Semantic association analysis method and system based on academic literature
By extracting basic semantic features of academic literature and calling pre-trained models, the semantic correlation feature representation between documents is generated, which solves the problem of difficulty in digging deep into the semantic relationship of literature in the existing methods, and realizes efficient semantic correlation analysis and accurate result generation.
Patent Information
- Application Number
- CN202510816613.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-18
- Publication Date
- 2025-08-22
- Estimated Expiration
- 2045-06-18
AI Technical Summary
Existing academic literature analysis methods are difficult to deeply explore the semantic relationships and logical relationships between documents, resulting in inefficiency and susceptibility to subjective factors.
By obtaining the data set of academic literature texts, basic semantic features are extracted and cleaned, the pre-trained semantic correlation analysis model is called, and the semantic correlation feature representations between literature text units are generated, including concept co-occurrence correlation degree and semantic logical connection degree, and accurate semantic correlation analysis results are generated.
It has achieved in-depth mining and accurate analysis of complex semantic relationships between academic literature, improved automation level and accuracy, helped scientific researchers to quickly discover potential connections between literature, and promoted academic innovation and knowledge dissemination.
Smart Images

Figure CN120337936B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a semantic association analysis method and system based on academic literature. Background Art
[0002] In the field of academic research, the rapid development of information technology and the increasing abundance of digital literature resources have led to an explosive growth in the volume of academic literature. This literature covers a wide range of disciplines and provides researchers with a valuable source of knowledge. However, faced with this massive volume of academic literature, efficiently mining the inherent connections between them and identifying potential research trends and knowledge connections has become a major challenge facing academic research.
[0003] Traditional methods for analyzing academic literature rely primarily on manual reading and expert judgment. This approach is not only inefficient but also susceptible to subjective factors, making it difficult to fully and accurately reveal the semantic connections between documents. Although natural language processing technology and machine learning algorithms have been widely used in academic literature analysis in recent years, most existing methods focus on surface-level analysis of document content, such as keyword extraction and text classification, and fail to deeply explore the semantic connections and logical relationships between documents. Summary of the Invention
[0004] In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides a semantic association analysis method based on academic documents, the method comprising:
[0005] Obtaining an academic document text data set to be analyzed, wherein the academic document text data set includes a plurality of document text units with subject tags;
[0006] Performing basic semantic feature extraction on the academic document text data set to obtain basic semantic features of each document text unit, wherein the basic semantic features include core concept features and contextual semantic relationship features of the document content;
[0007] Calling a pre-trained semantic association analysis model to perform association feature modeling on the basic semantic features to generate semantic association feature representations between document text units, wherein the semantic association feature representations include concept co-occurrence associations and semantic logical cohesion of the document content;
[0008] Determining a semantic association analysis result in the academic document text data set based on the semantic association feature representation, wherein the semantic association analysis result includes description information of the semantic association type and association strength between documents;
[0009] Based on the semantic association analysis result, analysis feedback information containing document association relationships is generated, and the analysis feedback information is output to a target document management system to support document content association retrieval operations.
[0010] On the other hand, an embodiment of the present invention also provides a semantic association analysis system based on academic literature, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.
[0011] Based on the above aspects, the embodiments of the present invention achieve in-depth mining and precise analysis of complex semantic associations between academic documents by comprehensively applying technical means such as basic semantic feature extraction, association feature modeling, and semantic association analysis. It can not only automatically extract the core concept features and contextual semantic relationship features of the document content, but also effectively capture the concept co-occurrence correlation and semantic logical connection between documents through pre-trained semantic association analysis models, and generate accurate semantic association feature representations. Based on the semantic association feature representation, it can further determine the semantic association type and association strength description information between documents, provide scientific researchers with intuitive and comprehensive document association analysis results, significantly improve the automation level and accuracy of semantic association analysis of academic documents, and help scientific researchers quickly discover potential connections between documents, grasp research trends, and promote academic innovation and knowledge dissemination. BRIEF DESCRIPTION OF THE DRAWINGS
[0012] Figure 1 It is a schematic diagram of the execution flow of the semantic association analysis method based on academic documents provided by an embodiment of the present invention.
[0013] Figure 2 Schematic diagram of exemplary hardware and software components of a semantic association analysis system based on academic documents provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0014] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 This is a flow chart of a semantic association analysis method based on academic documents provided by an embodiment of the present invention. The semantic association analysis method based on academic documents is introduced in detail below.
[0015] Step S110: Acquire an academic document text data set to be analyzed, wherein the academic document text data set includes a plurality of document text units with subject tags.
[0016] In this embodiment, the academic document text data set is composed of multiple document text units, and each document text unit is attached with a subject tag. The subject tag is a general identifier of the core content of the document. For example, in the field of medical research, the subject tags can be "cardiovascular disease treatment", "tumor immunotherapy", etc. Among them, these document text units can be obtained through various channels, such as academic databases, professional library websites, etc. From these data sources, document text units that meet the requirements are screened out according to the set search conditions. For example, in a well-known academic database, keywords and subject tags are used to search, and the document text units with the subject tag "cardiovascular disease treatment" that are searched are collected to form an academic document text data set. These document text units may come from different eras, different research institutions and authors, covering various research perspectives and results under this topic.
[0017] Step S120: performing basic semantic feature extraction processing on the academic document text data set to obtain basic semantic features of each document text unit, wherein the basic semantic features include core concept features and contextual semantic relationship features of the document content.
[0018] After obtaining the academic document text data set, the next step is to extract and process the basic semantic features to obtain the basic semantic features of each document text unit. The basic semantic features are mainly composed of core concept features and contextual semantic relationship features. The core concept features represent the most critical concepts and ideas in the document, while the contextual semantic relationship features reflect the logical relationship and semantic connection between these concepts in the document context. For example, in a document on "The application of artificial intelligence in medical imaging diagnosis", the core concept features may include "artificial intelligence algorithms", "medical imaging", "diagnostic models", etc., and the contextual semantic relationship features reflect the interaction between these concepts, such as how "artificial intelligence algorithms" are applied to the processing of "medical imaging" and then construct a "diagnostic model".
[0019] Step S121: performing text cleaning processing on the academic document text data set to remove formatting mark information and non-content symbol information in the document text unit to obtain a plain text data unit with a unified format.
[0020] When extracting basic semantic features, the first step is to perform text cleaning on the academic document text data set. During storage and transmission, document text units may contain some formatting markup information and non-content symbol information, which will interfere with subsequent semantic analysis and therefore need to be removed. Formatting markup information includes header and footer marks, chapter title formatting marks, and formula number marks, while non-content symbol information includes special symbols, hyperlink addresses, and footnote reference marks. The specific operations are as follows:
[0021] Step S1211: Identify formatting markup information in a document text unit, where the formatting markup information includes header and footer marks, chapter title formatting marks, and formula number marks.
[0022] In this step, the formatting markup information in the document text unit needs to be identified. Different documents may use different formatting markup methods, but common ones are specific symbol combinations or typesetting rules to represent headers, footers, chapter titles, and formula numbers. For example, headers and footers may appear at the top and bottom of the page in a specific font, font size, or position; chapter titles may use specific heading level marks, such as "###1. Introduction" and "###2. Related Work"; formula numbers may appear next to the formula in the form of brackets and numbers, such as "(1)" and "(2)". These formatting markup information can be identified through methods such as regular expression matching and text pattern recognition. Regular expressions are a powerful text matching tool that can find and identify specific content in the text based on preset patterns. For example, for the chapter title formatting markup, the regular expression "###\d+\.[\u4e00-\u9fa5a-zA-Z]+" can be used to match chapter titles that start with "###", followed by numbers and periods, and then followed by Chinese or English characters.
[0023] Step S1212: batch delete the formatting markup information using a regular expression matching method, and retain the text content information in the document text unit.
