Double-layer adaptive fusion method and system for scientific and technical literature knowledge

Abstract

The invention provides a double-layer adaptive fusion method and system for scientific and technical literature knowledge, and relates to the technical field of language processing, and the method comprises the steps: constructing a double-layer adaptive fusion architecture; receiving a scientific and technical literature text and extracting a text basic feature vector; the general knowledge layer processing module performs general high-dimensional feature vector mapping, and the professional knowledge layer processing module performs professional high-dimensional feature vector mapping; and the adaptive fusion module performs field confidence calculation, configures an adaptive fusion weight parameter, performs weighted fusion on the general-text high-dimensional feature vector and the professional text high-dimensional feature vector, and applies an output module to perform processing of the scientific and technical literature by using the fused feature vector. According to the method and the device, the technical problem of inaccurate knowledge fusion caused by poor adaptability of a general model, limited generalization ability of a field-specific model and lack of adaptability of static knowledge integration in the prior art can be solved, and the accuracy of knowledge fusion is improved by constructing a double-layer adaptive fusion architecture.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of language processing, and particularly relates to a double-level adaptive fusion method and system for scientific literature knowledge. BACKGROUND

[0002] With the rapid development of artificial intelligence technology, the demand for domain knowledge in the field of natural language processing is increasing. As an important carrier of professional knowledge, the automatic processing of scientific literature faces the major challenge of the fusion of general knowledge and professional knowledge. However, existing natural language processing technologies still face some major challenges, especially in the understanding of professional terms and concepts in scientific literature. In the existing technology, the general language model performs poorly in the identification and semantic understanding of professional terms when processing professional field literature due to the lack of domain-specific training data. On the other hand, although the domain-specific model can provide better performance in a specific domain, its training data is limited to a single domain, lacking the ability to generalize across domains, making it difficult to handle multi-domain knowledge fusion tasks. In addition, the static knowledge integration method also cannot dynamically adjust the weight and contribution of knowledge according to the text content, resulting in insufficient accuracy of knowledge fusion, which affects the processing ability of different types of scientific literature.

[0003] In summary, the existing technology has the technical problem of insufficient accuracy of knowledge fusion due to the poor adaptability of general models in professional fields, the limited generalization ability of domain-specific models, and the lack of adaptability of static knowledge integration. SUMMARY

[0004] The present application provides a double-level adaptive fusion method and system for scientific literature knowledge, which is used to solve the technical problem of insufficient accuracy of knowledge fusion due to the poor adaptability of general models in professional fields, the limited generalization ability of domain-specific models, and the lack of adaptability of static knowledge integration in the existing technology.

[0005] In view of the above problems, the present application provides a double-level adaptive fusion method and system for scientific literature knowledge.

[0006] In a first aspect, the application provides a double-level adaptive fusion method of scientific literature knowledge, which is realized by a double-level adaptive fusion system of scientific literature knowledge. The double-level adaptive fusion method of scientific literature knowledge comprises: constructing a double-level adaptive fusion architecture, which comprises a general knowledge layer processing module, a professional knowledge layer processing module, an adaptive fusion module, and an application output module; the double-level adaptive fusion architecture receives scientific literature text and extracts a text basic feature vector; the general knowledge layer processing module performs general high-dimensional feature vector mapping on the text basic feature vector, outputs a general-text high-dimensional feature vector, and the professional knowledge layer processing module performs professional high-dimensional feature vector mapping on the text basic feature vector, outputs a professional text high-dimensional feature vector; the adaptive fusion module is used for performing domain confidence calculation on the text basic feature vector, configuring an adaptive fusion weight parameter according to the domain confidence calculation result, weighting and fusing the general-text high-dimensional feature vector and the professional text high-dimensional feature vector by using the adaptive fusion weight parameter, and outputting a text fusion feature vector, and the application output module performs processing of the scientific literature by using the fusion feature vector.

