Cross-domain knowledge transfer method and system for academic literature

Abstract

The present invention provides a cross-domain knowledge transfer method and system for academic documents, aiming to address the difficulty in effectively implementing cross-domain knowledge transfer in existing technologies. The method first acquires a multi-domain academic document collection containing multiple disciplinary categories; then, it extracts domain features from the multi-domain academic document collection to obtain a domain feature representation for each document unit; then, based on the domain feature representation, it constructs a cross-domain knowledge transfer channel and establishes an association mapping relationship between domain feature representations of different disciplinary fields; using this cross-domain knowledge transfer channel, it transfers the knowledge of the source disciplinary document unit to the target disciplinary document unit, generating a knowledge transfer result; finally, it verifies the validity of the knowledge transfer result and generates migration validity evaluation information, thus achieving automated and accurate cross-domain knowledge transfer, which helps promote interdisciplinary integration and academic research innovation.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a cross-domain knowledge transfer method and system for academic literature. Background Art

[0002] In today's academic research, with the deepening integration of disciplines, cross-disciplinary knowledge transfer has become a key research direction. While different disciplinary fields may have specialized differences, they often also contain much shared knowledge and methods. However, traditional academic research methods are often limited to a single discipline, making it difficult to effectively explore and utilize interdisciplinary knowledge connections, resulting in a waste of knowledge resources and low research efficiency.

[0003] Although natural language processing and machine learning technologies have made certain progress in the field of academic literature analysis in recent years, most existing methods focus on surface-level analysis of literature content or knowledge mining within the same field, and fail to effectively achieve cross-domain knowledge transfer. 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 cross-domain knowledge transfer method for academic literature, the method comprising:

[0005] Acquire a multi-disciplinary academic document collection comprising multiple subject categories, wherein the multi-disciplinary academic document collection is composed of document units corresponding to different disciplines, and each document unit includes text content and field identification information of the discipline to which it belongs;

[0006] Performing domain feature extraction processing on the multi-domain academic document collection to obtain a domain feature representation of each document unit, wherein the domain feature representation includes semantic features reflecting the core content of the document unit and structural features reflecting the structural association of the document unit;

[0007] Constructing a cross-domain knowledge migration channel based on the domain feature representation, wherein the cross-domain knowledge migration channel is used to establish an association mapping relationship between feature representations of different subject domains;

[0008] Using the cross-domain knowledge transfer channel, the knowledge of the source subject literature unit is transferred to the target subject literature unit to generate a knowledge transfer result of the target subject literature unit;

[0009] The knowledge transfer result is subjected to validity verification processing to generate transfer validity evaluation information indicating the reliability of the knowledge transfer.

[0010] On the other hand, an embodiment of the present invention also provides a cross-domain knowledge migration system for 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, through the comprehensive use of technical means such as domain feature extraction, cross-domain knowledge transfer channel construction and knowledge transfer validity verification, the automated and cross-domain transfer of knowledge in multi-domain academic literature is realized. It can not only accurately extract the domain feature representation of each document unit, covering semantic features and structural features, but also construct a cross-domain knowledge transfer channel, effectively establish the association mapping relationship between the feature representations of different subject domains, and realize the cross-domain transfer of knowledge; finally, through the validity verification processing of the knowledge transfer results, it generates evaluation information indicating the reliability of knowledge transfer, ensures the accuracy of the transferred knowledge, significantly improves the automation level and accuracy of cross-domain knowledge transfer, and helps to promote interdisciplinary integration and improve the efficiency and innovation of academic research. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 It is a schematic diagram of the execution flow of the cross-domain knowledge transfer method for academic literature provided by an embodiment of the present invention.

[0013] Figure 2 Schematic diagram of exemplary hardware and software components of a cross-domain knowledge transfer system for academic literature provided by 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 cross-domain knowledge transfer method for academic documents provided by an embodiment of the present invention. The cross-domain knowledge transfer method for academic documents is introduced in detail below.

[0015] Step S110: Acquire a multi-field academic document collection containing multiple subject categories, wherein the multi-field academic document collection is composed of document units corresponding to different subjects, and each document unit contains text content and field identification information of the subject to which it belongs.

[0016] In this embodiment, in order to carry out cross-domain knowledge transfer work for academic documents, it is first necessary to construct a multi-domain academic document collection. This multi-domain academic document collection needs to cover multiple subject categories, and each subject corresponds to a number of document units. For example, in a multi-domain academic document collection involving medicine, computer science, and sociology, the document unit under the medical discipline may be an academic paper on disease treatment, drug development, etc., and its text content contains specific case analysis, experimental data, etc., and has the domain identification information of the medical discipline; the document unit under the computer science discipline may be a research report on artificial intelligence algorithms, software development, etc., and the text content contains information such as algorithm design and code implementation, and also has the domain identification information of the computer science discipline; the document unit under the sociology discipline may be an academic work on social phenomenon analysis, group behavior research, etc., and the text content contains social survey data, case analysis, etc., and also has the domain identification information of the sociology discipline.

[0017] There are many ways to obtain these literature units, such as searching academic databases. You can use professional academic databases such as HowNet, Wanfang, Web of Science, etc. to search according to keywords of different disciplines. Taking the medical discipline as an example, use keywords such as "cancer treatment" and "cardiovascular disease research" to search, and then screen the relevant literature found and include it in the multi-field academic literature collection. You can also obtain literature units through channels such as the websites of academic institutions and academic conference proceedings. For computer science disciplines, obtain cutting-edge research papers from the websites of well-known academic institutions such as the Department of Computer Science at Stanford University, or collect relevant literature from the proceedings of important academic conferences such as ACM SIGKDD.

[0018] When collecting literature units, ensure that each unit accurately identifies the discipline to which it belongs. This can be determined by using the literature's keywords, abstract, classification number, and other information. For example, if a literature unit contains keywords such as "neural network" and "deep learning," and the abstract primarily discusses algorithmic applications in computer science, it can be identified as a literature unit in the computer science discipline.

[0019] In the above process, in order to ensure the legality and compliance of the data, it is necessary to strictly follow relevant laws and regulations to obtain data authorization. When obtaining document units from public academic databases, it is necessary to carefully study the database's terms of use and authorization agreement. Generally speaking, most public academic databases will clearly stipulate the scope, purpose and restrictions of data use. Before obtaining data, it is necessary to confirm that the cross-domain knowledge transfer research being conducted meets the database's usage requirements. For example, some databases may allow the use of data for non-commercial research, but prohibit the use of data for commercial profit or other unauthorized purposes. When obtaining data, it is necessary to follow the procedures specified by the database, such as registering an account, filling in a description of the purpose of use, etc., to obtain legal data access rights.

[0020] In the case of obtaining literature units from channels such as academic institution websites and academic conference proceedings, it is necessary to communicate directly with the relevant institutions or copyright owners to obtain authorization for data collection. You can send a formal request email, detailing the purpose, use, and research methods of data collection. The email must clearly state that you will comply with relevant laws, regulations, and ethical standards to protect data privacy and intellectual property rights. For example, when requesting the computer science department of a university to obtain research papers on its website, you should state in the email that these papers will be used for cross-disciplinary knowledge transfer research, aiming to promote knowledge exchange and sharing between different disciplines, and the content of the papers will not be used for commercial purposes or disclosed to third parties. Wait for a response from the relevant institution or copyright owner, and only proceed with data collection after obtaining clear authorization.

[0021] During the data collection process, strict privacy protection and anti-leakage technical measures must be taken for document units that may involve privacy-sensitive data. For medical document units containing patient personal information, these privacy-sensitive information must be specially handled during text preprocessing operations. Data desensitization technology can be used to replace sensitive information such as the patient's name, ID number, address, etc. with anonymous identifiers. For example, the patient's name is replaced with "Patient A" or "Patient B", and the ID number is replaced with a string of randomly generated characters. At the same time, the processed data is encrypted and stored, and advanced encryption algorithms, such as the AES encryption algorithm, are used to ensure the security of the data during storage and transmission. When accessing and using these processed data, a strict permission management mechanism is set up, and only authorized personnel can access the relevant data.

[0022] Step S120: performing domain feature extraction processing on the multi-domain academic document collection to obtain a domain feature representation of each document unit, wherein the domain feature representation includes semantic features reflecting the core content of the document unit and structural features reflecting the structural association of the document unit.