[0024] After identifying the formatting mark information, use the regular expression matching method to delete it in batches. Regular expressions can accurately find the formatting mark information according to the preset pattern and replace it with an empty string or other specified characters. For example, for the chapter title formatting mark identified earlier, use regular expression matching and replace it with an empty string. This will remove the chapter title formatting mark and retain only the text content information. During the processing, ensure that the regular expression pattern is set accurately to avoid accidental deletion of text content. At the same time, for some complex formatting marks, it may be necessary to use different regular expressions for matching and deletion multiple times to ensure that all formatting mark information is removed.
[0025] Step S1213: Identify non-content symbol information in the document text unit, where the non-content symbol information includes special symbols, hyperlink addresses, and footnote reference marks.
[0026] In addition to formatting markup information, document text units also contain some non-content symbol information, such as special symbols, hyperlink addresses, and footnote reference marks. Special symbols include various punctuation marks, mathematical symbols, emoticons, etc. Hyperlink addresses usually begin with "http: / / " or "https: / / ", and footnote reference marks may appear in the text in the form of numbers or other symbols. These non-content symbol information can be identified through character matching, string search, and other methods. For example, for hyperlink addresses, they can be identified by searching for strings that begin with "http: / / " or "https: / / "; for footnote reference marks, they can be identified by searching for specific numbers or symbol combinations in the text.
[0027] Step S1214: performing character-by-character detection processing on the non-content symbol information, and replacing the detected non-content symbol information with blank characters.
[0028] After identifying non-content symbols, they are processed character by character. This character-by-character detection ensures accurate identification and processing of each non-content symbol. Detected non-content symbols are replaced with whitespace. For example, special symbols can be replaced with spaces; hyperlink addresses can be replaced entirely with a single space; and footnote reference marks can also be replaced with spaces. During the replacement process, care must be taken to maintain text coherence to avoid incomplete text or semantic confusion caused by the replacement operation.
[0029] Step S1215: performing blank line merging processing on the processed document text unit, merging multiple consecutive blank lines into a single blank line, and obtaining a plain text data unit with a unified format.
[0030] After the previous processing, multiple consecutive blank lines may appear in the document text unit. To make the text format more uniform, the processed document text unit is subjected to blank line merging. This can be done by traversing each line of the text, detecting consecutive blank lines, and merging them into a single blank line. For example, if two consecutive blank lines are detected, one of them is deleted, leaving only one blank line. Through the above method, a uniformly formatted plain text data unit is ultimately obtained, providing clean, standardized text data for subsequent semantic analysis.
[0031] Step S122: performing word segmentation processing on the plain text data unit, extracting core term units and key phrase units in the document text unit, wherein the core term units include professional terms in the document research field, and the key phrase units include phrases describing the logical relationship between terms.
[0032] After obtaining uniformly formatted plain text data units, they are segmented to extract core term units and key phrase units from the document text units. Word segmentation involves dividing continuous text into individual words or phrases according to predefined rules. Core term units are specialized terms within the research field, representing the document's core concepts and research directions. Key phrase units contain phrases describing the logical relationships between terms, reflecting the semantic connections and logical structure between the core terms. For example, in a document on "gene editing technology," core term units might include "CRISPR / Cas9," "gene editing," and "off-target effects," while key phrase units might include "CRISPR / Cas9-based gene editing" and "methods for reducing off-target effects." Professional word segmentation tools, such as Jieba (Chinese) and NLTK (English), can be used to segment plain text data units. During the word segmentation process, the segmentation results should be screened and adjusted based on the research field and professional characteristics of the document to ensure that the correct core term units and key phrase units are extracted. For example, some professional terms in specific fields may need to be added to the dictionary of the word segmentation tool to improve the accuracy of word segmentation.
[0033] Step S123: Calling a pre-trained semantic encoding model to perform contextual semantic encoding processing on the core term unit and the key phrase unit to generate an initial semantic vector of the document text unit. The initial semantic vector includes the co-occurrence relationship between the term units and the logical connection relationship between the phrase units.
[0034] After extracting the core term units and key phrase units, a pre-trained semantic encoding model is used to perform contextual semantic encoding on them to generate the initial semantic vector for the document text unit. The pre-trained semantic encoding model is trained on a large amount of text data and can learn the semantic relationships between words and phrases. The initial semantic vector contains the co-occurrence relationships between term units and the logical cohesion relationships between phrase units. These logical cohesion relationships reflect the semantic information of the document text unit. The specific operation is as follows:
[0035] Step S1231: Arrange the core term units and key phrase units in the order of appearance in the document text unit to generate a term phrase sequence.
[0036] First, the extracted core term units and key phrase units are arranged in the order of their appearance in the document text unit to generate a term phrase sequence. This arrangement preserves the contextual order of terms and phrases in the original text, facilitating subsequent semantic encoding. For example, in a document on "quantum computing," after the core term units and key phrase units are arranged in the order of their appearance, the generated term phrase sequence might be "quantum bit," "quantum algorithm," "quantum algorithm design based on quantum bits," and so on.
[0037] Step S1232: Input the term phrase sequence into the word embedding layer of the semantic encoding model, and generate a corresponding word vector representation for each term unit and phrase unit, wherein the word vector representation contains basic semantic information of the term phrase.
[0038] The generated term phrase sequence is input into the word embedding layer of the semantic encoding model. The role of the word embedding layer is to convert each term unit and phrase unit into a corresponding word vector representation. A word vector is a numerical vector that can map the semantic information of words and phrases into a low-dimensional vector space. Each word vector representation contains the basic semantic information of the term phrase, and the position and distance of different word vectors in the vector space reflect the semantic similarity between them. For example, in the semantic encoding model, the word vectors of "qubit" and "qubit" may be close in the vector space because they have similar semantics.
[0039] Step S1233: performing sequence modeling processing on the word vector representation through the context encoder of the semantic encoding model, capturing the contextual dependencies between terms and phrases in the sequence, and generating a context encoding vector containing contextual semantic information.
[0040] After obtaining the word vector representation, it is processed through the context encoder of the semantic encoding model for sequence modeling. The context encoder can capture the contextual dependencies between terms and phrases in the sequence. That is, it considers the contextual information of terms and phrases in the original text to generate a context encoding vector that contains contextual semantic information. For example, in the sentence "Quantum computing utilizes the characteristics of quantum bits for high-speed calculations," the context encoding vector of "quantum bit" will take into account the contextual information such as "quantum computing" and "high-speed computing," thereby more accurately reflecting its semantic meaning in the sentence. Context encoders can adopt model structures such as recurrent neural networks (RNNs), long short-term memory networks (LSTMs), and gated recurrent units (GRUs). These models can effectively process sequential data and capture long-term dependencies within the sequence.
[0041] Step S1234: Perform attention weighting processing on the context encoding vector, assign attention weights according to the importance of term phrases in the document text unit, enhance the semantic representation of key term phrases, and weaken the semantic representation of secondary term phrases.
[0042] To highlight the semantic information of key term phrases, attention weighting is applied to the generated context encoding vector. The attention mechanism assigns attention weights based on the importance of the term phrase within the document text unit, with important term phrases receiving higher attention weights and less important term phrases receiving lower attention weights. This attention weighting process can enhance the semantic representation of key term phrases and weaken those of less important term phrases. For example, in a paper on "artificial intelligence algorithm optimization," key terms such as "genetic algorithm" and "particle swarm algorithm" might receive higher attention weights, while some auxiliary descriptive phrases might receive lower attention weights. The attention mechanism can be implemented in various ways, such as dot product attention and additive attention.
[0043] Step S1235: Perform dimension splicing processing on the attention-weighted context encoding vector to generate the initial semantic vector of the document text unit.
[0044] After performing attention weighting, the weighted context encoding vectors undergo dimensional concatenation. Dimension concatenation involves concatenating multiple vectors according to predefined rules to form a higher-dimensional vector. Through dimensional concatenation, the context encoding vectors of individual terms and phrases are integrated to generate an initial semantic vector for the document unit. This initial semantic vector incorporates the co-occurrence relationships between term units and the logical cohesion between phrase units, comprehensively reflecting the semantic information of the document unit.
