by classifying the key semantic indicator set from the description purpose of the key semantic indicator in the review manuscript.

[0088] Optionally, an empty module keyword list is created for each of the multiple indicator modules in the key semantic indicator set, the keywords in the M keyword sets are traversed, and the keywords are added to the corresponding keyword list for clustering according to the matching results of each keyword attribute with the multiple indicator modules, thereby summarizing the obtained multiple keyword lists to obtain the M indicator module identification sets. The M module keyword clusters include M indicator module identification sets, each module keyword cluster includes multiple module keyword sets, each module keyword set corresponds to an indicator module identification, and each indicator module identification corresponds to a manuscript module identification.

[0089] Optionally, a first indicator module is extracted from the multiple indicator modules, and M indicator module identifier sets are screened with the first indicator module as an index, and module keyword sets corresponding to module indicator identifiers that meet the first indicator module are extracted to obtain M first module keyword sets.

[0090] After obtaining the M first module keyword sets, the M first module keyword sets are iteratively associated and memorized and integrated according to the paper decreasing relationship sequence to obtain the first integrated semantic information. Further, based on the same acquisition principle of obtaining the first integrated semantic information, the M module keyword clusters are iteratively associated and memorized and integrated according to multiple indicator modules to obtain the multiple integrated semantic information.

[0091] In a possible implementation, according to the order of the M related papers in the paper descending sequence, the first module keyword set corresponding to the first related paper is extracted from the M first module keyword sets, and is used as the first first module keyword set, and the text feature set corresponding to the first first module keyword set is added to the first memory unit. The first memory unit is a unit for storing initial text features. Optionally, the text feature set is obtained by extracting features from the text containing the keywords in the corresponding related papers using the first first module keyword set as an index.

[0092] Then, the first module keyword set corresponding to the second-ranked related paper is extracted from the M first module keyword sets as the second first module keyword set. The relevance between the second-ranked related paper and the target technical problem is less than the relevance between the first-ranked related paper and the target technical problem.

[0093] In one embodiment, the first first module keyword set and the first memory unit are used as matching targets, that is, the first first module keyword set and the first memory unit are used as indexes to match and screen the texts with higher similarity in the second related papers to achieve text feature association matching and obtain a first matching result.

[0094] Feature extraction is performed on the text of the position of the keywords in the second first module keyword set in the second related paper to obtain a second text feature set, and then the second text feature set and the first matching result are added to the second memory unit.

[0095] By analogy, the M related papers are iteratively associated and memorized based on the first memory unit and the second memory unit until the text features in the M related papers are iteratively extracted. At this time, the iteration is stopped to obtain the Mth memory unit, and the Mth memory unit is used as the first integrated semantic information. The technical effect of in-depth iterative semantic relationship mining of the M related papers and improving the quality of semantic information is achieved.

[0096] S600: Inputting the multiple integrated semantic information into multiple generation branches of the automatically generated network layer according to the multiple manuscript module identifiers to generate module drafts, and using the integrated fully connected network layer of the automatically generated network layer to integrate and analyze the multiple module drafts to obtain the target review manuscript.

[0097] In one possible implementation, the multiple manuscript module identifiers and the corresponding multiple integrated semantic information obtained in the previous step are respectively input into the automatic generation network layer (a functional model for automatically generating a complete review manuscript based on the integrated semantic information corresponding to each manuscript module) to obtain multiple module drafts. Each module draft corresponds to the specific description content of a writing module in the review manuscript. A variety of review manuscript templates are embedded in the automatic generation network layer, and the review manuscript templates can be adaptively matched and adjusted based on the input multiple module drafts to obtain the target review manuscript.