[0023] After acquiring a multi-disciplinary academic document collection, the next step is to extract domain features from each document unit in the collection to obtain a domain feature representation that represents the document unit. The domain feature representation consists of two parts: semantic features that reflect the core content of the document unit, and structural features that reflect the structural associations of the document unit.

[0024] Semantic features can reflect the meaning of key concepts and the expression of core arguments within a document unit. For example, in a medical article, semantic features can reflect core content such as the pathogenesis and treatment of a disease; in computer science literature, semantic features can reflect the principles and application scenarios of an algorithm. Structural features focus on the organizational structure of a document unit and the connections between its parts. For example, in medical literature, the hierarchical relationship between different chapters, the citation relationship between figures and the main text, etc.; in computer science literature, the code structure of different modules, and the connection between references and the main text, etc.

[0025] Step S121: performing a text preprocessing operation on each document unit in the multi-field academic document collection, wherein the text preprocessing operation includes removing non-content symbols, filtering common stop words, and splitting continuous text paragraphs into processable text segment units.

[0026] In order to more accurately extract the domain features of document units, we must first perform text preprocessing on each document unit in the multi-domain academic document collection.

[0027] Removing non-content symbols is the first step in preprocessing. Within the text of a document, there are many non-content symbols, such as punctuation marks (commas, periods, exclamation marks, etc.) and special characters (@, #, $, etc.). While these symbols play a role in text expression, they do not provide key information for semantic analysis and feature extraction, and may even interfere with subsequent processing. For example, in a medical document, a large number of punctuation marks can complicate the semantic structure of the text and hinder the identification of key concepts. Therefore, these non-content symbols need to be removed to make the text more concise. Regular expressions can be used to remove non-content symbols, for example, by defining a regular expression pattern that matches all punctuation marks and special characters and then replacing them with an empty string.

[0028] Filtering common stop words is the second step of preprocessing. Common stop words refer to those words that frequently appear in natural language but contribute little to the core semantic expression of the text, such as "de", "shi", "zai", "he", etc. In literature, the large presence of these words will increase the complexity of text processing and may obscure key semantic information. For example, in computer science literature, the word "de" may appear in many places, but it does not help to understand the core principles of algorithms. Therefore, these common stop words need to be filtered out. A stop word list can be established to include common stop words, and then during the text processing, the words in the text are compared with the stop word list, and the matching stop words are removed from the text.

[0029] Splitting continuous text paragraphs into processable text fragment units is the third step of preprocessing. A literature usually consists of multiple continuous text paragraphs, and these paragraphs may contain different topics and information. To facilitate subsequent semantic analysis and feature extraction, it is necessary to split the continuous text paragraphs into smaller, processable text fragment units. The splitting can be done according to the topic of the paragraph, the logical relationship of the sentences, etc. For example, in a medical literature, the paragraphs describing disease symptoms and the paragraphs introducing treatment methods can be split into independent text fragment units respectively. In this way, each text fragment unit can be used as a relatively independent processing object, which is more convenient for subsequent semantic analysis.

[0030] Step S122: Invoke the pre-trained semantic encoding model to perform context semantic analysis on the text fragment units, and generate semantic features reflecting the core content of the literature unit. The semantic features include the context association information of key concepts in the text fragment units and the semantic expression vectors of the core arguments.

[0031] After text preprocessing, processable text fragment units are obtained. Next, invoke the pre-trained semantic encoding model to perform context semantic analysis on these text fragment units to generate semantic features reflecting the core content of the literature unit.

[0032] The pre-trained semantic encoding model is trained on a large amount of text data and can learn the semantic information and context relationships of natural language. For example, in the medical field, the pre-trained semantic encoding model can learn the semantic associations between disease names, symptoms, treatment methods, etc.; in the field of computer science, it can learn the semantic relationships between algorithm names, data structures, programming languages, etc.

[0033] Step S1221: Input the text fragment units into the word embedding layer of the semantic encoding model to generate the initial word vector representations of the words in each text fragment unit.

[0034] First, the text snippet unit is input into the word embedding layer of the semantic encoding model. The role of the word embedding layer is to convert each word in the text into a corresponding vector representation, namely the initial word vector representation. During this process, the word embedding layer maps the words into a high-dimensional vector space based on the pre-trained word vector model. For example, in the text snippet unit of medical literature, the word "cancer" will be mapped to a specific vector, which contains the information of the word "cancer" in the semantic space. The position and direction of different words in the vector space reflect the semantic relationship between them. For example, the positions of the two words "cancer" and "tumor" in the vector space may be relatively close because they have a certain semantic connection.

[0035] Step S1222: Perform temporal context modeling on the initial word vector representation through the bidirectional long short-term memory network layer of the semantic encoding model to capture the semantic dependency relationship between words in the text fragment unit in the context of the previous and next words, and generate an intermediate semantic vector containing context information.

[0036] Next, the generated initial word vector representation is input into the bidirectional long short-term memory (Bi-LSTM) layer of the semantic encoding model. The Bi-LSTM layer is a recurrent neural network that processes sequential data and captures temporal dependencies between elements in a sequence. In text processing, the Bi-LSTM layer can capture the semantic dependencies between words in a text segment unit within their surrounding context. For example, in the sentence "This drug has a significant effect on cancer treatment," there is a semantic dependency between "drug" and "cancer treatment," and this relationship is learned by the Bi-LSTM layer. The Bi-LSTM layer processes the text sequence in two directions (forward and backward), capturing the forward and backward context of the word, respectively. This information is then combined to generate an intermediate semantic vector that incorporates contextual information. This intermediate semantic vector contains not only the semantic information of the word itself, but also the semantic information of the word within its surrounding context.

[0037] Step S1223: Utilize the attention mechanism layer of the semantic encoding model to perform key information focusing processing on the intermediate semantic vector, identify key concept words in the text fragment unit that are strongly related to the core content of the document unit, and generate an attention weight distribution of the key concept words.

[0038] After obtaining the intermediate semantic vector, the attention mechanism layer of the semantic encoding model is used to focus on key information. The role of the attention mechanism layer is to identify key concept words in the text fragment unit that are strongly related to the core content of the document unit and assign attention weights to these key concept words. For example, in the text fragment unit of a medical document, "cancer treatment" and "drug efficacy" may be key concept words. The attention mechanism layer calculates the relevance of each word to the core content based on the information in the intermediate semantic vector, and then assigns an attention weight to each word. The higher the relevance, the greater the attention weight. In this way, the importance of key concept words can be highlighted and the interference of irrelevant information can be reduced. Finally, the attention weight distribution of key concept words is generated, which reflects the importance of each key concept word in the text fragment unit.

[0039] Step S1224: performing weighted aggregation processing on the intermediate semantic vectors according to the attention weight distribution to generate local semantic features reflecting the contextual association information of key concept vocabulary.

[0040] According to the generated attention weight distribution, the intermediate semantic vectors are weighted aggregated. Specifically, each intermediate semantic vector is multiplied by the corresponding attention weight, and then these weighted vectors are aggregated. For example, for a text fragment unit containing multiple words, each word has a corresponding intermediate semantic vector and attention weight. Each intermediate semantic vector is multiplied by its attention weight and then added together to obtain a local semantic feature that reflects the contextual association information of the key concept vocabulary. The local semantic feature not only contains the semantic information of the key concept vocabulary, but also contains their association information in the context. Through the above-mentioned weighted aggregation processing, the key semantic information in the text fragment unit can be extracted more accurately.

[0041] Step S1225: Input the local semantic features into the fully connected layer of the semantic encoding model for dimensionality compression processing to generate a semantic expression vector with a fixed dimensional representation, and the semantic expression vector maintains semantic information consistency with the local semantic features.

[0042] Finally, the local semantic features are input into the fully connected layer of the semantic encoding model for dimensionality compression. The fully connected layer is a neural network layer that can map the input vector into a vector space of fixed dimensionality. During this process, the fully connected layer performs linear transformation and nonlinear activation on the local semantic features, converting them into semantic expression vectors with fixed dimensional representation. For example, the dimension of the local semantic features may be relatively high, which is not conducive to subsequent processing and analysis. The fully connected layer compresses them to a lower fixed dimension. At the same time, the fully connected layer ensures that the semantic expression vector maintains semantic information consistency with the local semantic features, that is, the semantic expression vector can still accurately reflect the core semantic information of the text segment unit.

[0043] Step S123: parsing the text structure information of the document unit, wherein the text structure information includes the hierarchical relationship of chapter titles, the citation association relationship between figures and text, and the subject distribution relationship of references.