[0045] Step S124: performing feature screening processing on the initial semantic vector, retaining semantic dimension information that is strongly related to the document subject tag, removing redundant semantic dimension information that is weakly related to the document subject tag, and obtaining the core concept features of the document text unit.
[0046] After generating the initial semantic vector of the document text unit, it is subjected to feature screening. The initial semantic vector may contain some redundant semantic dimension information that is weakly correlated with the document subject label. This information will increase the complexity of subsequent analysis and may interfere with the extraction of core concepts. Therefore, it is necessary to retain the semantic dimension information that is strongly correlated with the document subject label and remove redundant information to obtain the core concept features of the document text unit. Feature screening can be performed by calculating the correlation between each dimension of the initial semantic vector and the document subject label. The correlation calculation can use methods such as cosine similarity and Pearson correlation coefficient. For example, for a document about "new energy vehicles", its subject label is "new energy vehicle technology innovation". The dimensions related to "new energy" and "automotive technology" in the initial semantic vector may have a high correlation with the subject label, while the dimensions with other unrelated fields have a low correlation. Remove the dimensions with low correlation and retain the dimensions with high correlation to obtain the core concept features.
[0047] Step S125: Analyze the position distribution relationship of the core term units and key phrase units in the document text unit, and extract the contextual semantic relationship features of the document content. The contextual semantic relationship features include the proximity relationship of the term units within the paragraph and the progressive relationship of the phrase units between chapters.
[0048] In addition to the core concept features, it is also necessary to extract the contextual semantic relationship features of the document content. The contextual semantic relationship features can reflect the positional distribution relationship between term units and key phrase units in the document text unit, including the proximity relationship between term units within a paragraph and the progressive relationship between phrase units between chapters. For example, in a paragraph, adjacent term units may have closer semantic connections; in different chapters, the order of appearance of key phrase units may reflect the progressive relationship of the research content. The contextual semantic relationship features can be extracted by analyzing the positional information of core term units and key phrase units in the document text unit, such as the paragraphs, sentences, chapters, etc. in which they are located. For the proximity relationship between term units within a paragraph, the distance between term units can be calculated. The closer the distance, the stronger the proximity relationship; for the progressive relationship between phrase units between chapters, the order of appearance and logical connection of phrase units in different chapters can be analyzed.
[0049] Step S126: normalize the core concept features and the contextual semantic relationship features to generate basic semantic features with unified dimensional representation.
[0050] After obtaining the core concept features and contextual semantic relationship features, they need to be standardized. After standardization, the core concept features and contextual semantic relationship features are spliced together to generate basic semantic features with a unified dimensional representation. Basic semantic features contain the core concept features and contextual semantic relationship features of the document content and can comprehensively reflect the semantic information of the document text unit.
[0051] Step S130: calling a pre-trained semantic association analysis model to perform association feature modeling on the basic semantic features, generating a semantic association feature representation between document text units, wherein the semantic association feature representation includes the concept co-occurrence association degree and semantic logical cohesion degree of the document content.
[0052] After obtaining the basic semantic features of each document text unit, the pre-trained semantic association analysis model is called to model its association features. The semantic association analysis model is obtained through pre-training on a large amount of document data, and it can learn the semantic association patterns between documents. The basic semantic features are processed by this model to generate a semantic association feature representation between document text units. This representation includes the concept co-occurrence correlation and semantic logical cohesion of the document content. The concept co-occurrence correlation reflects the co-occurrence of core concepts in different documents, and the semantic logical cohesion reflects the semantic logical relationship between different documents. For example, in a group of documents about "big data analysis", if two documents both mention "data mining algorithm", then their concept co-occurrence correlation is high; if the research conclusion of one document is the basis for the research of another document, then their semantic logical cohesion is high.
[0053] Step S131: inputting the basic semantic features into the feature input layer of the semantic association analysis model, performing feature standardization processing, and obtaining standardized basic semantic features.
[0054] First, the obtained basic semantic features are input into the feature input layer of the semantic association analysis model. The main function of the feature input layer is to preprocess the input basic semantic features, including feature standardization. Feature standardization can unify the basic semantic features of different document text units to the same scale, avoiding the problem of unstable model training or inaccurate results caused by different feature scales. Feature standardization can be performed using a variety of methods, such as z-score standardization and minimum-maximum standardization. Z-score standardization converts the feature value into a standard normal distribution by calculating the mean and standard deviation of the feature; minimum-maximum standardization scales the feature value to a fixed interval, such as [0, 1]. Through feature standardization, standardized basic semantic features are obtained, providing unified input data for subsequent association feature modeling.
[0055] Step S132: Performing intra-document local association analysis on the standardized basic semantic features through the local association modeling layer of the semantic association analysis model, extracting the local association pattern of the core concept features and the contextual semantic relationship features in each document text unit, wherein the local association pattern includes the importance weight of the concept features in the contextual relationship.
[0056] After obtaining the standardized basic semantic features, the local association modeling layer of the semantic association analysis model is used to perform local association analysis within the document. The purpose of the local association modeling layer is to extract the local association pattern between the core concept features and the contextual semantic relationship features in each document text unit. The specific operation is as follows:
[0057] Step S1321: Separate the core concept features and contextual semantic relationship features from the standardized basic semantic features of each document text unit.
[0058] The standardized basic semantic features obtained after feature standardization processing contain two parts of information: core concept features and contextual semantic relationship features. In order to perform subsequent local association analysis, these two parts of features need to be separated. Since the representation or positional relationship of the core concept features and contextual semantic relationship features in the standardized basic semantic features has been clarified in the previous steps, they can be separated based on this information. For example, if the standardized basic semantic feature is in a vector form, and it is known that the first part of the dimension represents the core concept feature and the second part of the dimension represents the contextual semantic relationship feature, then the two parts of the feature can be extracted separately according to the dimension division rule.
[0059] Step S1322: Calculate the dot product similarity between each concept dimension in the core concept feature and each relationship dimension in the contextual semantic relationship feature to generate a similarity matrix.
[0060] After separating the core concept features and contextual semantic relationship features, the similarity between them needs to be calculated. Here, the dot product similarity calculation method is adopted. For each concept dimension in the core concept feature, a dot product operation must be performed with each relational dimension in the contextual semantic relationship feature. The dot product operation can measure the similarity between two vectors. In this case, it measures the semantic similarity between the concept dimension and the relational dimension. By performing dot product operations on all concept dimensions and relational dimensions, the results are arranged into a similarity matrix. For example, if the core concept feature has m concept dimensions and the contextual semantic relationship feature has n relational dimensions, then the generated similarity matrix is a matrix with m rows and n columns, and each element in the matrix represents the dot product similarity between a concept dimension and a relational dimension.
[0061] Step S1323: performing row normalization processing on the similarity matrix to obtain the importance weight of each concept dimension in different relationship dimensions.
[0062] After obtaining the similarity matrix, row normalization is required to clarify the importance of each concept dimension across different relationship dimensions. Row normalization processes each row of the matrix so that the sum of the elements is 1. Through row normalization, the similarity of each concept dimension across different relationship dimensions can be converted into importance weights. After normalization, the value of each row element represents the relative importance of that concept dimension across each relationship dimension. For example, for a row corresponding to a concept dimension, the larger the value of an element, the more important that concept dimension is in the corresponding relationship dimension.
[0063] Step S1324: performing column normalization processing on the similarity matrix to obtain the coverage weight of each relationship dimension on different concept dimensions.
[0064] In addition to calculating the importance weight of each concept dimension on different relationship dimensions, it is also necessary to calculate the coverage weight of each relationship dimension on different concept dimensions. This requires column normalization of the similarity matrix. Column normalization is to process each column element of the matrix so that the sum of each column element is 1. Through column normalization, the similarity of each relationship dimension on different concept dimensions can be converted into coverage weight. After normalization, the value of each column element represents the relative coverage of the relationship dimension on each concept dimension. For example, for a column corresponding to a certain relationship dimension, the larger the value of a certain element, the wider the coverage of the relationship dimension on the corresponding concept dimension.