[0098] Optionally, each generation branch is responsible for generating a draft corresponding to the module based on the integrated semantic information of a specific module identifier. These generation branches can be based on deep learning models, such as recurrent neural networks (RNN), long short-term memory networks (LSTM), Transformers, etc., which can learn and generate coherent text. Each generation branch outputs a module draft, which is independent of each other but closely related to its corresponding module identifier. Exemplarily, a branch framework is constructed based on a convolutional neural network, and multiple sample module drafts and multiple sample integrated semantic information are obtained as construction data. The branch framework is supervised and trained using the construction data, and the mapping relationship between the integrated semantic information and the module draft is learned. The network parameters of the framework are updated according to the output results during training until the model converges and a generation branch with training is obtained.

[0099] Optionally, the multiple generating branches are connected in parallel, and the integrated fully connected network layer is fully connected to the output layers of the multiple generating branches. The integrated fully connected network layer is a complex neural network structure that can capture the correlation and dependency between the first drafts of different modules, as well as their position and role in the entire review manuscript. The integrated fully connected network layer matches the corresponding review manuscript template by analyzing the content, style, logical order, etc. of the first drafts of each module, and then outputs a complete, integrated and analyzed target review manuscript.

[0100] Optionally, multiple sample module draft sets and multiple sample review manuscripts are obtained as training data, and supervised training is performed on the network layer constructed based on the recurrent neural network. During the training, the network parameters are updated according to the output results until the model output converges, thereby obtaining the integrated fully connected network layer that has completed the training.

[0101] By using multiple generation branches to accurately capture the integrated semantic information of the corresponding manuscript module and generate a matching draft, the fully connected network layer is then used to effectively integrate the drafts of each module and generate a high-quality review manuscript. The large model is used to automatically summarize, summarize, and organize, and finally form a draft of the review paper, reducing the time for writing the paper, and mining the decreasing relationship of the paper to improve the technical effect of automatically writing the review manuscript.

[0102] The present application provides a method and device for generating a review manuscript based on semantic relationship mining, which has the following technical effects:

[0103] 1. This application obtains a set of key semantic indicators, builds a semantic content extraction model, and then uses the semantic content extraction model to extract content, obtains N key semantic content sets, analyzes the N key semantic content sets, performs descending serialization processing on N papers, obtains a descending relationship sequence of papers, and obtains multiple integrated semantic information through iterative associative memory integration. Then, according to multiple manuscript module identifiers, multiple integrated semantic information is input into multiple generation branches of the automatic generation network layer to generate module drafts, and the integrated fully connected network layer of the automatic generation network layer is used to integrate and analyze multiple module drafts to obtain the target review manuscript. The technical effect of improving the writing depth and writing quality of the review manuscript is achieved.

[0104] 2. By extracting keywords from the target technical issues and target technical backgrounds and expanding derived keywords, the goal of improving the reliability of paper retrieval is achieved. Then, the N key semantic content sets are centrally screened based on the expansion results, and M related papers are determined based on the centralized screening. The papers in the paper pool are screened to improve the reliability of subsequent review manuscript writing and analysis. Then, the screening results are serialized in descending order to obtain a descending relationship sequence of papers, thereby paving the way for subsequent iterative associative memory integration and improving the technical effect of semantic integration quality.

[0105] Embodiment 2: Based on the same inventive concept as the method for generating a review manuscript based on semantic relationship mining in the above embodiment, Figure 2 As shown, the present application provides a review manuscript generation device based on semantic relationship mining, the device comprising:

[0106] A key semantic indicator set acquisition module 11 is used to acquire a key semantic indicator set, wherein the key semantic indicator set includes a semantic sentence indicator set and a semantic phrase indicator set;

[0107] A semantic content extraction module construction module 12 is used to extract content from the memory paper library based on the semantic sentence index set and the semantic phrase index set, obtain a memory content extraction result set, and construct a semantic content extraction model based on the memory content extraction result set and the memory paper library;

[0108] A key semantic content set acquisition module 13 is used to extract content from N papers in the paper pool using the semantic content extraction model to obtain N key semantic content sets;