[0044] While extracting semantic features, it is also necessary to parse and process the text structure information of the document unit. Text structure information includes multiple aspects. For cross-domain knowledge transfer, the hierarchical relationship between chapter titles, the citation relationship between figures and text, and the disciplinary distribution of references are relatively important information.

[0045] The hierarchical relationship between chapter titles reflects the organizational structure of a document and the logical hierarchy of its content. For example, a medical article may have chapter titles such as "Introduction," "Disease Overview," "Treatment Methods," "Experimental Results," and "Conclusion." These chapter titles have a certain hierarchical relationship. For example, "Treatment Methods" may be a subsection under "Disease Overview." By analyzing the hierarchical relationship between chapter titles, we can understand the overall framework and logical order of the document's content. Text parsing technology can be used to identify the formatting of chapter titles (such as font size and indentation) and determine the hierarchical relationship between chapter titles based on this formatting information.

[0046] The citation relationship between a figure and the main text reflects the figure's role in the document and its connection to the main text. Medical literature may contain figures and tables showing data such as disease incidence and treatment efficacy. These figures and tables are often cited in the main text to support the main text's arguments. By analyzing the citation relationship between figures and the main text, we can understand the information flow and mutual support between the figures and the text. Text matching technology can be used to find sentences in the main text that cite figures and tables, and then associate the figures with these citations.

[0047] The disciplinary distribution of references reflects the research context and knowledge sources of a document. In a computer science paper, references may span different subfields of computer science, as well as other related disciplines (such as mathematics and physics). By analyzing the disciplinary distribution of references, we can understand the document's citations across interdisciplinary fields. References can be categorized and annotated to determine the discipline to which each reference belongs. References from different disciplines can then be counted to analyze their disciplinary distribution.

[0048] Step S124: extracting structural features reflecting the structural association of document units based on the text structure information, wherein the structural features include a hierarchical depth parameter of chapter titles, a citation frequency parameter of tables and figures and text, and a subject concentration parameter of references.

[0049] Based on the text structure information obtained from the analysis, structural features that reflect the structural association of the document units are extracted. These structural features include the hierarchical depth parameters of chapter titles, the citation frequency parameters of tables and figures and the disciplinary concentration parameters of references.

[0050] Step S1241: perform hierarchical depth calculation on the hierarchical relationship of chapter titles in the text structure information, count the hierarchical distance of each chapter title relative to the root directory, and generate a hierarchical depth parameter for the chapter title. The hierarchical depth parameter is used to indicate the core degree of the chapter content in the document unit.

[0051] Perform hierarchical depth calculation on the hierarchical relationship of chapter titles. First, determine the root directory of the document, which can usually be the main title of the document. Then, count the hierarchical distance of each chapter title relative to the root directory. For example, in a medical document, the main title is the root directory, "Disease Overview" is the first-level chapter title, and its hierarchical depth is 1; "Disease Symptoms" is the second-level chapter title under "Disease Overview", and its hierarchical depth is 2. The hierarchical depth parameter can reflect the core degree of the chapter content in the document unit. Generally speaking, the content corresponding to the chapter title with a shallower hierarchical depth may be more core. Through hierarchical depth calculation processing, a hierarchical depth parameter can be generated for each chapter title. These parameters constitute the characteristics that reflect the hierarchical structure of the chapter title.

[0052] Step S1242: Perform citation frequency statistics on the citation association relationship between the charts and the text in the text structure information, count the number of times each chart is cited by the text paragraphs, and generate citation frequency parameters for the charts and the text. The citation frequency parameters are used to indicate the degree of association between the chart content and the text content.

[0053] Citation frequency statistics are performed on the citation relationship between charts and the main text. By searching for statements in the main text that cite charts, the number of times each chart is cited is counted. For example, in a computer science paper, there is a chart showing algorithm performance that is cited five times in the main text, so the citation frequency parameter of this chart is 5. The citation frequency parameter can indicate the close connection between the content of the chart and the main text. The higher the citation frequency, the closer the connection between the chart and the main text. By counting the citation frequency of each chart, the citation frequency parameters of the chart and the main text are generated. These parameters constitute the characteristics that reflect the citation structure of the chart and the main text.

[0054] Step S1243: Calculate the subject concentration of the subject distribution relationship of the references in the text structure information, count the proportion of the number of documents belonging to the same subject in the references, and generate a subject concentration parameter for the references. The subject concentration parameter is used to indicate the degree of subject focus of the research content of the document unit.

[0055] The subject concentration of the references is calculated. First, the references are classified to determine the subject to which each reference belongs. Then, the proportion of references belonging to the same subject to the total number of references is calculated. For example, in a medical document, there are a total of 20 references, 15 of which belong to the medical discipline. Then, the reference subject concentration parameter of the document is 15 / 20. The subject concentration parameter can indicate the degree of subject focus of the research content of the document unit. The higher the subject concentration, the more focused the research content of the document is on a certain discipline. By calculating the subject concentration parameter, a feature reflecting the subject distribution structure of the reference is generated.

[0056] Step S1244: normalize the hierarchical depth parameter, the citation frequency parameter, and the discipline concentration parameter to eliminate the dimensional differences between different parameters and generate a set of structural features with comparable metrics.

[0057] Since the hierarchical depth parameters, citation frequency parameters, and discipline concentration parameters have different dimensions, they need to be standardized in order to be able to effectively compare and analyze these parameters. The purpose of standardization is to eliminate the dimensional differences between different parameters so that they have comparable measurements. Common standardization methods can be used, such as z-score standardization. For the hierarchical depth parameters, citation frequency parameters, and discipline concentration parameters, their means and standard deviations are calculated respectively, and then the mean is subtracted from each parameter value and then divided by the standard deviation to obtain the standardized parameter value. In the above manner, the hierarchical depth parameters, citation frequency parameters, and discipline concentration parameters are converted into numerical values with the same dimensions, generating a set of structural features with comparable measurements.

[0058] Step S1245: performing feature fusion processing on the structural feature set, integrating the hierarchical depth parameter, the citation frequency parameter and the subject concentration parameter into a single-dimensional structural feature by weighted summation, and the structural feature retains the original semantic information of each parameter.

[0059] Perform feature fusion processing on the standardized set of structural features. In order to obtain a feature that can comprehensively reflect the structural association of document units, it is necessary to integrate the hierarchical depth parameter, citation frequency parameter, and discipline concentration parameter. Feature fusion can be performed by weighted summation, assigning a weight to each parameter, and then multiplying each parameter by its weight and adding them together to obtain a single-dimensional structural feature. For example, assign a weight w1 to the hierarchical depth parameter, a weight w2 to the citation frequency parameter, and a weight w3 to the discipline concentration parameter. Multiply the hierarchical depth parameter by w1, the citation frequency parameter by w2, and the discipline concentration parameter by w3, and then add them together to obtain the structural feature. When assigning weights, the importance of each parameter needs to be considered to ensure that the structural feature can retain the original semantic information of each parameter.

[0060] Step S125: normalizing the semantic features and the structural features to generate a domain feature representation with a unified representation form, wherein the domain feature representation includes both semantic association information and structural association information.

[0061] After obtaining the semantic features and structural features, they need to be standardized so that they can be processed and compared under the same framework. The dimensions and value ranges of semantic features and structural features may be different. These differences can be eliminated through standardization to generate domain feature representations with a unified representation form. A standardization method similar to step S1244 can be used, such as z-score standardization. The mean and standard deviation of the semantic features and structural features are calculated respectively, and then each element of the semantic features and structural features is subtracted from its mean and then divided by its standard deviation to obtain the standardized semantic features and structural features.

[0062] Next, the standardized semantic features and structural features are concatenated. Since semantic features and structural features represent different aspects of a document unit, concatenation can integrate these two types of information into a unified domain feature representation. For example, assuming that the semantic feature is an m-dimensional vector and the structural feature is an n-dimensional vector, the concatenated domain feature representation is an m+n-dimensional vector. This domain feature representation contains both semantic and structural association information, and can more comprehensively represent the characteristics of a document unit.

[0063] Step S130: constructing a cross-domain knowledge transfer channel based on the domain feature representation, wherein the cross-domain knowledge transfer channel is used to establish an association mapping relationship between feature representations of different subject domains.