[0065] Step S1325: Based on the importance weight and the coverage weight, a local association pattern of the core concept features and the contextual semantic relationship features in the document text unit is constructed, and the local association pattern includes the relationship importance distribution of the concept dimension and the concept coverage distribution of the relationship dimension.
[0066] After obtaining the importance weight of each concept dimension in different relationship dimensions and the coverage weight of each relationship dimension in different concept dimensions, we can construct a local association pattern between the core concept features and the contextual semantic relationship features in the document text unit. This local association pattern combines the relationship importance distribution of the concept dimension and the concept coverage distribution of the relationship dimension. The relationship importance distribution of the concept dimension describes the importance of each concept dimension in different contextual relationships, and the concept coverage distribution of the relationship dimension describes the coverage of each contextual relationship for different concept dimensions. Through this local association pattern, we can clearly understand the local association between the core concept features and the contextual semantic relationship features within the document, such as which concepts are more important in certain contexts, which contextual relationships cover more concepts, etc.
[0067] Step S133: The standardized basic semantic features are subjected to inter-document global association analysis processing through the global association modeling layer of the semantic association analysis model, and the similarity matching degree of the core concept features between different document text units and the connection consistency of the contextual semantic relationship features are calculated to obtain the global association matching results between the documents.
[0068] After completing the local correlation analysis within a document, further analysis is needed to determine the global correlation between different document text units. The global correlation modeling layer of the semantic correlation analysis model is responsible for performing global correlation analysis between documents using standardized basic semantic features. For the core concept features between different document text units, the similarity matching degree is calculated. Various methods can be used to calculate similarity matching, such as cosine similarity and Euclidean distance. Cosine similarity measures the cosine of the angle between two vectors. A value closer to 1 indicates greater similarity between the two vectors, which in turn indicates greater similarity between the core concept features of the two documents. Euclidean distance calculates the spatial distance between two vectors. A smaller distance indicates greater similarity between the core concept features of the two documents. For contextual semantic relationship features, their cohesion consistency is calculated. This can be achieved by analyzing the logical order and semantic coherence of the contextual relationships across different documents. For example, if the contextual relationships of one document are sequentially developed according to the research steps, and another document discussing the same topic follows a similar logical order, then the contextual semantic relationship features of the two documents can be considered to have a high degree of cohesion consistency. By calculating the similarity matching degree of core concept features and the cohesion consistency of contextual semantic relationship features, the global correlation matching results between documents are finally obtained.
[0069] Step S134: Input the local association pattern and the global association matching result into the feature fusion layer of the semantic association analysis model, dynamically assign the fusion weight of the local association pattern and the global association matching result based on the attention mechanism, and generate a fusion association feature vector.
[0070] After obtaining the local correlation patterns within a document and the global correlation matching results between documents, these two pieces of information need to be fused. The feature fusion layer of the semantic correlation analysis model is responsible for completing this fusion task. An attention mechanism is used to dynamically assign fusion weights between the local correlation patterns and the global correlation matching results. The attention mechanism automatically adjusts the importance of local correlation patterns and global correlation matching results in the fusion process based on different situations. For example, in some cases, local correlation information within a document may be more important, so the attention mechanism assigns a higher weight to the local correlation patterns; in other cases, global correlation information between documents may be more critical, so the attention mechanism assigns a higher weight to the global correlation matching results. By dynamically assigning fusion weights, the local correlation patterns and global correlation matching results are weighted and concatenated to generate a fused correlation feature vector. This fused correlation feature vector combines the local correlation information within a document and the global correlation information between documents.
[0071] Step S135: Perform feature dimensionality reduction processing on the fused association feature vector, retain key dimensional information that can characterize the semantic association between documents, and generate semantic association feature representations between document text units, wherein the semantic association feature representations include concept co-occurrence association degree and semantic logical cohesion degree.
[0072] The generated fused correlation feature vector may have a high dimensionality and contain some redundant information that is not very useful for characterizing the semantic associations between documents. To simplify the data and highlight key information, the fused correlation feature vector needs to be subjected to feature dimensionality reduction. Feature dimensionality reduction can be achieved using various methods, such as principal component analysis (PCA) and linear discriminant analysis (LDA). PCA projects high-dimensional data into a low-dimensional space by identifying the principal components while preserving the data's variance as much as possible. LDA reduces dimensionality by finding the projection direction that maximizes the separation between different categories. During feature dimensionality reduction, key dimensional information that characterizes the semantic associations between documents is retained. After dimensionality reduction, a semantic correlation feature representation is generated between document text units. This semantic correlation feature representation includes concept co-occurrence correlation and semantic logical cohesion. Concept co-occurrence correlation reflects the degree of co-occurrence of core concepts across different documents, while semantic logical cohesion reflects the degree of semantic logical coherence between different documents.
[0073] Step S140: Determine the semantic association analysis result in the academic document text data set based on the semantic association feature representation, wherein the semantic association analysis result includes semantic association type and association strength description information between documents.
[0074] After obtaining the semantic association feature representation between document text units, it is necessary to further determine the semantic association analysis results in the academic document text data set. The semantic association analysis results include the semantic association type and association strength description information between documents.
[0075] Step S141: parsing the concept co-occurrence correlation and semantic logical cohesion in the semantic association feature representation, and extracting the correlation value distribution information between the document text units.
[0076] First, the semantic association feature representation is parsed to extract information about concept co-occurrence correlation and semantic logical cohesion. These values are present in the semantic association feature representation as numerical values. These values are extracted and organized into correlation numerical distribution information. This distribution information describes the distribution of correlations between different document units, such as which documents have high and low correlations.
[0077] Step S142: performing cluster analysis on the correlation value distribution information to identify document text unit groups with strong correlation and document text unit groups with weak correlation.
[0078] This step aims to distinguish between strongly and weakly associated document text unit groups based on the distribution of correlation values through cluster analysis. The following are the specific sub-steps and detailed descriptions:
[0079] Step S1421: Select the density clustering algorithm as the cluster analysis method, and set the cluster radius parameter and the minimum sample number parameter.
[0080] Among many clustering algorithms, density clustering algorithms cluster data based on the density distribution of data points. They are suitable for processing data with varying density regions and are particularly useful for data with potentially irregular distributions of association values. The cluster radius parameter determines the neighborhood of each data point, and its size affects the number of data points within the neighborhood. The minimum number of samples parameter specifies the minimum number of neighborhood data points required for a data point to qualify as a core point. The settings of these two parameters need to be adjusted based on the distribution characteristics of the association values. If the association values are more dispersed, a larger cluster radius and a smaller minimum number of samples may be required; conversely, a smaller cluster radius and a larger minimum number of samples may be required. For example, within a research field, if the distribution of association values varies significantly between papers, it may be necessary to flexibly adjust these two parameters to achieve accurate clustering results.
[0081] Step S1422: taking the association value of each pair of document text units in the association value distribution information as a sample point, and constructing a sample point set.
[0082] The correlation value for each pair of document text units is the basic data unit for cluster analysis. These correlation values are considered sample points, and all sample points are integrated together to form a sample point set. This sample point set is the basis for subsequent clustering operations and contains information about the correlation between all pairs of documents. For example, in an academic document text dataset containing multiple medical articles, every two documents have a correlation value. Collecting these values to form a sample point set will cover the correlation information of all document pairs in the entire document collection.
[0083] Step S1423: Calculate the Euclidean distance between each sample point and other sample points in the sample point set, and determine the neighborhood sample points of each sample point.
[0084] Euclidean distance is a commonly used distance measurement method that can measure the distance between two sample points in space. For each sample point in the sample point set, the Euclidean distance between it and all other sample points is calculated. According to the pre-set cluster radius parameter, the neighborhood range of each sample point is determined, and the sample points within the neighborhood are the neighborhood sample points of the sample point. By calculating the Euclidean distance and determining the neighborhood sample points, the data distribution around each sample point can be understood. For example, in a sample point set composed of correlation values of literature in different disciplines, after calculating the Euclidean distance between each sample point and other sample points, it can be known which sample points are close to each other and may belong to the same cluster group.