[0109] The paper decreasing relationship sequence acquisition module 14 is used to obtain the target technical problem and the target technical background, perform derivative keyword expansion according to the target technical problem and the target technical background, and centrally screen the N key semantic content sets according to the expansion results, and perform decreasing serialization processing on the screening results to obtain the paper decreasing relationship sequence, wherein the paper decreasing relationship sequence includes M related papers, ;

[0110] The integrated semantic information acquisition module 15 is used to use the key semantic indicator set as an index, traverse the paper decreasing relationship sequence to perform iterative associative memory integration, and obtain multiple integrated semantic information, wherein the multiple integrated semantic information includes multiple manuscript module identifiers;

[0111] The target review manuscript acquisition module 16 is used to input the multiple integrated semantic information into the multiple generation branches of the automatically generated network layer according to the multiple manuscript module identifiers to generate module drafts, and use the integrated fully connected network layer of the automatically generated network layer to integrate and analyze the multiple module drafts to obtain the target review manuscript.

[0112] Furthermore, the paper decreasing relationship sequence obtaining module 14 is used to perform the following steps:

[0113] Traversing the N key semantic content sets to extract keywords and obtain N keyword sets;

[0114] Extracting keywords from the target technical problem and the target technical background to obtain an initial keyword set;

[0115] Derivatively expanding the initial keyword set according to the neighborhood association index and the cross-domain association index to generate a target neighborhood keyword set and a target cross-domain keyword set;

[0116] The N keyword sets are centrally screened using the target adjacent keyword set and the target cross-domain keyword set to obtain screening results, wherein the screening results include M related papers.

[0117] Furthermore, the paper decreasing relationship sequence obtaining module 14 is used to perform the following steps:

[0118] Performing similarity identification on the target adjacent keyword set and the target cross-domain keyword set and the keywords in the N keyword sets to obtain N similarity sets;

[0119] Performing mean processing on the N similarity sets to obtain N similarity centers;

[0120] Taking the N similarity centers as starting points, iteratively screening is performed in the N similarity sets according to a preset centralized screening bandwidth, and the N similarity centers are updated according to the screening results to obtain N updated similarity centers;

[0121] Centralized screening is performed based on the N updated similarity centers to determine the screening result.

[0122] Furthermore, the paper decreasing relationship sequence obtaining module 14 is used to perform the following steps:

[0123] Constructing N approximate neighborhoods, wherein the N approximate neighborhoods are regions constructed with the N updated proximity centers as the center and the preset centralized screening bandwidth as the radius;

[0124] Performing approximation mean processing on the N approximate neighborhoods to obtain N target approximations;

[0125] The N target proximities are sorted in descending order, and M papers corresponding to the first M target proximities are selected as the M related papers to obtain the screening result.

[0126] Furthermore, the integrated semantic information obtaining module 15 is used to perform the following steps:

[0127] Taking multiple indicator modules in the key semantic indicator set as indexes, clustering the M keyword sets of the M related papers in the paper decreasing relationship sequence to obtain M module keyword clusters, wherein the M module keyword clusters include M indicator module identifier sets, each module keyword cluster includes multiple module keyword sets, and each module keyword set corresponds to an indicator module identifier;

[0128] Extracting a first indicator module from the multiple indicator modules, and using the first indicator module as an index, screening M indicator module identifier sets to obtain M first module keyword sets;

[0129] Iteratively associating and memorizing the M first module keyword sets according to the decreasing relationship sequence of the paper to obtain first integrated semantic information;

[0130] The M module keyword clusters are iteratively associated and memorized according to multiple indicator modules to obtain the multiple integrated semantic information.

[0131] Furthermore, the integrated semantic information obtaining module 15 is used to perform the following steps:

[0132] According to the order of the M related papers in the paper decreasing relationship sequence, extract the first module keyword set corresponding to the first related paper from the M first module keyword sets as the first first module keyword set, and add the text feature set corresponding to the first first module keyword set into the first memory unit;

[0133] Extracting the first module keyword set corresponding to the second-ranked related paper from the M first module keyword sets again as the second first module keyword set;

[0134] Taking the first first module keyword set and the first memory unit as matching targets, performing text feature association matching on the second related paper to obtain a first matching result;

[0135] Acquire a text feature set corresponding to the second first module keyword set to obtain a second text feature set, and add the second text feature set and the first matching result into a second memory unit;

[0136] Based on the first memory unit and the second memory unit, the M related papers are iteratively associated and integrated to obtain the Mth memory unit, and the Mth memory unit is used as the first integrated semantic information.