[0064] After obtaining the domain feature representation of each document unit, the next step is to build a cross-domain knowledge transfer channel based on these domain feature representations. The role of this cross-domain knowledge transfer channel is to establish an association mapping relationship between the domain feature representations of different disciplines, enabling knowledge transfer between different disciplines.

[0065] Step S131: selecting interdisciplinary document pairs containing a common research topic from the multi-field academic document collection, wherein the interdisciplinary document pairs are composed of document units belonging to different disciplines but with overlapping research topics.

[0066] First, interdisciplinary document pairs are selected from a multi-disciplinary academic literature collection. These interdisciplinary document pairs need to contain common research themes, that is, they are composed of document units belonging to different disciplines but with overlapping research themes. For example, in the fields of medicine and biology, there may be research on gene therapy for diseases. In this case, document pairs involving gene therapy can be selected from the medical document unit and the biological document unit. By analyzing the title, abstract, keywords and other information of the document units, document unit pairs with overlapping research themes can be found. For example, using a text matching algorithm, the keywords of different document units are compared. If a certain number of keywords are the same or similar, it is considered that the two document units may have a common research theme.

[0067] Step S132: extracting the domain feature representations of the source subject document unit and the target subject document unit in the interdisciplinary document pair.

[0068] After selecting an interdisciplinary document pair, extract the domain feature representations of the source and target subject document units. The source subject document unit is the source of knowledge, and the target subject document unit is the target of knowledge transfer. For example, in an interdisciplinary document pair between medicine and computer science, if disease diagnosis knowledge from the medical field is to be transferred to an intelligent diagnosis system in the computer science field, the medical document unit is the source subject document unit, and the computer science document unit is the target subject document unit. Based on the domain feature representation of each document unit obtained in the previous step, the domain feature representations of the source and target subjects are directly extracted from the interdisciplinary document pair.

[0069] Step S133: Calculate the feature similarity value between the domain feature representation of the source subject document unit and the domain feature representation of the target subject document unit, wherein the feature similarity value is calculated using a cosine similarity algorithm.

[0070] After obtaining the domain feature representations of the source and target disciplines, calculate the feature similarity values between them. The cosine similarity algorithm can be used to calculate the feature similarity value. Cosine similarity is a commonly used method for calculating vector similarity, which measures the degree of similarity between two vectors by calculating the cosine value of the angle between them. For the domain feature representation vector A of the source discipline document unit and the domain feature representation vector B of the target discipline document unit, first calculate the dot product of vector A and vector B, then calculate the modulus of vector A and vector B respectively, and finally divide the dot product by the product of the modulus of the two vectors to obtain the cosine similarity value. The closer the cosine similarity value is to 1, the more similar the two vectors are, that is, the more similar the domain feature representations of the source and target disciplines are; the closer the value is to 0, the less similar the two vectors are.

[0071] Step S134: screening interdisciplinary document pairs with similarity values higher than a preset threshold as key mapping samples based on the feature similarity values.

[0072] Based on the calculated feature similarity values, interdisciplinary document pairs with similarity values above a preset threshold are selected as key mapping samples. The preset threshold is a pre-defined criterion used to determine whether the similarity between interdisciplinary document pairs is high enough for subsequent association rule mining. For example, if the preset threshold is a set similarity value, when the feature similarity value of an interdisciplinary document pair exceeds the threshold, it is selected as a key mapping sample. By selecting key mapping samples, unnecessary calculations and analysis can be reduced, allowing focus to be placed on interdisciplinary document pairs with high similarity.

[0073] Step S135: performing association rule mining on the source subject domain feature representation and the target subject domain feature representation of the key mapping sample to generate a migration rule set reflecting the cross-disciplinary feature correspondence.

[0074] The association rule mining process is performed on the source subject domain feature representation and the target subject domain feature representation of the selected key mapping samples. The purpose of association rule mining is to find the corresponding relationship between the source subject domain feature representation and the target subject domain feature representation and generate a set of migration rules.

[0075] Step S1351: performing feature decomposition processing on the source subject domain feature representation of the key mapping sample to separate the semantic feature part and the structural feature part.

[0076] First, the source subject domain feature representation of the key mapping sample is subjected to feature decomposition. Since the domain feature representation is composed of semantic features and structural features, it can be separated into semantic feature parts and structural feature parts. Based on the dimensional information during the previous splicing, the domain feature representation vector can be divided according to the dimension to obtain a semantic feature vector and a structural feature vector. For example, if the domain feature representation is an m+n dimensional vector, where the first m dimensions are semantic features and the last n dimensions are structural features, then the first m dimensions can be extracted as the semantic feature part, and the last n dimensions can be extracted as the structural feature part.

[0077] Step S1352: performing synchronous feature decomposition processing on the target subject domain feature representation of the key mapping sample to separate the corresponding semantic feature part and structural feature part.

[0078] Similarly, synchronized feature decomposition is performed on the target subject domain feature representation of the key mapping sample. Following the same method as step S1351, the target subject domain feature representation vector is separated into semantic and structural features. This decomposition of the domain feature representations of both the source and target subjects into semantic and structural features facilitates subsequent correspondence analysis.

[0079] Step S1353: Establishing a semantic correspondence between the semantic feature part of the source discipline and the semantic feature part of the target discipline, wherein the semantic correspondence is generated by matching feature sub-vectors representing the same research concept in the two parts.

[0080] Establish a semantic correspondence between the semantic feature components of the source discipline and the semantic feature components of the target discipline. This correspondence can be generated by matching feature subvectors representing the same research concepts in the two components. For example, in an interdisciplinary literature pair between medicine and computer science, the semantic feature component of the source discipline (medicine) may have a feature subvector representing "disease diagnosis," while the semantic feature component of the target discipline (computer science) may have a feature subvector representing "intelligent diagnostic algorithm." Although these two subvectors come from different disciplines, they are both related to diagnosis, so a semantic correspondence can be established between them. Vector matching algorithms, such as the nearest neighbor algorithm, can be used to find similar feature subvectors in the semantic feature components of the source and target disciplines and match them.

[0081] Step S1354: establishing a structural correspondence between the source discipline structural feature portion and the target discipline structural feature portion, wherein the structural correspondence is generated by matching feature sub-vectors representing the same logical organization mode in the two portions.

[0082] Establish a structural correspondence between the source discipline's structural features and the target discipline's structural features. This correspondence is generated by matching the characteristic subvectors representing the same logical organization in the two sections. For example, the structural features of the source discipline (medicine) may have a characteristic subvector representing the chapter hierarchy, while the structural features of the target discipline (computer science) may have a characteristic subvector representing the code module hierarchy. If the logical organization represented by these two subvectors is similar (e.g., both are hierarchical structures), then a structural correspondence can be established between them. Vector matching algorithms can also be used to find similar characteristic subvectors and establish a correspondence.

[0083] Step S1355: Perform rule induction processing on the semantic correspondence and the structural correspondence to generate a migration rule set including semantic migration rules and structural migration rules. The semantic migration rules are used to guide the conversion of the semantic features of the source discipline to the semantic features of the target discipline, and the structural migration rules are used to guide the conversion of the structural features of the source discipline to the structural features of the target discipline.

[0084] Rule induction is performed on the established semantic and structural correspondences. Rule induction involves summarizing these correspondences into universal rules. For semantic correspondences, semantic transfer rules are summarized. These rules guide the transformation of the semantic features of the source discipline into those of the target discipline. For example, if a concept in the semantic features of the source discipline corresponds to a concept in the semantic features of the target discipline, a rule can be summarized to describe how to transform this concept in the source discipline into the corresponding concept in the target discipline during knowledge transfer. For structural correspondences, structural transfer rules are summarized to guide the transformation of the structural features of the source discipline into those of the target discipline. For example, if the chapter hierarchy of the source discipline corresponds to the code module hierarchy of the target discipline, a rule can be summarized to describe how to transform the chapter hierarchy of the source discipline into the code module hierarchy of the target discipline. The semantic transfer rules and structural transfer rules are integrated to form a transfer rule set.

[0085] Step S1356: performing conflict detection processing on the migration rule set, eliminating contradictory descriptions between different rules, and generating a final migration rule set.

[0086] Perform conflict detection on the generated migration rule set. During the rule induction process, contradictory descriptions may appear between different rules. For example, one semantic migration rule may stipulate that a concept in the source discipline be converted to concept A in the target discipline, while another rule may stipulate that it be converted to concept B in the target discipline. Through conflict detection, these contradictory rules are identified and adjusted or eliminated. A rule comparison algorithm can be used to compare the rules in the migration rule set pairwise to identify the contradictions between the rules. These rules are then processed according to the set strategy (such as prioritizing more specific rules), ultimately eliminating the contradictory descriptions and generating the final migration rule set.