[0085] Step S1424: Determine the type of sample points according to the number of the neighborhood sample points, where the types of sample points include core points, boundary points, and noise points.
[0086] Based on the number of sample points in its neighborhood, sample points can be divided into core points, boundary points, and noise points. If the number of sample points in the neighborhood of a sample point is greater than or equal to the minimum number of sample points parameter, then the sample point is a core point. The core point is the center of the cluster and represents a data area with higher density. If the number of sample points in the neighborhood of a sample point is less than the minimum number of sample points parameter, but it is located in the neighborhood of a core point, then the sample point is a boundary point. Boundary points surround the core point. If a sample point is neither a core point nor a boundary point, then it is a noise point. Noise points are usually isolated data points that do not belong to any cluster group. For example, in a set of sample points about the relevance of scientific and technological literature, some sample points are surrounded by many neighboring sample points, which are core points; some sample points are surrounded by fewer neighboring sample points but are close to core points, which are boundary points; and those isolated sample points are noise points.
[0087] Step S1425: Merge the core point and its neighborhood sample points into a cluster group, assign the boundary points to the cluster group to which the nearest core point belongs, and exclude noise points.
[0088] After determining the type of sample point, the core point and its neighboring sample points are merged into a cluster group. The core point serves as the center of the cluster, and its neighboring sample points have a strong correlation with the core point, so they are grouped together. For boundary points, since they are close to a core point, they are assigned to the cluster group to which the nearest core point belongs. This makes the clustering results more complete. Noise points, however, have a weak correlation with other data points and have little impact on the clustering results, so they are excluded. For example, in a sample point set containing the correlation between literature in different research directions in computer science, each core point, its neighboring sample points, and the corresponding boundary points are merged into different cluster groups. After excluding the noise points, a clear clustering result is obtained.
[0089] Step S1426: Based on the average value of the correlation values of the sample points in the cluster group, the cluster group whose average value of the correlation value is higher than the preset threshold is marked as a document text unit group with strong correlation, and the cluster group whose average value of the correlation value is lower than the preset threshold is marked as a document text unit group with weak correlation.
[0090] The preset threshold is the boundary that distinguishes strong associations from weak associations. Calculate the average value of the correlation values of the sample points in each cluster group and compare the average value with the preset threshold. If the average value is higher than the preset threshold, it means that the correlation between the document text units in the cluster group is high, and it is marked as a document text unit group with strong correlation; conversely, if the average value is lower than the preset threshold, it is marked as a document text unit group with weak correlation. The setting of the preset threshold needs to take into account the research purpose and data characteristics. For example, in an association analysis of biomedical research literature, if you want to more strictly screen out document groups with strong correlation, you can appropriately increase the preset threshold.
[0091] Step S143: Analyze the distribution characteristics of concept co-occurrence correlation and semantic logical cohesion in the document text unit group with strong correlation, and determine the semantic correlation type between documents, which includes concept extension correlation type and logical argumentation correlation type.
[0092] After identifying strongly correlated groups of document text units, we further analyze the distribution characteristics of concept co-occurrence correlation and semantic logical cohesion within these groups. Regarding concept co-occurrence correlation, if a large number of the same core concepts frequently appear in certain documents, and these concepts are further expanded and deepened in different documents, then we can determine that these documents have a concept extension type of correlation. For example, if core concepts such as "word vector model" and "attention mechanism" frequently appear in a series of documents on "artificial intelligence natural language processing," and different documents study and expand on these concepts from different perspectives, then these documents have a concept extension type of correlation. Regarding semantic logical cohesion, if there is a clear logical deduction and argumentation relationship between certain documents, for example, if one document proposes a theoretical hypothesis and another document verifies the hypothesis through experiments, then we can determine that these documents have a logical argumentation type of correlation. By analyzing the concept co-occurrence correlation and semantic logical cohesion within strongly correlated document text units, we can accurately determine the semantic correlation type between the documents.
[0093] Step S144: Calculate the average and variance of the correlation values of each pair of document text units in the document text unit group with strong correlation, and generate correlation strength description information, which includes correlation central tendency parameter and dispersion degree parameter.
[0094] To more comprehensively describe the strength of association between documents in a group of strongly associated document text units, the mean and variance of the association values for each pair of document text units were calculated. The mean reflects the central tendency of the association, that is, the overall level of association between documents in the group; the variance reflects the dispersion of the association, that is, the fluctuation of the association between documents in the group. For example, if the mean is high and the variance is small, it means that the association between most documents in the group is high and relatively stable; if the mean is high but the variance is large, it means that there are large differences in the association between documents in the group. The mean and variance values are used to generate descriptive information on the strength of association. The descriptive information on the strength of association includes the central tendency parameter of the association (i.e., the mean value) and the dispersion parameter (i.e., the variance value).
[0095] Step S145: combining the semantic association type and the association strength description information to generate a semantic association analysis result in the academic document text data set.
[0096] The determined semantic association types and the generated association strength description information are combined to generate semantic association analysis results for the academic literature text dataset. The semantic association analysis results combine the semantic association types (such as concept extension association types and logical argument association types) and association strength description information (such as the association central tendency parameter and dispersion parameter) between documents. Through this semantic association analysis result, we can clearly understand the semantic associations between different documents in the academic literature text dataset, including their association types and the strength of the associations.
[0097] Step S150: generating analysis feedback information including document association relationships based on the semantic association analysis results, and outputting the analysis feedback information to a target document management system to support document content association retrieval operations.
[0098] After obtaining the semantic association analysis results, it is necessary to generate analysis feedback information containing document association relationships based on the results.
[0099] Step S151: extracting semantic association type and association strength description information from the semantic association analysis result, and constructing structured description data of document association relationships, wherein the structured description data includes an identifier of an associated document pair and a corresponding association type label.
[0100] Semantic association type and association strength descriptive information are extracted from the semantic association analysis results. Based on this information, structured descriptive data of document association relationships is constructed. This structured descriptive data records the association relationships between documents in a standardized manner, including associated document pair identifiers and corresponding association type labels. The associated document pair identifiers are used to uniquely identify each pair of associated documents, while the association type labels specify the semantic association type between the pair of documents, such as the concept extension association type or the logical argument association type. By constructing structured descriptive data, complex document association relationships can be represented in a clear and orderly manner.
[0101] Step S152: Perform visual conversion processing on the structured description data to convert the document association relationship into a visual association graph containing nodes and edges, where the nodes represent document text units, the edges represent semantic association relationships between documents, and the attributes of the edges include association type labels and association strength description information.
[0102] This step converts the structured document association data into an intuitive visual association graph. The following are the specific sub-steps and detailed descriptions:
[0103] For example, step S1521: create a node for each document text unit and create an edge for each document association relationship.
[0104] In a visual association graph, nodes are intuitive representations of document text units, with each node representing a document. Edges represent semantic relationships between documents, with each pair of associated documents connected by an edge. By constructing these nodes and edges, the basic framework for a visual association graph has been established. For example, in an analysis of the associations between documents in different branches of physics, each document corresponds to a node, and the associations between documents correspond to an edge. This allows complex document associations to be presented graphically.
[0105] Step S1522: configuring node attribute information for the node, wherein the node attribute information includes document title, subject tag and publication time information.
[0106] Node attribute information can enrich the information of the document represented by the node. The document title allows users to quickly identify the main content of the document; the subject tag clarifies the subject area of the document, helping users to grasp the research direction of the document from a macro perspective; the publication time information can reflect the timeliness of the document, which is very important for users who need to pay attention to the latest research results. In the visual map, when users interact with the nodes, this attribute information can be displayed. For example, in a visual association map about new energy research documents, users can click on a node to see the title, subject tag (such as solar energy, wind energy, etc.) and publication time of the document, thereby better understanding the basic situation of the document.