[0137] Furthermore, the semantic content extraction module construction module 12 is used to perform the following steps:

[0138] Extracting text from the memory paper library according to the semantic sentence index set and the semantic phrase index set, and marking the extraction results with semantic tags to obtain the memory content extraction result set;

[0139] The memory content extraction result set and the memory paper library are used as training data, and supervised training is performed on the framework built based on the convolutional neural network. The loss function is used to identify the loss of the training results until the loss is less than or equal to the preset loss value, thereby obtaining the trained semantic content extraction model.

[0140] Through the above-mentioned detailed description of a method for generating a review manuscript based on semantic relationship mining in this specification, those skilled in the art can clearly know a method and device for generating a review manuscript based on semantic relationship mining in this embodiment. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the description of the method part.

[0141] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for generating review manuscripts based on semantic relationship mining, characterized in that: The method comprises: Acquire a key semantic indicator set, wherein the key semantic indicator set includes a semantic sentence indicator set and a semantic phrase indicator set; Extracting content from a memory paper library based on the semantic sentence index set and the semantic phrase index set to obtain a memory content extraction result set, and constructing a semantic content extraction model based on the memory content extraction result set and the memory paper library; Using the semantic content extraction model to extract content from N papers in the paper pool, obtaining N key semantic content sets; Obtaining a target technical problem and a target technical background, performing derivative keyword expansion according to the target technical problem and the target technical background, and centrally screening the N key semantic content sets according to the expansion results, and performing descending serialization processing on the screening results to obtain a descending relationship sequence of papers, wherein the descending relationship sequence of papers includes M related papers, M≤N; Taking the key semantic indicator set as an index, traversing the paper decreasing relationship sequence to perform iterative associative memory integration to obtain a plurality of integrated semantic information, wherein the plurality of integrated semantic information includes a plurality of manuscript module identifiers; According to the multiple manuscript module identifiers, the multiple integrated semantic information is input into multiple generation branches of the automatically generated network layer to generate module drafts, and the integrated fully connected network layer of the automatically generated network layer is used to integrate and analyze the multiple module drafts to obtain the target review manuscript; Wherein, the method further comprises: Taking multiple indicator modules in the key semantic indicator set as indexes, clustering the M keyword sets of the M related papers in the paper decreasing relationship sequence to obtain M module keyword clusters, wherein the M module keyword clusters include M indicator module identifier sets, each module keyword cluster includes multiple module keyword sets, and each module keyword set corresponds to an indicator module identifier; Extracting a first indicator module from the multiple indicator modules, and using the first indicator module as an index, screening M indicator module identifier sets to obtain M first module keyword sets; Iteratively associating and memorizing the M first module keyword sets according to the decreasing relationship sequence of the paper to obtain first integrated semantic information; Iteratively associating and memorizing the M module keyword clusters according to multiple indicator modules to obtain the multiple integrated semantic information; Wherein, the method further comprises: According to the order of the M related papers in the paper decreasing relationship sequence, extract the first module keyword set corresponding to the first related paper from the M first module keyword sets as the first first module keyword set, and add the text feature set corresponding to the first first module keyword set into the first memory unit; Extracting the first module keyword set corresponding to the second-ranked related paper from the M first module keyword sets again as the second first module keyword set; Taking the first first module keyword set and the first memory unit as matching targets, performing text feature association matching on the second related paper to obtain a first matching result; Acquire a text feature set corresponding to the second first module keyword set to obtain a second text feature set, and add the second text feature set and the first matching result into a second memory unit; Based on the first memory unit and the second memory unit, the M related papers are iteratively associated and integrated to obtain the Mth memory unit, and the Mth memory unit is used as the first integrated semantic information.