[0087] Step S136: performing structured storage processing on the migration rule set to generate a cross-domain knowledge migration channel for establishing an association mapping relationship between feature representations of different subject areas.

[0088] The final set of migration rules is stored in a structured manner. Structured storage allows the migration rule set to be stored in an ordered, queryable manner, facilitating subsequent use and management. A database can be used to store the migration rule set, with each rule as a record in the database, documenting the rule type (semantic migration rule or structural migration rule), source discipline feature information, target discipline feature information, and so on. By storing the migration rule set in a structured manner, a cross-domain knowledge migration channel is generated to establish associative mapping relationships between feature representations of different disciplinary domains. This cross-domain knowledge migration channel serves as a bridge for knowledge migration, enabling knowledge from the source discipline to be mapped to the target discipline through these rules.

[0089] Step S140: Utilize the cross-domain knowledge transfer channel to transfer the knowledge of the source subject literature unit to the target subject literature unit, and generate a knowledge transfer result of the target subject literature unit.

[0090] After constructing the cross-domain knowledge transfer channel, the cross-domain knowledge transfer channel can be used to transfer the knowledge of the source subject literature unit to the target subject literature unit to generate the knowledge transfer results of the target subject literature unit.

[0091] Step S141: selecting source subject document units and target subject document units that require knowledge transfer from the multi-field academic document collection, wherein the source subject document units contain target knowledge content to be transferred.

[0092] First, select source subject literature units and target subject literature units that require knowledge transfer from a collection of multi-disciplinary academic literature. The source subject literature units contain the target knowledge content to be transferred. For example, in an interdisciplinary knowledge transfer scenario between medicine and computer science, if you want to transfer knowledge about disease diagnosis in the medical field to an intelligent diagnosis system in the computer science field, then select literature units containing disease diagnosis-related knowledge from the medical literature units as source subject literature units, and select literature units that need to supplement knowledge from the computer science literature units as target subject literature units. Based on specific knowledge transfer needs, you can screen appropriate source subject literature units and target subject literature units using information such as the document title, abstract, and keywords.

[0093] Step S142: extracting the domain feature representation of the source subject document unit, wherein the domain feature representation includes source subject semantic features and source subject structural features.

[0094] After selecting the source disciplinary document unit, its domain feature representation is extracted. This domain feature representation includes the source disciplinary semantic features and source disciplinary structural features. In the previous step, domain feature extraction processing has been performed on each document unit in the multi-disciplinary academic document collection. Therefore, the domain feature representation of the source disciplinary document unit can be directly extracted from the stored domain feature representation. The source disciplinary semantic features reflect the core content semantic information of the source disciplinary document unit, and the source disciplinary structural features reflect the structural association information of the source disciplinary document unit.

[0095] Step S143: The source subject semantic features are converted using the semantic transfer rules in the cross-domain knowledge transfer channel to generate target semantic features that are dimensionally aligned with the target subject semantic features.

[0096] The semantic features of the source discipline are converted using the semantic transfer rules in the cross-domain knowledge transfer channel. The semantic transfer rules are generated in step S135, which stipulate how to convert the semantic features of the source discipline into the semantic features of the target discipline. According to the semantic transfer rules, each feature element in the semantic features of the source discipline is converted. For example, if the semantic transfer rules stipulate that a concept in the source discipline is converted into a corresponding concept in the target discipline, then the feature elements representing the concept in the semantic features of the source discipline are replaced or adjusted accordingly. During the conversion process, it is necessary to ensure that the generated target semantic features are aligned with the dimensions of the semantic features of the target discipline. Dimension alignment can ensure that the target semantic features can be integrated with the knowledge system of the target discipline. Through the above conversion process, the semantic features of the source discipline are converted into target semantic features that match the semantic features of the target discipline.

[0097] Step S144: converting the source discipline structural features through the structural migration rules in the cross-domain knowledge migration channel to generate target structural features that are dimensionally aligned with the target discipline structural features.

[0098] Similarly, the structural migration rules in the cross-domain knowledge migration channel are used to convert the source discipline structural features. The structural migration rules specify how to convert the source discipline structural features into the target discipline structural features. According to the structural migration rules, each characteristic element in the source discipline structural features is converted. For example, if the structural migration rules stipulate that the chapter hierarchy structure of the source discipline is converted into the code module hierarchy structure of the target discipline, then the characteristic elements representing the chapter hierarchy structure in the source discipline structural features are adjusted according to the rules. During the conversion process, it is necessary to ensure that the generated target structural features are aligned with the dimensions of the target discipline structural features, so that the target structural features can adapt to the knowledge structure of the target discipline. Through the above conversion process, the source discipline structural features are converted into target structural features that match the target discipline structural features.

[0099] Step S145: performing feature fusion processing on the target semantic features and the target structural features to generate a target subject domain feature representation.

[0100] After obtaining the target semantic features and target structural features, they are subjected to feature fusion processing. Feature fusion can integrate the target semantic features and target structural features to form a unified target subject domain feature representation. A similar splicing method as step S125 can be used to splice the target semantic feature vector and the target structural feature vector. Since the target semantic features and target structural features have been aligned with the semantic feature and structural feature dimensions of the target subject respectively, the spliced target subject domain feature representation can accurately reflect the characteristic information of the target subject.

[0101] Step S146: Adaptively adjust the target knowledge content of the source subject document unit based on the target subject domain feature representation to generate a knowledge transfer result that conforms to the target subject knowledge expression specification, and the knowledge transfer result includes adjusted semantic content and adjusted structural organization.

[0102] Based on the generated target subject domain feature representation, the target knowledge content of the source subject literature unit is adaptively adjusted.

[0103] Step S1461: Decompose the target knowledge content of the source subject literature unit to separate the core argument part and the auxiliary argument part.

[0104] First, the target knowledge content of the source subject literature unit is decomposed. The target knowledge content is separated into a core argument and supporting arguments. The core argument is the core of the knowledge content, expressing the main viewpoint and conclusion of the knowledge; the supporting arguments include evidence, examples, and other content used to support the core argument. For example, in the target knowledge content of a medical article, the core argument may be the effectiveness of a treatment for a certain disease, while the supporting arguments may include relevant case analysis, experimental data, etc. Text analysis techniques can be used to decompose the target knowledge content based on sentence importance, logical relationships, and other factors.

[0105] Step S1462: Based on the target semantic features in the target subject domain feature representation, term replacement processing is performed on the core argument part, replacing the source subject-specific terms with equivalent terms in the target subject, and generating semantically adapted core argument content.

[0106] Based on the target semantic features in the target discipline's domain feature representation, terminology is replaced in the core argument. The source and target disciplines may use different terms to express the same or similar concepts. For example, in interdisciplinary knowledge transfer between medicine and computer science, the term "disease diagnosis" in medicine might be expressed as "intelligent diagnostic algorithm" in computer science. Using the information in the target semantic features, equivalent terms for source discipline-specific terms in the target discipline are identified. The source discipline-specific terms in the core argument are then replaced with equivalent target discipline terms. This ensures that the core argument content is semantically compatible with the target discipline, making it easier for readers in the target discipline to understand.

[0107] Step S1463: Based on the target structural features in the target subject field feature representation, the auxiliary argument part is restructured, and the presentation order of the auxiliary argument content is reorganized according to the chapter hierarchy relationship and citation association relationship of the target subject literature unit to generate structurally adapted auxiliary argument content.

[0108] The supporting argument section is restructured based on the target structural features in the target discipline domain feature representation. The target discipline literature unit may have its own specific chapter hierarchy and citation association relationships. In order to make the supporting argument content structurally compatible with the target discipline, the presentation order of the supporting argument content needs to be reorganized according to the structural requirements of the target discipline. For example, the target discipline literature unit may require the experimental results to be presented first, followed by the results analysis, while the supporting argument content of the source discipline literature unit may require the analysis to be performed first and then the results to be presented. Based on the target structural features, the order of the supporting argument content is adjusted to generate structurally compatible supporting argument content.

[0109] Step S1464: Perform content consistency check on the semantically adapted core argument content and the structurally adapted auxiliary argument content to ensure the continuity of the two parts in logical expression.