[0007] Optionally, the general knowledge layer processing module comprises a general knowledge graph, the general knowledge graph comprises general concept nodes and semantic relationship acquisition defined based on general knowledge samples; the general concept nodes are matched with the text basic feature vector by using the general knowledge graph, matched concept nodes are acquired, and concept nodes are expanded according to the semantic relationship acquisition; the text basic feature vector is mapped to a general high-dimensional feature vector in a high-dimensional vector space by using the matched concept nodes and the expanded concept nodes, and a general-text high-dimensional feature vector is outputted.

[0008] Optionally, the professional knowledge layer processing module comprises multi-domain knowledge samples; multi-source text domain feature vectors of the text basic feature vector are extracted, the multi-source text domain feature vectors are processed by professional text conversion according to the multi-domain knowledge samples, and standardized professional text feature vectors are obtained; the standardized professional text feature vectors are mapped to professional high-dimensional feature vectors in a high-dimensional vector space, and a professional text high-dimensional feature vector is outputted.

[0009] Optionally, a multi-label domain classifier is trained, wherein the multi-label domain classifier is trained by confidence training samples of a plurality of domain labels; the text basic feature vector is input into the multi-label domain classifier to perform confidence calculation on the plurality of domain labels respectively, and a plurality of domain confidences corresponding to the plurality of domain labels are output; and the plurality of domain confidences are normalized, and the normalized plurality of normalized domain confidences are output as domain confidence calculation results.

[0010] Optionally, the multi-label domain classifier is trained by confidence training samples of a plurality of domain labels, and the confidence training samples include corresponding domain professional term density samples, domain professional term proportion samples, citation mode samples, expression mode samples, and text structure samples under the plurality of domain labels.

[0011] Optionally, a hierarchical attention network is constructed, and the hierarchical attention network is connected with the multi-label domain classifier, wherein the hierarchical attention network includes a plurality of layers of domain features, and the plurality of layers of domain features include word-level domain features, sentence-level domain features, and paragraph-level domain features; domain feature extraction is performed on the text basic feature vector according to the hierarchical attention network, and a domain feature vector is output, and the domain feature vector is calculated by the multi-label domain classifier.

[0012] Optionally, a domain label of a first domain confidence in the domain confidence calculation result is obtained; and general knowledge layer processing weight and professional knowledge layer processing weight parameters are initialized according to the first domain confidence and the domain label, and adaptive fusion weight parameters are obtained from the general knowledge layer processing weight and the professional knowledge layer processing weight.

[0013] Optionally, a weight adjustment factor is constructed, the weight adjustment factor includes a text complexity adjustment factor, a domain professional term density adjustment factor, a context consistency adjustment factor, and a historical performance feedback adjustment factor; the adaptive fusion weight parameters are updated according to the weight adjustment factor, and updated adaptive fusion weight parameters are output; the updated adaptive fusion weight parameters are subjected to weight smoothing processing, and the general-text high-dimensional feature vector and the professional text high-dimensional feature vector are weighted and fused by the smoothed adaptive fusion weight parameters.

[0014] Optionally, a reward function is constructed, a real-time performance score result is output by performing performance scoring on the smoothed adaptive fusion weight parameters according to the reward function; the real-time performance score result is used as a feedback signal of adaptive reinforcement learning to perform adaptive updating on the smoothed adaptive fusion weight parameters until a preset performance score threshold is met, and an adaptive fusion weight parameter optimal solution is output.