2. A method for generating review manuscripts based on semantic relationship mining as claimed in claim 1, characterized in that: include: Traversing the N key semantic content sets to extract keywords and obtain N keyword sets; Extracting keywords from the target technical problem and the target technical background to obtain an initial keyword set; Derivatively expanding the initial keyword set according to the neighborhood association index and the cross-domain association index to generate a target neighborhood keyword set and a target cross-domain keyword set; The N keyword sets are centrally screened using the target adjacent keyword set and the target cross-domain keyword set to obtain screening results, wherein the screening results include M related papers.

3. A method for generating review manuscripts based on semantic relationship mining as claimed in claim 2, characterized in that: include: Performing similarity identification on the target adjacent keyword set and the target cross-domain keyword set and the keywords in the N keyword sets to obtain N similarity sets; Performing mean processing on the N similarity sets to obtain N similarity centers; Taking the N similarity centers as starting points, iteratively screening is performed in the N similarity sets according to a preset centralized screening bandwidth, and the N similarity centers are updated according to the screening results to obtain N updated similarity centers; Centralized screening is performed based on the N updated similarity centers to determine the screening result.

4. A method for generating review manuscripts based on semantic relationship mining as claimed in claim 3, characterized in that: include: Constructing N approximate neighborhoods, wherein the N approximate neighborhoods are regions constructed with the N updated proximity centers as the center and the preset centralized screening bandwidth as the radius; Performing approximation mean processing on the N approximate neighborhoods to obtain N target approximations; The N target proximities are sorted in descending order, and M papers corresponding to the first M target proximities are selected as the M related papers to obtain the screening result.

5. The method for generating review manuscripts based on semantic relationship mining according to claim 1, characterized in that: include: Extracting text from the memory paper library according to the semantic sentence index set and the semantic phrase index set, and marking the extraction results with semantic tags to obtain the memory content extraction result set; The memory content extraction result set and the memory paper library are used as training data, and supervised training is performed on the framework built based on the convolutional neural network. The loss function is used to identify the loss of the training results until the loss is less than or equal to the preset loss value, thereby obtaining the trained semantic content extraction model.

6. A review manuscript generation device based on semantic relationship mining, characterized in that: The device is used to implement the steps of the method according to any one of claims 1 to 5, and the device comprises: A key semantic indicator set acquisition module, used to acquire a key semantic indicator set, wherein the key semantic indicator set includes a semantic sentence indicator set and a semantic phrase indicator set; A semantic content extraction module construction module is used to extract content from the memory paper library based on the semantic sentence index set and the semantic phrase index set, obtain a memory content extraction result set, and construct a semantic content extraction model based on the memory content extraction result set and the memory paper library; A key semantic content set acquisition module is used to extract content from N papers in the paper pool using the semantic content extraction model to obtain N key semantic content sets; A paper decreasing relationship sequence acquisition module is used to obtain a target technical problem and a target technical background, perform derivative keyword expansion according to the target technical problem and the target technical background, and centrally screen the N key semantic content sets according to the expansion results, and perform decreasing serialization processing on the screening results to obtain a paper decreasing relationship sequence, wherein the paper decreasing relationship sequence includes M related papers, M≤N; An integrated semantic information acquisition module is used to use the key semantic indicator set as an index, traverse the paper decreasing relationship sequence to perform iterative associative memory integration, and obtain multiple integrated semantic information, wherein the multiple integrated semantic information includes multiple manuscript module identifiers; The target review manuscript acquisition module is used to input the multiple integrated semantic information into the multiple generation branches of the automatically generated network layer according to the multiple manuscript module identifiers to generate module drafts, and use the integrated fully connected network layer of the automatically generated network layer to integrate and analyze the multiple module drafts to obtain the target review manuscript.

Images