[0110] A content coherence check is performed on the semantically aligned core argument and structurally aligned supporting argument content. Although the core argument and supporting argument content have been semantically and structurally aligned with the target discipline, respectively, after terminology replacement and structural reorganization, the logical coherence between the two components may be affected. For example, after terminology replacement, the logical relationship between the core argument and supporting argument may no longer be clear. A content coherence check examines the logical relationship between the two components to ensure their logical continuity. Logical reasoning algorithms can be used to analyze the rationality of the logical connection between the two components and to adjust any inconsistencies.

[0111] Step S1465: Perform format standardization on the content that has passed the coherence check, unify the punctuation usage specifications and paragraph separation rules, and generate knowledge transfer results that conform to the target subject knowledge expression specifications.

[0112] After the coherence check is passed, the content is formatted and standardized. Different disciplines may have different punctuation and paragraph separation rules. For example, in medicine, punctuation may focus more on expressing strict logical relationships, while in literature, punctuation may be more used to create emotional atmosphere. Paragraph separation rules may also vary by discipline; some disciplines may prefer long paragraphs, while others are more accustomed to short paragraphs.

[0113] Punctuation usage needs to be standardized according to the target discipline's standards. A target discipline punctuation rule library can be established, storing common punctuation usage patterns in the target discipline. Then, the semantically aligned core argument content and structurally aligned auxiliary argument content are examined sentence by sentence, and non-standard punctuation is replaced according to the rules in the rule library. For example, if the target discipline typically uses a semicolon to separate parallel sentence components, while the source discipline uses a comma, the comma should be replaced with a semicolon.

[0114] It's also crucial to unify paragraph separation rules. You can determine the standard for paragraph separation based on the paragraph characteristics of the target disciplinary literature unit. For example, the target disciplinary literature unit may use a paragraph break at the beginning of each new point or argument. Then, adjust the content according to this standard. You can determine whether paragraph breaks are necessary by analyzing the subject and logical relationships of the sentences, and re-segment any content that doesn't conform to the target disciplinary paragraph separation rules.

[0115] After uniformly processing punctuation and paragraph separation rules, the generated content conforms to the target discipline's knowledge expression standards. This knowledge transfer result includes adjusted semantic content and adjusted structural organization, which not only adapts to the target discipline in terms of semantics but also follows the target discipline's expression conventions in terms of structure and format.

[0116] Step S150: performing validity verification processing on the knowledge transfer result to generate transfer validity evaluation information indicating the reliability of the knowledge transfer.

[0117] In order to ensure the quality and reliability of knowledge transfer, it is necessary to verify the validity of the knowledge transfer results and generate corresponding transfer validity evaluation information.

[0118] Step S151: selecting a target discipline benchmark document unit related to the research topic of the knowledge transfer result from the multi-field academic document collection, wherein the target discipline benchmark document unit contains authoritative knowledge content verified by domain experts.

[0119] Select target discipline benchmark literature units related to the research topic of knowledge transfer results from a collection of multi-disciplinary academic literature. These benchmark literature units contain authoritative knowledge content verified by domain experts, which can serve as reference standards for verifying the effectiveness of knowledge transfer results. For example, in an interdisciplinary knowledge transfer scenario between medicine and computer science, if the knowledge transfer result is about intelligent diagnostic algorithm knowledge in the field of computer science, then authoritative literature units related to the research topic of intelligent diagnostic algorithms, reviewed and approved by domain experts, are selected from the computer science literature units as target discipline benchmark literature units. Appropriate target discipline benchmark literature units can be screened out by analyzing information such as the title, abstract, and keywords of the literature units and combining them with recommendations from domain experts.

[0120] Step S152: extracting the domain feature representation of the target subject benchmark document unit as a benchmark feature representation, wherein the benchmark feature representation includes a benchmark semantic feature and a benchmark structural feature.

[0121] After selecting the target discipline benchmark document unit, its domain feature representation is extracted as the benchmark feature representation. This benchmark feature representation includes benchmark semantic features and benchmark structural features. In the previous step, domain feature extraction processing has been performed on each document unit in the multi-domain academic document collection. Therefore, the domain feature representation of the target discipline benchmark document unit can be directly extracted from the stored domain feature representation. The benchmark semantic features reflect the core content semantic information of the target discipline benchmark document unit, and the benchmark structural features reflect the structural association information of the target discipline benchmark document unit.

[0122] Step S153: extracting the domain feature representation of the knowledge transfer result as a transfer feature representation, wherein the transfer feature representation includes a transfer semantic feature and a transfer structural feature.

[0123] Similarly, a domain feature representation of the knowledge transfer results is extracted as a transfer feature representation. This transfer feature representation includes both transferred semantic features and transferred structural features. The knowledge transfer results undergo the same domain feature extraction process as the previous literature unit, including text preprocessing, semantic feature extraction, structural feature extraction, and normalization. The resulting transfer feature representation reflects the characteristic information of the knowledge transfer results and is used for comparison with the baseline feature representation.

[0124] Step S154: Calculate the semantic matching degree between the transferred semantic feature and the benchmark semantic feature, where the semantic matching degree is generated by comparing the overlapping ratio of feature subvectors representing the same research concept in the two features.

[0125] Calculate the semantic matching degree between the transferred semantic features and the baseline semantic features. The semantic matching degree is used to measure the degree of similarity between the semantic content of the knowledge transfer results and the semantic content of the benchmark literature unit in the target discipline. To calculate the semantic matching degree, it is first necessary to find the feature subvectors in the transferred semantic features and the baseline semantic features that represent the same research concept. Vector matching algorithms, such as the cosine similarity algorithm, can be used to compare each feature subvector of the transferred semantic features and the baseline semantic features pairwise to find feature subvector pairs with high similarity. Then, count the number of these feature subvector pairs representing the same research concept and calculate their overlap ratio in the total number of feature subvectors of the transferred semantic features and the baseline semantic features. This overlap ratio is the semantic matching degree. The closer it is to 1, the more similar the semantic content of the transferred semantic features is to the baseline semantic features.

[0126] Step S155: Calculating the structural matching degree between the migration structural feature and the reference structural feature, wherein the structural matching degree is generated by comparing the overlapping ratio of feature subvectors representing the same logical organization mode in the two features.

[0127] Calculate the structural match between the transferred structural features and the baseline structural features. The structural match is used to measure the degree of similarity between the structural organization of the knowledge transfer results and the structural organization of the benchmark literature units in the target discipline. Similar to calculating the semantic match, a vector matching algorithm is first used to find the feature subvectors in the transferred structural features and the baseline structural features that represent the same logical organization. For example, if both the transferred structural features and the baseline structural features have feature subvectors representing chapter hierarchical relationships, by comparing the similarities of these feature subvectors, pairs of feature subvectors representing the same chapter hierarchical relationships are found. Then, the number of these feature subvector pairs representing the same logical organization is counted, and the overlap ratio of these feature subvectors in the total number of feature subvectors in the transferred structural features and the baseline structural features is calculated. This overlap ratio is the structural match; the closer it is to 1, the more similar the structural organization of the transferred structural features and the baseline structural features is.

[0128] Step S156: Calculate a comprehensive matching value based on the semantic matching degree and the structural matching degree, wherein the comprehensive matching value integrates the semantic matching degree and the structural matching degree by weighted average.

[0129] Based on the calculated semantic and structural match values, a comprehensive match value is calculated. This comprehensive match value comprehensively reflects the degree of semantic and structural match between the knowledge transfer results and the benchmark literature units in the target discipline. To calculate the comprehensive match value, weights need to be assigned to the semantic and structural match values. The weight assignment needs to consider the importance of semantic content and structural organization in the target discipline. For example, if semantic content is more critical in the target discipline, a higher weight can be assigned to semantic match; if structural organization is more important for the expression and understanding of knowledge, a higher weight can be assigned to structural match. The comprehensive match value is obtained by multiplying the semantic match value by its weight and the structural match value by its weight, and then adding them together.

[0130] Step S157: Generate transfer effectiveness evaluation information indicating the reliability of knowledge transfer according to the comprehensive matching value, wherein the transfer effectiveness evaluation information includes the comprehensive matching value and the corresponding reliability level description.

[0131] Step S1571: establishing a mapping relationship table between comprehensive matching degree values and reliability levels, wherein the mapping relationship table includes reliability level labels corresponding to different comprehensive matching degree value intervals.