[0107] Step S1523: configuring edge attribute information for the edge, where the edge attribute information includes an association type label and association strength description information.
[0108] Edge attribute information further describes the association relationship between documents. The association type label clarifies the semantic association type between documents. For example, the concept extension association type indicates that one document expands a concept based on another document; the logical argument association type indicates that one document provides logical support for the viewpoint of another document. The association strength description information reflects the closeness of the association between documents and can be expressed in terms of association values. By configuring these attribute information for edges, users can gain a deeper understanding of the nature and strength of the association between documents. For example, in a visual association map of artificial intelligence algorithm research documents, the edge association type label may be displayed as "algorithm improvement association", and the association strength description information can use a value to indicate the closeness of the association between two documents in terms of algorithm improvement.
[0109] Step S1524: color-code the nodes according to the subject labels of the document text units. Nodes with the same subject label use the same color, and nodes with different subject labels use different colors.
[0110] Color coding is an intuitive visualization method that allows users to quickly identify documents with the same subject label by distinguishing by color. Nodes with the same subject label use the same color, so that in the atlas, document nodes with the same research topic will be clustered together with the same color, making it easier for users to grasp the overall distribution of documents with different topics. Nodes with different subject labels use different colors to clearly distinguish the differences between different topics. For example, in a visual association map covering documents in multiple disciplines, document nodes in the field of computer science can be represented by blue, and document nodes in the field of biology can be represented by green. This allows users to see the distribution and association of documents in different disciplines at a glance.
[0111] Step S1525: performing width coding processing on the edges according to the association strength description information. The width of the edge with higher association strength is larger, and the width of the edge with lower association strength is smaller.
[0112] Edge width encoding can intuitively demonstrate differences in the strength of associations between documents. Edges with strong associations have larger widths and are more prominent in the graph, indicating a close connection between the two documents. Edges with weak associations have smaller widths and are less noticeable, indicating a weaker connection. This allows users to quickly determine the strength of associations between documents based on edge width. For example, in a visual association graph of financial market research papers, the edge between two closely related papers on market trend forecasting would be wider, while the edge between papers with weaker associations would be narrower.
[0113] Step S1526: Arrange the nodes and edges in space according to the force-directed layout algorithm to generate a visual association graph containing the nodes and edges.
[0114] The force-directed layout algorithm, based on the principles of physics, treats nodes as charged particles and edges as springs. Nodes repel each other, while edges attract each other. By iteratively adjusting the positions of nodes, the entire visual association graph reaches a state of equilibrium. In this process, closely connected nodes move closer together, while weakly connected nodes move further away. The resulting visual association graph intuitively displays the association structure and density relationships between documents. For example, in a visual association graph of literature research documents, the force-directed layout algorithm clusters nodes belonging to the same literary genre, while nodes belonging to different literary genres are relatively dispersed, forming a clear document association network.
[0115] Step S153: configuring an interactive operation interface for the visual association graph, wherein the interactive operation interface supports click query operations on nodes and screening and filtering operations on edges.
[0116] To enhance the practicality and interactivity of the visual association graph, an interactive interface is configured. This interface supports click-to-query operations on nodes. When a user clicks a node, detailed information about the corresponding document, such as title, subject tags, publication date, and core content summary, pops up, allowing the user to further understand the specific details of the document. The interactive interface also supports edge filtering. Users can filter edges based on conditions such as association type label and association strength, displaying only those relationships that meet specific criteria. This allows the user to focus on document relationships of interest and improves information retrieval efficiency.
[0117] Step S154: performing data encapsulation processing on the structured description data and the visual association graph to generate analysis feedback information including document association relationships.
[0118] The structured description data and the visual association graph are encapsulated. Data encapsulation is the process of integrating different types of data into a unified data package. Through data encapsulation, the structured description data and the visual association graph are combined into a complete analysis feedback information containing the document association relationships. This analysis feedback information contains both the structured data of the document association relationships and the intuitive visual graph, providing users with multi-dimensional document association information.
[0119] Step S155: Perform format adaptation processing on the analysis feedback information to make it conform to the data input format requirements of the target document management system.
[0120] After generating analysis feedback information, it needs to be output to the target document management system to support document content-related retrieval operations. Because different document management systems may have different data input format requirements, the analysis feedback information needs to be format-adapted. According to the target document management system's data input format specifications, the analysis feedback information is converted and adjusted accordingly, such as changing the data storage format and adjusting the data field order. This format adaptation ensures that the analysis feedback information can be smoothly input into the target document management system.
[0121] During the entire process, if privacy-sensitive data is collected, such as when acquiring academic literature text datasets, it may contain personal information of the author. Data desensitization can be used to protect and prevent the leakage of this privacy-sensitive data. Data desensitization techniques include substitution, masking, and encryption. Substitution replaces sensitive data with a virtual value, such as replacing the author's real name with an anonymous number; masking partially hides sensitive data, such as displaying only the last four digits of the author's mobile phone number; and encryption uses an encryption algorithm to encrypt sensitive data, making it accessible only to authorized users. These techniques can effectively protect privacy and prevent leakage. In addition to data desensitization, access control can also be used to strictly manage access to privacy-sensitive data, ensuring that only authorized users have access to data of a specific level. Furthermore, during data transmission, encrypted transmission protocols such as SSL / TLS should be used to prevent data theft or tampering during transmission. For servers storing privacy-sensitive data, secure storage systems should be used, and regular data backups and security audits should be conducted to promptly identify and address potential security risks.
[0122] When it comes to building and training AI models, pre-trained semantic encoding models and semantic association analysis models play a key role in the overall technical solution. Essential modules for the semantic encoding model include a word embedding layer, a context encoder, and an attention mechanism module. The word embedding layer is responsible for converting the input core term units and key phrase units into corresponding word vector representations, which serves as the basis for subsequent semantic encoding. The context encoder typically employs a recurrent neural network structure such as an LSTM or GRU to capture the contextual dependencies between terms and phrases in a sequence, generating a context encoding vector containing contextual semantic information. The attention mechanism module then weights the context encoding vector to highlight the semantic representation of key terms and phrases.
[0123] The hierarchical structure of the semantic encoding model is as follows: the input layer receives a sequence of term phrases generated by arranging core term units and key phrase units in sequence; the word embedding layer converts the input term phrases into word vectors; the context encoder performs sequence modeling on the word vectors; the attention mechanism module weights the context encoding vectors; and finally, the attention-weighted context encoding vectors are output, which are then dimensionally spliced to obtain the initial semantic vector of the document text unit.
[0124] When training a semantic encoding model, the training data can be a large amount of academic literature text data. This data needs to be preprocessed similarly to the previous steps, including text cleaning, word segmentation, and other operations. The specific steps of training are as follows: First, the preprocessed text data is divided into a training set, a validation set, and a test set. The training set is used to learn the model's parameters, the validation set is used to adjust the model's hyperparameters, and the test set is used to evaluate the model's final performance. Set training parameters such as the learning rate, batch size, and number of training rounds. The learning rate controls the step size of the model's parameter updates in each iteration, the batch size determines the number of data samples used in each training, and the number of training rounds indicates the number of times the entire training data is traversed by the model.
[0125] During training, stochastic gradient descent (SGD) or its variants, such as the Adam optimization algorithm, is used to update the model's parameters. Each time, a batch of data is taken from the training set and fed into the model. The model's output is calculated through forward propagation. Then, based on the loss function between the output and the true label, the gradient is calculated through backpropagation. Finally, the model parameters are updated based on the gradient. This process is repeated until the model's performance on the validation set reaches stability or meets a preset stopping condition.