[0015] In a second aspect, the present application also provides a two-level adaptive fusion system of scientific literature knowledge, configured to perform the two-level adaptive fusion method of scientific literature knowledge as described in the first aspect. The two-level adaptive fusion system of scientific literature knowledge comprises: an architecture organization unit configured to construct a two-level adaptive fusion architecture, the two-level adaptive fusion architecture comprising a general knowledge layer processing module, a professional knowledge layer processing module, an adaptive fusion module, and an application output module; a vector extraction unit configured to receive a scientific literature text by the two-level adaptive fusion architecture and extract a text basic feature vector; a feature mapping unit configured to perform general high-dimensional feature vector mapping on the text basic feature vector by the general knowledge layer processing module, output a general-text high-dimensional feature vector, and perform professional high-dimensional feature vector mapping on the text basic feature vector by the professional knowledge layer processing module, output a professional-text high-dimensional feature vector; a text fusion unit configured to perform domain confidence calculation on the text basic feature vector by the adaptive fusion module, configure adaptive fusion weight parameters according to the domain confidence calculation result, perform weighted fusion on the general-text high-dimensional feature vector and the professional-text high-dimensional feature vector by the adaptive fusion weight parameters, output a text fusion feature vector, and perform processing of the scientific literature by the application output module with the fusion feature vector.

[0016] The technical solutions provided in the present application have at least the following beneficial effects:

[0017] By constructing a double-level adaptive fusion architecture, the double-level adaptive fusion architecture includes a general knowledge layer processing module, a professional knowledge layer processing module, an adaptive fusion module and an application output module; the double-level adaptive fusion architecture receives scientific literature text and extracts a text basic feature vector; the general knowledge layer processing module maps the text basic feature vector to a general high-dimensional feature vector, outputs a general-text high-dimensional feature vector, and the professional knowledge layer processing module maps the text basic feature vector to a professional high-dimensional feature vector, and outputs a professional text high-dimensional feature vector; the adaptive fusion module is used for domain confidence calculation on the text basic feature vector, and the adaptive fusion weight parameter is configured according to the domain confidence calculation result, the general-text high-dimensional feature vector and the professional text high-dimensional feature vector are weighted and fused with the adaptive fusion weight parameter, and a text fusion feature vector is output, and the application output module executes the processing of the scientific literature with the fusion feature vector. That is, by constructing a double-level adaptive fusion architecture, containing a general knowledge layer processing module, a professional knowledge layer processing module, an adaptive fusion module and an application output module, the text basic feature vector is mapped to a general high-dimensional feature vector and a professional high-dimensional feature vector, the fusion weight is dynamically calculated according to the domain confidence calculation result, the contribution of general and professional knowledge is adjusted based on the domain feature, the adaptability of the general model in the professional field is enhanced, the cross-domain adaptability of the domain special-purpose model is ensured, the adaptability of knowledge fusion is improved, and the accuracy of knowledge fusion is improved. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the application or the prior art, the drawings needed in the embodiments or the prior art description will be briefly introduced below. Obviously, the drawings in the following description are only exemplary, and other drawings can be obtained by those skilled in the art without creative labor on the basis of the provided drawings.

[0019] Figure 1 A flowchart of a double-level adaptive fusion method for scientific literature knowledge of the application.

[0020] Figure 2 A general flowchart of the double-level adaptive fusion architecture of the application.

[0021] Figure 3 A structure diagram of a double-level adaptive fusion system for scientific literature knowledge of the application.

[0022] Explanation of reference signs: architecture organization unit 11, vector extraction unit 12, feature mapping unit 13, text fusion unit 14. DETAILED DESCRIPTION

[0023] This application provides a two-level adaptive fusion method and system for scientific and technological literature knowledge to address the technical problems in existing technologies, such as poor adaptability of general-purpose models in professional domains, limited generalization ability of domain-specific models, and lack of adaptability in static knowledge integration, leading to insufficient accuracy in knowledge fusion. By constructing a two-level adaptive fusion architecture, including a general knowledge layer processing module, a professional knowledge layer processing module, an adaptive fusion module, and an application output module, the method performs general high-dimensional feature vector mapping and professional high-dimensional feature vector mapping on the basic feature vectors of the text. It dynamically calculates the fusion weights based on the domain confidence calculation results, adjusting the contribution of general and professional knowledge based on domain features. This enhances the adaptability of the general model in professional domains while ensuring the cross-domain adaptability of the domain-specific model, improving the adaptability of knowledge fusion and thus enhancing the accuracy of knowledge fusion.