[0132] In this step, in order to accurately evaluate the reliability of knowledge transfer based on the comprehensive matching value, it is necessary to establish a mapping relationship table between the comprehensive matching value and the reliability level. This mapping relationship table will associate different comprehensive matching value intervals with corresponding reliability level labels. For example, setting a higher comprehensive matching value interval to correspond to the "high reliability" level label means that the comprehensive matching value within this interval indicates that the knowledge transfer result is very similar to the target subject benchmark document unit, and the reliability of the knowledge transfer is high; and a lower comprehensive matching value interval corresponds to the "low reliability" level label, indicating that the comprehensive matching value within this interval represents that the knowledge transfer result is quite different from the target subject benchmark document unit, and the reliability of the knowledge transfer is low. The establishment of this mapping relationship table needs to be based on a large amount of experimental data and field experience to ensure its rationality and accuracy.

[0133] Step S1572: Search the mapping relationship table according to the comprehensive matching value to determine the corresponding reliability level label.

[0134] After establishing the mapping relationship table, the calculated comprehensive matching value is searched in the mapping relationship table to determine the corresponding reliability level label. By comparing the comprehensive matching value with different intervals in the mapping relationship table, the interval in which the comprehensive matching value falls is found, and then the corresponding reliability level label is determined. This process can intuitively reflect the reliability of knowledge transfer.

[0135] Step S1573: extract the semantic mismatch part in the knowledge transfer result whose semantic matching degree is lower than a preset threshold, and record the specific content of the semantic mismatch part and the corresponding conversion difference between the source subject terminology and the target subject terminology.

[0136] To more comprehensively evaluate the results of knowledge transfer, it is necessary to extract semantically mismatched portions of the results where the semantic match falls below a preset threshold. The preset threshold is a pre-determined standard used to determine whether the semantic match meets the threshold. By comparing the transferred semantic features with the baseline semantic features, we identify portions where the semantic match falls below the threshold. For these semantically mismatched portions, we record their specific content in detail, including relevant text paragraphs and sentences. We also record the conversion differences between the corresponding source and target subject terms to analyze potential issues in the term replacement process.

[0137] Step S1574: extracting the structural mismatching portion of the knowledge transfer result whose structural matching degree is lower than a preset threshold, and recording the specific content of the structural mismatching portion and the corresponding organizational differences between the source discipline structure and the target discipline structure.

[0138] Similarly, we extract structural mismatches from the knowledge transfer results, where the degree of structural fit falls below a preset threshold. We compare the transferred structural features with the baseline structural features to identify those where the structural fit falls below the preset threshold. We record the specific content of these structural mismatches, such as chapter organization and paragraph arrangement. We also analyze the organizational differences between the source and target discipline structures to identify potential irregularities that may have arisen during the restructuring process, allowing us to improve the knowledge transfer process.

[0139] Step S1575: Integrate the comprehensive matching value, the reliability level label, the semantic mismatching part record, and the structural mismatching part record to generate migration effectiveness evaluation information including quantitative matching data and qualitative difference analysis.

[0140] Finally, the comprehensive match value, reliability level label, and records of semantic mismatches and structural mismatches are integrated. The comprehensive match value and reliability level label provide quantitative information on the reliability of knowledge transfer, while the records of semantic mismatches and structural mismatches provide qualitative difference analysis. This integration of information generates a comprehensive assessment of transfer effectiveness. This assessment provides users with a clear understanding of the effectiveness of knowledge transfer and any semantic and structural issues.

[0141] Throughout the entire process, the principles of legality, fairness, and privacy protection in data processing must be strictly adhered to to ensure the feasibility and compliance of the technical solution. When collecting privacy-sensitive data, technical measures such as data encryption and anonymization are employed to prevent its leakage and misuse. For example, during text preprocessing of document units, sections that may contain privacy-sensitive information are specially marked and processed, protecting data privacy without compromising knowledge transfer. During the construction and training of the AI model, the necessary modules, layers, and connections of the semantic encoding model, as well as the specific training steps and parameters, were clarified. The pretrained semantic encoding model includes modules such as a word embedding layer, a bidirectional long short-term memory network layer, an attention mechanism layer, and a fully connected layer. Trained on a large amount of text data, it learns the semantics and contextual information of natural language. When applying this semantic encoding model, text fragment units are used as input and output semantic features reflecting the core content of the document unit, effectively integrating the semantic encoding model with the academic literature domain.

[0142] Figure 2A schematic diagram of exemplary hardware and software components of a cross-domain knowledge transfer system 100 for academic literature that can implement the concepts of the present application, provided in some embodiments of the present application, is shown. For example, a processor 120 can be used in the cross-domain knowledge transfer system 100 for academic literature and used to perform the functions of the present application.

[0143] The cross-domain knowledge migration system 100 for academic documents can be a general-purpose server or a special-purpose server, both of which can be used to implement the cross-domain knowledge migration method for academic documents 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.

[0144] For example, the cross-domain knowledge migration system 100 for 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 cross-domain knowledge migration system 100 for 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 cross-domain knowledge migration system 100 for academic documents also includes an input / output I / O interface 150 between the computer and other input and output devices.

[0145] For ease of explanation, only one processor is described in the cross-domain knowledge transfer system 100 for academic documents. However, it should be noted that the cross-domain knowledge transfer system 100 for academic documents in this application may also include multiple processors, so the steps performed by one processor described in this application may also be performed jointly or individually by multiple processors. For example, if the processor of the cross-domain knowledge transfer system 100 for academic documents executes step A and step B, it should be understood that step A and step B may also be performed 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.

[0146] 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 cross-domain knowledge transfer method for academic literature is implemented.

[0147] 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 cross-domain knowledge transfer method for academic literature, characterized by: The method comprises: Acquire a multi-disciplinary academic document collection comprising multiple subject categories, wherein the multi-disciplinary academic document collection is composed of document units corresponding to different disciplines, and each document unit includes text content and field identification information of the discipline to which it belongs; Performing domain feature extraction processing on the multi-domain academic document collection to obtain a domain feature representation of each document unit, wherein the domain feature representation includes semantic features reflecting the core content of the document unit and structural features reflecting the structural association of the document unit; Constructing a cross-domain knowledge migration channel based on the domain feature representation, wherein the cross-domain knowledge migration channel is used to establish an association mapping relationship between feature representations of different subject domains; Using the cross-domain knowledge transfer channel, the knowledge of the source subject literature unit is transferred to the target subject literature unit to generate a knowledge transfer result of the target subject literature unit; Performing validity verification processing on the knowledge transfer result to generate transfer validity evaluation information indicating the reliability of the knowledge transfer; The domain feature extraction process is performed on the multi-domain academic document collection to obtain a domain feature representation of each document unit, wherein the domain feature representation includes semantic features reflecting the core content of the document unit and structural features reflecting the structural association of the document unit, including: Performing a text preprocessing operation on each document unit in the multi-field academic document collection, wherein the text preprocessing operation includes removing non-content symbols, filtering common stop words, and splitting continuous text paragraphs into processable text segment units; Calling a pre-trained semantic encoding model to perform contextual semantic analysis on the text fragment unit to generate semantic features reflecting the core content of the document unit, wherein the semantic features include contextual association information of key concepts in the text fragment unit and a semantic expression vector of the core argument; Parsing the text structure information of the document unit, wherein the text structure information includes the hierarchical relationship of chapter titles, the citation association relationship between tables and figures and the text, and the subject distribution relationship of references; Extracting structural features reflecting the structural association of document units based on the text structure information, wherein the structural features include a hierarchical depth parameter of chapter titles, a citation frequency parameter of tables and figures and text, and a subject concentration parameter of references; The semantic features and the structural features are standardized to generate a domain feature representation with a unified representation form, wherein the domain feature representation includes both semantic association information and structural association information.

2. The cross-domain knowledge transfer method for academic literature according to claim 1 is characterized in that: The calling of the pre-trained semantic encoding model to perform contextual semantic analysis on the text segment unit to generate semantic features reflecting the core content of the document unit includes: Inputting the text segment units into the word embedding layer of the semantic encoding model to generate initial word vector representations of the vocabulary in each text segment unit; Performing temporal context modeling on the initial word vector representation through the bidirectional long short-term memory network layer of the semantic encoding model to capture the semantic dependency of words in the text segment unit in the context of the preceding and following texts, and generating an intermediate semantic vector containing context information; Using the attention mechanism layer of the semantic encoding model to perform key information focusing processing on the intermediate semantic vector, identifying key concept words in the text segment unit that are strongly related to the core content of the document unit, and generating an attention weight distribution of the key concept words; Performing weighted aggregation processing on the intermediate semantic vectors according to the attention weight distribution to generate local semantic features reflecting contextual association information of key concept words; The local semantic features are input into the fully connected layer of the semantic encoding model for dimensionality compression processing to generate a semantic expression vector with a fixed dimensional representation, and the semantic expression vector maintains semantic information consistency with the local semantic features.