[0126] The essential modules of the semantic association analysis model include the feature input layer, local association modeling layer, global association modeling layer, feature fusion layer, and feature dimensionality reduction layer. The feature input layer normalizes the basic semantic features of the input; the local association modeling layer extracts the local association patterns between the core concept features and the contextual semantic relationship features within the document; the global association modeling layer calculates the global association matching results between different document text units; the feature fusion layer fuses the local association patterns and global association matching results based on the attention mechanism; and the feature dimensionality reduction layer reduces the dimensionality of the fused feature vector to generate a semantic association feature representation between document text units.
[0127] The hierarchical structure of the semantic association analysis model is as follows: the input layer receives standardized basic semantic features; the local association modeling layer and the global association modeling layer perform local and global association analysis respectively; the feature fusion layer fuses the results of the two; the feature dimensionality reduction layer reduces the dimensionality of the fusion result; and finally outputs the semantic association feature representation between document text units.
[0128] When training a semantic association analysis model, the training data is academic literature text data that has been processed using basic semantic feature extraction. The training steps are similar to those for the semantic encoding model, requiring the data to be divided into training, validation, and test sets. Training parameters such as the learning rate, batch size, and number of training rounds are set. The model parameters are updated using an appropriate optimization algorithm. During training, the loss function can be designed based on the characteristics of semantic association feature representation. For example, the loss function can be constructed by combining the difference between the actual values of concept co-occurrence correlation and semantic logical cohesion and the model's predicted values.
[0129] When applying these AI models in specific fields or scenarios, such as academic document management and retrieval, the model's input data is preprocessed and feature-extracted academic document text data, including the basic semantic features of each document's text unit. The model's output data is a representation of the semantic association features between document text units, along with the semantic association analysis results generated based on this, and analysis feedback information containing document association relationships. The inherent relationship between these input and output data lies in the fact that the model processes the basic semantic features of the input, extracts semantic association information between documents, and ultimately outputs it in the form of analysis feedback information, providing support for the document management system's associated retrieval operations.
[0130] Regarding data collection, ensure that the data collection process complies with laws, regulations, and ethical standards. Academic literature text data can be collected through legitimate academic database interfaces. When acquiring data, authorization from the database owner must be obtained and compliance with the database's terms of use is required. Furthermore, during the data collection process, preliminary screening and filtering are performed to remove documents that clearly do not meet the requirements, such as duplicates and low-quality documents.
[0131] In terms of tag management, accurate subject tags are assigned to each unit of document text. This can be accomplished through a combination of manual and automatic tagging. Manual tagging is performed by domain experts to ensure accuracy and professionalism. Automatic tagging uses machine learning algorithms to automatically assign subject tags to documents by extracting and classifying document text features. Furthermore, a tag management system is established to uniformly manage and maintain subject tags, ensuring consistency and standardization.
[0132] In terms of rule setting, clear rules should be established to guide the entire semantic association analysis process. For example, during feature extraction, word segmentation rules should be set to determine which words should be extracted as core term units and key phrase units. During cluster analysis, the value range and adjustment rules for the cluster radius parameter and the minimum sample number parameter should be set. These rules should be adjusted and optimized based on the specific application scenario and data characteristics.
[0133] In terms of recommendation decisions, the system provides users with recommendations for document-related searches based on the generated semantic association analysis results and analysis feedback. For example, when a user searches for a document in the document management system, the system can recommend related documents based on the semantic associations between that document and other documents. The recommendation decision rules can be configured based on the association type and association strength, prioritizing documents with strong associations and specific association types (such as concept extension associations).
[0134] Throughout the implementation of the technical solution, performance evaluation and optimization must be continuously conducted. Model performance is assessed by evaluating the model's output, such as accuracy, recall, and F1 score. Based on the evaluation results, adjustments and optimizations are made to the model's parameters, structure, or training process to improve performance and accuracy. Simultaneously, the entire semantic association analysis process is optimized, such as by adjusting feature extraction methods and optimizing cluster analysis algorithms, to enhance the efficiency and effectiveness of the entire technical solution.
[0135] In practical applications, further expansion and utilization of analysis feedback information can be considered. For example, analysis feedback information can be combined with the citation relationships of documents to construct a more comprehensive literature knowledge graph. The literature knowledge graph can intuitively display the semantic associations and citation relationships between documents, providing academic researchers with deeper knowledge discovery and research support. Furthermore, intelligent retrieval systems based on the literature knowledge graph can be developed to support more complex literature association retrieval operations, such as multi-condition retrieval based on semantic associations and citation relationships, and association path queries.
[0136] To store and manage analytical feedback, a dedicated database is needed to store structured descriptive data containing document relationships and visualized association maps. Database design should consider data storage efficiency, query efficiency, and data security. Either a relational or non-relational database, such as MySQL or MongoDB, can be used, depending on the data characteristics and application requirements. Within the database, appropriate table structures should be established for different types of data to ensure orderly data organization and management.
[0137] Figure 2 The following diagram illustrates exemplary hardware and software components of a system 100 for analyzing semantic associations based on academic documents, which can implement the concepts of the present application, according to some embodiments of the present application. For example, the processor 120 can be used in the system 100 for analyzing semantic associations based on academic documents and performing the functions described in the present application.
[0138] The academic document-based semantic association analysis system 100 can be a general-purpose server or a special-purpose server, both of which can be used to implement the academic document-based semantic association analysis method of this application. Although only one server is shown in this application, for convenience, the functions described in this application can be implemented in a distributed manner on multiple similar platforms to balance the processing load.
[0139] For example, the semantic association analysis system 100 based on academic documents may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and storage media 140 in different forms, such as a disk, ROM, or RAM, or any combination thereof. Exemplarily, the semantic association analysis system 100 based on academic documents may also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The method of the present application can be implemented according to these program instructions. The semantic association analysis system 100 based on academic documents also includes an I / O interface 150 between the computer and other input and output devices.
[0140] For ease of explanation, only one processor is described in the semantic association analysis system 100 based on academic documents. However, it should be noted that the semantic association analysis system 100 based on academic documents in the present application may also include multiple processors, so the steps performed by one processor described in the present application may also be performed jointly or individually by multiple processors. For example, if the processor of the semantic association analysis system 100 based on academic documents executes step A and step B, it should be understood that step A and step B may also be executed jointly by two different processors or individually in one processor. For example, the first processor executes step A, the second processor executes step B, or the first processor and the second processor execute steps A and B together.
[0141] In addition, an embodiment of the present invention further provides a readable storage medium, in which computer-executable instructions are preset. When a processor executes the computer-executable instructions, the above-mentioned semantic association analysis method based on academic literature is implemented.
[0142] It should be noted that in order to simplify the description of the present invention and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, multiple features are sometimes combined into one embodiment, figure or description thereof.