[0024] Example 1, as Figure 1 As shown, this application provides a two-level adaptive fusion method for scientific and technological literature knowledge. This method is applied to a two-level adaptive fusion system for scientific and technological literature knowledge. The specific steps of this method are as follows:

[0025] A two-level adaptive fusion architecture is constructed, which includes a general knowledge layer processing module, a professional knowledge layer processing module, an adaptive fusion module, and an application output module.

[0026] Specifically, a two-tiered adaptive fusion architecture is constructed, comprising a general knowledge layer processing module, a professional knowledge layer processing module, an adaptive fusion module, and an application output module. This two-tiered adaptive fusion architecture processes general knowledge and professional knowledge separately, ultimately fusing them through an adaptive algorithm. It can adapt to diverse text inputs, ensuring that the optimal knowledge fusion method is dynamically selected based on text features. The general knowledge layer processing module is built upon large-scale general knowledge graphs such as WordNet, ConceptNet, and Wikidata, covering basic concepts and common relationships in scientific literature, providing general semantic understanding capabilities. The general knowledge layer contains approximately 5 million concept nodes, covering basic scientific concepts, common academic terms, and common research methods. It establishes 25 basic semantic relationship types, including is-a, part-of, related-to, causes, and used-for, totaling over 20 million relation triples. Each relationship is weighted using a confidence score. Knowledge graph embedding algorithms such as TransE and RotatE are used to map the concepts and relationships of the general knowledge layer into a 512-dimensional continuous vector space, supporting efficient semantic similarity calculation and retrieval. These steps complete the construction of the general knowledge layer processing module.

[0027] The professional knowledge layer integrates professional ontologies and knowledge bases from multiple disciplines, covering professional terms, concept systems, and professional relationships in each field. It integrates professional ontologies from 15 main disciplines, including: medical field: MeSH (Medical Subject Headings), computer science: CSO (Computer Science Ontology), physics: Physics Subject Classification, chemistry: ChEBI (Chemical Entities of Biological Interest), biology: Gene Ontology, mathematics: Mathematics Subject Classification, engineering: Engineering Ontology. Standardize professional ontologies from different sources, unify concept identifiers, attribute descriptions, and relationship types. Establish multilingual mapping and synonym expansion for professional concepts to support processing of Chinese and English scientific literature. The professional knowledge layer uses the same 512-dimensional vector space as the general knowledge layer to ensure semantic representation compatibility between the two layers. At the same time, domain identification vectors are added to professional concepts to facilitate domain recognition and weight allocation.

[0028] The general knowledge layer processing module provides broad basic semantic understanding capabilities, and the professional knowledge layer processing module provides deep domain-specific knowledge support. The two-level adaptive fusion architecture takes into account the breadth and depth of knowledge, and supports flexible knowledge expansion and updating. The adaptive fusion module realizes intelligent fusion of general knowledge and professional knowledge through dynamic weight allocation and graph neural network mapping. The application output module provides unified semantic representation and knowledge enhancement results, supporting multiple downstream application tasks.

[0029] In addition, a cross-knowledge layer processing module can be added based on the two-level adaptive fusion architecture to specifically handle cross-disciplinary concepts, forming a three-level structure of general knowledge layer processing module, cross-knowledge layer processing module, and professional knowledge layer processing module. The cross-knowledge layer processing module contains concepts from interdisciplinary fields such as bioinformatics and computational physics. When processing interdisciplinary literature, the cross-knowledge layer processing module can specifically understand and map these cross-disciplinary concepts, thereby improving performance in handling complex interdisciplinary problems.