3. The cross-domain knowledge transfer method for academic literature according to claim 1 is characterized in that: The structural features reflecting the structural association of the document units are extracted based on the text structure information, and the structural features include the hierarchical depth parameters of the chapter titles, the citation frequency parameters of the tables and texts, and the subject concentration parameters of the references, including: Performing hierarchical depth calculation processing on the hierarchical relationship of chapter titles in the text structure information, counting the hierarchical distance of each chapter title relative to the root directory, and generating a hierarchical depth parameter for the chapter title, wherein the hierarchical depth parameter is used to indicate the core degree of the chapter content in the document unit; Performing citation frequency statistics on the citation association relationship between the charts and the text in the text structure information, counting the number of times each chart is cited by the text paragraph, and generating citation frequency parameters for the charts and the text, wherein the citation frequency parameters are used to indicate the closeness of the association between the chart content and the text content; Performing discipline concentration calculation on the discipline distribution relationship of the references in the text structure information, counting the proportion of documents belonging to the same discipline in the references, and generating a discipline concentration parameter for the references, wherein the discipline concentration parameter is used to indicate the degree of discipline focus of the research content of the document unit; Standardizing the hierarchical depth parameter, the citation frequency parameter, and the discipline concentration parameter to eliminate dimensional differences between different parameters and generate a set of structural features with comparable metrics; The structural feature set is subjected to feature fusion processing, and the hierarchical depth parameter, the citation frequency parameter and the subject concentration parameter are integrated into a single-dimensional structural feature by weighted summation, and the structural feature retains the original semantic information of each parameter.

4. The cross-domain knowledge transfer method for academic literature according to claim 1 is characterized in that: The cross-domain knowledge transfer channel is constructed based on the domain feature representation, and the cross-domain knowledge transfer channel is used to establish an association mapping relationship between feature representations of different subject domains, including: Selecting interdisciplinary document pairs containing a common research topic from the multi-field academic document collection, wherein the interdisciplinary document pairs are composed of document units belonging to different disciplines but with overlapping research topics; Extracting domain feature representations of source subject document units and target subject document units in the interdisciplinary document pair; Calculating a feature similarity value between the domain feature representation of the source subject document unit and the domain feature representation of the target subject document unit, wherein the feature similarity value is calculated using a cosine similarity algorithm; According to the feature similarity value, interdisciplinary document pairs with similarity values higher than a preset threshold are selected as key mapping samples; Performing association rule mining on the source subject domain feature representation and the target subject domain feature representation of the key mapping sample to generate a migration rule set reflecting the cross-disciplinary feature correspondence; The migration rule set is subjected to structured storage processing to generate a cross-domain knowledge migration channel for establishing an association mapping relationship between feature representations of different subject areas.

5. The cross-domain knowledge transfer method for academic literature according to claim 4 is characterized in that: The performing of association rule mining on the source subject domain feature representation and the target subject domain feature representation of the key mapping sample to generate a migration rule set reflecting the cross-disciplinary feature correspondence includes: Performing feature decomposition processing on the source subject domain feature representation of the key mapping sample to separate the semantic feature part and the structural feature part; Performing synchronous feature decomposition processing on the target subject domain feature representation of the key mapping sample to separate the corresponding semantic feature part and structural feature part; Establishing a semantic correspondence between the source subject semantic feature part and the target subject semantic feature part, wherein the semantic correspondence is generated by matching feature sub-vectors representing the same research concept in the two parts; Establishing a structural correspondence between the source discipline structural feature portion and the target discipline structural feature portion, wherein the structural correspondence is generated by matching feature subvectors representing the same logical organization mode in the two portions; Performing rule induction processing on the semantic correspondence and the structural correspondence to generate a migration rule set including semantic migration rules and structural migration rules, wherein the semantic migration rules are used to guide the conversion of source subject semantic features to target subject semantic features, and the structural migration rules are used to guide the conversion of source subject structural features to target subject structural features; Conflict detection is performed on the migration rule set to eliminate conflicting descriptions between different rules and generate a final migration rule set.

6. The cross-domain knowledge transfer method for academic literature according to claim 1 is characterized in that: The step of utilizing the cross-domain knowledge transfer channel to transfer the knowledge of the source subject literature unit to the target subject literature unit to generate the knowledge transfer result of the target subject literature unit includes: Selecting source subject document units and target subject document units that require knowledge transfer from the multi-field academic document collection, wherein the source subject document units contain target knowledge content to be transferred; Extracting domain feature representation of the source subject document unit, wherein the domain feature representation includes source subject semantic features and source subject structural features; The source subject semantic features are converted and processed by the semantic transfer rules in the cross-domain knowledge transfer channel to generate target semantic features that are dimensionally aligned with the target subject semantic features; The source discipline structural features are converted by using the structural migration rules in the cross-domain knowledge migration channel to generate target structural features that are dimensionally aligned with the target discipline structural features; Performing feature fusion processing on the target semantic features and the target structural features to generate a target subject domain feature representation; Based on the target subject domain feature representation, the target knowledge content of the source subject document unit is adaptively adjusted to generate a knowledge transfer result that conforms to the target subject knowledge expression specification. The knowledge transfer result includes adjusted semantic content and adjusted structural organization.

7. The cross-domain knowledge transfer method for academic literature according to claim 6 is characterized in that: The adaptive adjustment of the target knowledge content of the source subject literature unit based on the target subject domain feature representation to generate a knowledge transfer result that conforms to the target subject knowledge expression specification includes: Decomposing the target knowledge content of the source subject literature unit to separate the core argument part and the auxiliary argument part; According to the target semantic features in the target subject domain feature representation, term replacement processing is performed on the core argument part, replacing the source subject-specific terms with equivalent terms in the target subject, and generating semantically adapted core argument content; According to the target structural features in the target subject field feature representation, the auxiliary argument part is restructured, and the presentation order of the auxiliary argument content is reorganized according to the chapter hierarchy and citation association of the target subject literature unit to generate structurally adapted auxiliary argument content; The core argument content of the semantic adaptation and the auxiliary argument content of the structural adaptation are checked for content consistency, and the content that passes the consistency check is formatted and standardized, with unified punctuation usage specifications and paragraph separation rules, to generate knowledge transfer results that meet the knowledge expression specifications of the target discipline.

8. The cross-domain knowledge transfer method for academic literature according to claim 1 is characterized in that: The performing of validity verification processing on the knowledge transfer result to generate transfer validity evaluation information indicating the reliability of the knowledge transfer includes: Selecting a target subject benchmark document unit related to the research topic of the knowledge transfer result from the multi-field academic document collection, wherein the target subject benchmark document unit contains authoritative knowledge content verified by domain experts; Extracting a domain feature representation of the target subject benchmark document unit as a benchmark feature representation, wherein the benchmark feature representation includes a benchmark semantic feature and a benchmark structural feature; Extracting a domain feature representation of the knowledge transfer result as a transfer feature representation, wherein the transfer feature representation includes a transfer semantic feature and a transfer structural feature; Calculating a semantic matching degree between the transferred semantic feature and the benchmark semantic feature, wherein the semantic matching degree is generated by comparing an overlapping ratio of feature subvectors representing the same research concept in the transferred semantic feature and the benchmark semantic feature; Calculating a structural matching degree between the migration structural feature and the reference structural feature, wherein the structural matching degree is generated by comparing an overlapping ratio of feature subvectors representing the same logical organization mode in the migration structural feature and the reference structural feature; Calculating a comprehensive matching value based on the semantic matching degree and the structural matching degree, wherein the comprehensive matching value integrates the semantic matching degree and the structural matching degree by weighted average; The migration effectiveness evaluation information indicating the reliability of knowledge migration is generated according to the comprehensive matching value, and the migration effectiveness evaluation information includes the comprehensive matching value and the corresponding reliability level description.

9. A cross-domain knowledge transfer system for academic literature, characterized by: 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 cross-domain knowledge transfer method for academic literature as described in any one of claims 1 to 8 above.

Images