Claims
1. A semantic association analysis method based on academic literature, characterized in that: The method comprises: Obtaining an academic document text data set to be analyzed, wherein the academic document text data set includes a plurality of document text units with subject tags; Performing basic semantic feature extraction on the academic document text data set to obtain basic semantic features of each document text unit, wherein the basic semantic features include core concept features and contextual semantic relationship features of the document content; Calling a pre-trained semantic association analysis model to perform association feature modeling on the basic semantic features to generate semantic association feature representations between document text units, wherein the semantic association feature representations include concept co-occurrence associations and semantic logical cohesion of the document content; Determining a semantic association analysis result in the academic document text data set based on the semantic association feature representation, wherein the semantic association analysis result includes description information of the semantic association type and association strength between documents; Generating analysis feedback information containing document association relationships based on the semantic association analysis results, and outputting the analysis feedback information to a target document management system to support document content association retrieval operations; The calling of the pre-trained semantic association analysis model to perform association feature modeling processing on the basic semantic features to generate semantic association feature representations between document text units includes: Inputting the basic semantic features into the feature input layer of the semantic association analysis model and performing feature standardization processing to obtain standardized basic semantic features; Performing document local association analysis on the standardized basic semantic features through the local association modeling layer of the semantic association analysis model, extracting the local association pattern of the core concept features and the contextual semantic relationship features in each document text unit, wherein the local association pattern includes the importance weight of the concept features in the contextual relationship; The global association modeling layer of the semantic association analysis model performs a global association analysis on the standardized basic semantic features between documents, calculates the similarity matching degree of the core concept features between different document text units and the cohesion consistency of the contextual semantic relationship features, and obtains the global association matching results between documents; Inputting the local association pattern and the global association matching result into the feature fusion layer of the semantic association analysis model, dynamically allocating the fusion weight of the local association pattern and the global association matching result based on the attention mechanism, and generating a fused association feature vector; Performing feature dimensionality reduction processing on the fused association feature vector, retaining key dimensional information that can characterize the semantic association between documents, and generating semantic association feature representations between document text units, wherein the semantic association feature representations include concept co-occurrence association degree and semantic logical cohesion degree; Determining the semantic association analysis result in the academic document text data set according to the semantic association feature representation includes: Analyzing the concept co-occurrence correlation and semantic logical cohesion in the semantic correlation feature representation, and extracting correlation numerical distribution information between document text units; Performing cluster analysis on the correlation value distribution information to identify document text unit groups with strong correlation and document text unit groups with weak correlation; Analyze the distribution characteristics of concept co-occurrence correlation and semantic logical cohesion in the document text unit group with strong correlation, and determine the semantic correlation type between the documents, wherein the semantic correlation type includes concept extension correlation type and logical argumentation correlation type; Calculating the average value and variance value of the correlation values of each pair of document text units in the document text unit group with strong correlation, and generating correlation strength description information, wherein the correlation strength description information includes a correlation central tendency parameter and a dispersion degree parameter; The semantic association type and the association strength description information are combined to generate a semantic association analysis result in the academic document text data set.
2. The semantic association analysis method based on academic literature according to claim 1 is characterized in that: The basic semantic feature extraction process is performed on the academic document text data set to obtain the basic semantic features of each document text unit, including: Performing text cleaning on the academic document text data set to remove formatting markup information and non-content symbol information in the document text units to obtain plain text data units with a unified format; Performing word segmentation on the plain text data unit to extract core term units and key phrase units from the document text unit, wherein the core term units include professional terms in the document research field, and the key phrase units include phrases describing the logical relationship between terms; Calling a pre-trained semantic encoding model to perform contextual semantic encoding processing on the core term units and key phrase units to generate an initial semantic vector of the document text unit, wherein the initial semantic vector includes the co-occurrence relationship between the term units and the logical cohesion relationship between the phrase units; Performing feature screening on the initial semantic vector, retaining semantic dimension information that is strongly related to the document subject label, removing redundant semantic dimension information that is weakly related to the document subject label, and obtaining the core concept features of the document text unit; Analyze the position distribution relationship of the core term units and key phrase units in the document text unit, and extract the contextual semantic relationship features of the document content, wherein the contextual semantic relationship features include the proximity relationship of the term units within the paragraph and the progressive relationship of the phrase units between chapters; The core concept features and the contextual semantic relationship features are standardized to generate basic semantic features with unified dimensional representation.
3. The semantic association analysis method based on academic literature according to claim 2 is characterized in that: The text cleaning process is performed on the academic document text data set to remove format mark information and non-content symbol information in the document text unit to obtain a plain text data unit with a unified format, including: Identifying formatting markup information in a document text unit, wherein the formatting markup information includes header and footer marks, chapter title formatting marks, and formula number marks; Deleting the formatting markup information in batches by using a regular expression matching method, and retaining the text content information in the document text unit; Identifying non-content symbol information in a document text unit, wherein the non-content symbol information includes special symbols, hyperlink addresses, and footnote reference marks; Performing character-by-character detection processing on the non-content symbol information, and replacing the detected non-content symbol information with blank characters; The processed document text unit is subjected to blank line merging processing, and multiple consecutive blank lines are merged into a single blank line to obtain a plain text data unit with a unified format.
4. The semantic association analysis method based on academic literature according to claim 2 is characterized in that: The calling of the pre-trained semantic encoding model to perform contextual semantic encoding processing on the core term unit and the key phrase unit to generate an initial semantic vector of the document text unit includes: Arranging the core term units and key phrase units in the order of appearance in the document text units to generate a term phrase sequence; Inputting the term phrase sequence into the word embedding layer of the semantic encoding model, generating corresponding word vector representations for each term unit and phrase unit, wherein the word vector representations contain basic semantic information of the term phrase; Performing sequence modeling on the word vector representation through the context encoder of the semantic encoding model to capture the contextual dependencies between terms and phrases in the sequence and generate a context encoding vector containing contextual semantic information; Performing attention weighting processing on the context encoding vector, allocating attention weights according to the importance of term phrases in the document text unit, enhancing the semantic representation of key term phrases, and weakening the semantic representation of secondary term phrases; The attention-weighted context encoding vector is dimensionally spliced to generate the initial semantic vector of the document text unit.
5. The semantic association analysis method based on academic literature according to claim 4 is characterized in that: The local association modeling layer of the semantic association analysis model performs intra-document local association analysis on the standardized basic semantic features to extract the local association pattern between the core concept features and the contextual semantic relationship features in each document text unit, including: Based on the standardized basic semantic features of each document text unit, the core concept features and contextual semantic relationship features are separated; Calculate the dot product similarity between each concept dimension in the core concept feature and each relationship dimension in the context semantic relationship feature to generate a similarity matrix; Performing row normalization on the similarity matrix to obtain the importance weight of each concept dimension in different relationship dimensions; Performing column normalization on the similarity matrix to obtain coverage weights of each relationship dimension on different concept dimensions; Based on the importance weight and the coverage weight, a local association pattern of core concept features and contextual semantic relationship features in a document text unit is constructed, wherein the local association pattern includes a relationship importance distribution of a concept dimension and a concept coverage distribution of a relationship dimension.
6. The semantic association analysis method based on academic literature according to claim 1 is characterized in that: The cluster analysis process is performed on the correlation value distribution information to identify document text unit groups with strong correlation and document text unit groups with weak correlation, including: Select the density clustering algorithm as the cluster analysis method, set the cluster radius parameter and the minimum number of samples parameter; Taking the correlation value of each pair of document text units in the correlation value distribution information as a sample point, and constructing a sample point set; Calculate the Euclidean distance between each sample point and other sample points in the sample point set, and determine the neighborhood sample points of each sample point; Determine the type of sample points according to the number of the neighborhood sample points, where the types of sample points include core points, boundary points, and noise points; Merge the core point and its neighboring sample points into a cluster group, assign the boundary points to the cluster group to which the nearest core point belongs, and exclude noise points; According to the average value of the correlation values of the sample points in the cluster group, the cluster group whose average value of the correlation value is higher than the preset threshold is marked as a document text unit group with strong correlation, and the cluster group whose average value of the correlation value is lower than the preset threshold is marked as a document text unit group with weak correlation.
7. The semantic association analysis method based on academic literature according to claim 1 is characterized in that: The generating of analysis feedback information containing document association relationships based on the semantic association analysis results includes: Extracting semantic association type and association strength description information from the semantic association analysis result, and constructing structured description data of document association relationships, wherein the structured description data includes an identifier of an associated document pair and a corresponding association type label; Performing visual conversion processing on the structured description data to convert the document association relationship into a visual association graph containing nodes and edges, wherein the nodes represent document text units, the edges represent semantic association relationships between documents, and the attributes of the edges include association type labels and association strength description information; Configuring an interactive operation interface for the visual association graph, wherein the interactive operation interface supports click query operations on nodes and screening and filtering operations on edges; Performing data encapsulation processing on the structured description data and the visual association graph to generate analysis feedback information including document association relationships; The analysis feedback information is format-adapted to meet the data input format requirements of the target document management system.
8. A semantic association analysis system based on academic literature, characterized in that: It includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the semantic association analysis method based on academic documents as described in any one of claims 1 to 7 above.
Citation Information
Patent Citations
Keyword-based document research hotspot recommending method
CN106682172A
Platform content intelligent recommendation method and system based on natural language processing
CN118551031A