[0030] Alternatively, by constructing a knowledge pyramid structure from coarse granularity to fine granularity, different levels of concepts are processed, including four granularity levels of subject category layer, subject subdivision layer, professional direction layer, and specific concept layer. The subject category layer processes a wide range of subject categories (such as medicine, computer science, etc.); the subject subdivision layer processes the subdivision fields within the subject (such as genetics, ecology, etc. in biology); the professional direction layer processes more specific directions under the subject (such as cancer research, data science, etc.); and the specific concept layer processes specific concepts or terms (such as tumor markers, machine learning algorithms). When processing scientific literature, according to the needs of the content of the literature, the appropriate granularity layer is selected for knowledge processing. For example, a medical literature may need to start from the subject category layer (such as medicine), gradually refine to the professional direction layer (such as cancer research), and finally focus on the specific concept layer. The processing accuracy and granularity of each level increase as the level deepens.

[0031] Alternatively, the knowledge layer is dynamically constructed according to the characteristics of the input text, without pre-setting a fixed hierarchical structure, but temporarily combining related knowledge sources as needed. According to the content of the input text, the appropriate knowledge level is selected and combined in real time, and instead of fixedly using the general knowledge layer processing module, the cross-knowledge layer processing module or the professional knowledge layer processing module, the most suitable knowledge source is flexibly selected for processing. When processing a certain literature, the system first analyzes the characteristics of the text and determines the required knowledge layer according to the content. For example, if the literature discusses new methods for cancer, the professional layer knowledge base (such as the oncology ontology) is selected first, and then the general layer and cross-layer knowledge bases are dynamically introduced according to the interdisciplinary concepts mentioned in the literature (such as gene editing technology), which can be flexibly adjusted according to specific needs, without fixedly using the pre-set hierarchical structure.

[0032] By increasing the cross-knowledge layer, the multi-granularity knowledge pyramid, and dynamically constructing the knowledge level, different solutions are provided to improve the effect of literature processing, flexibly cope with the knowledge fusion task of interdisciplinary and complex literature, and exhibit stronger processing capacity and adaptability in multi-field applications.

[0033] The double-level adaptive fusion architecture receives scientific literature text and extracts a text basic feature vector.

[0034] Specifically, as shown in the accompanying drawings Figure 2As shown, the double-level adaptive fusion architecture receives scientific literature text input, supports multiple formats and encodings, and performs standardized preprocessing. Natural language processing techniques such as word segmentation, part-of-speech tagging, named entity recognition, etc. are used to preliminarily process the input scientific literature text. The processed scientific literature text is converted into a vector representation through word embedding algorithms such as Word2Vec, GloVe, BERT, etc., forming the basic feature vector of the scientific literature text. After word segmentation and processing, the scientific literature text is represented by words. For example, using BERT, each word or phrase is converted into a 512-dimensional vector that represents the semantic meaning of the word in the given context. Then, each word or phrase in the scientific literature text is mapped to a high-dimensional semantic vector, and the entire scientific literature text is aggregated or weighted summed to form a complete text basic feature vector, usually mapped to a high-dimensional space such as 512 dimensions, representing the overall semantic information of the scientific literature text. By converting the scientific literature text into a basic feature vector, the scientific literature text is efficiently represented and the core concepts and professional terms in the scientific literature text are understood.

[0035] The general knowledge layer processing module maps the text basic feature vector to a general high-dimensional feature vector, outputting a general-text high-dimensional feature vector, and the professional knowledge layer processing module maps the text basic feature vector to a professional high-dimensional feature vector, outputting a professional-text high-dimensional feature vector.

[0036] Further, the present application further includes the following steps: wherein the general knowledge layer processing module includes a general knowledge graph, the general knowledge graph includes general concept nodes and semantic relationships defined based on general knowledge samples; the general concept nodes are matched with the text basic feature vector using the general knowledge graph, the matching concept nodes are obtained, and the concept nodes are expanded according to the semantic relationship acquisition; the text basic feature vector is mapped to a general high-dimensional feature vector in a high-dimensional vector space by the matching concept nodes and the expanded concept nodes, and a general-text high-dimensional feature vector is outputted.

[0037] Further, the present application further includes the following steps: wherein the professional knowledge layer processing module includes multi-domain knowledge samples; multi-source text domain feature vectors of the text basic feature vector are extracted, the multi-source text domain feature vectors are processed into standardized professional text feature vectors according to the multi-domain knowledge samples; the standardized professional text feature vectors are mapped to professional high-dimensional feature vectors in a high-dimensional vector space, and professional-text high-dimensional feature vectors are outputted.

[0038] In particular, the general knowledge layer processing module includes a general knowledge graph, which is a knowledge graph covering a wide range of fields, containing general concept nodes and semantic relationships between general concept nodes. Based on existing large-scale knowledge graphs (such as WordNet, ConceptNet, Wikidata, etc.), a general knowledge graph is constructed, containing about 5 million general concept nodes, covering general knowledge such as basic scientific concepts, academic terminology, and common research methods. When processing input text, the general knowledge graph is used to match the text basic feature vector with general concept nodes to obtain matching concept nodes. According to the semantic relationship between nodes, related expansion concept nodes are obtained. Using the matching concept nodes and expansion concept nodes, the text basic feature vector is mapped to a general high-dimensional feature vector in a high-dimensional vector space (512-dimensional vector space). Each matching concept node and expansion concept node in the text will be mapped to a high-dimensional vector (usually a 512-dimensional vector), combined together to form the final general-text high-dimensional feature vector.

[0039] The professional knowledge layer processing module includes multi-field knowledge samples, i.e. knowledge datasets covering multiple fields, including professional terms, concepts and semantic relationships from different disciplines. The multi-source text field feature vector of the text basic feature vector is extracted, i.e. the text feature vector extracted from multiple fields, covering knowledge information from different fields. The multi-source text field feature vector is processed by the field knowledge sample, i.e. the multi-source text field feature vector is converted into a standardized text feature vector conforming to a certain specific field. That is, the different sources of professional ontology are standardized, and the concept identifier, attribute description and relationship type are unified. Multi-language mapping and synonym expansion of professional concepts are established to support processing of Chinese and English scientific literature. The standardized professional text feature vector is a text feature vector processed by the professional field conversion process, which ensures that text feature vectors in different fields have consistent representation and scale. The standardized professional text feature vector will be further mapped to a high-dimensional vector space, such as a 512-dimensional space, by using a professional knowledge graph or a deep learning embedding algorithm (such as TransE, RotatE, etc.), to ensure that each text feature vector can efficiently represent the semantic information of the professional field. In the high-dimensional vector space, each concept, term and relationship exists in the form of a vector, and more accurate semantic understanding can be obtained through vector calculation.

[0040] The text basic feature vector is a preliminary feature representation extracted from scientific literature text, usually including word frequency, word sequence, etc. The general high-dimensional feature vector mapping is a process of converting the text basic feature vector to a high-dimensional space, aiming to capture general knowledge features in the text. The professional high-dimensional feature vector mapping is similar to the general high-dimensional feature vector mapping, but focuses on capturing professional knowledge features in the text.

[0041] The adaptive fusion module is used for domain confidence calculation on the text-based feature vector, and an adaptive fusion weight parameter is configured according to the domain confidence calculation result, the general-text high-dimensional feature vector and the professional text high-dimensional feature vector are weighted and fused with the adaptive fusion weight parameter, and a text fusion feature vector is output, and the application output module executes the processing of the scientific literature with the fusion feature vector.

[0042] Further, the application further includes the following steps: training a multi-label domain classifier, wherein the multi-label domain classifier is trained by confidence training samples of a plurality of domain labels; inputting the text-based feature vector into the multi-label domain classifier to perform confidence calculation on the plurality of domain labels respectively, and outputting a plurality of domain confidences corresponding to the plurality of domain labels; and performing normalization processing on the plurality of domain confidences, and outputting the normalized domain confidences as the domain confidence calculation result.

[0043] The multi-label domain classifier is trained by confidence training samples of a plurality of domain labels, and the confidence training samples include corresponding domain professional term density samples, domain professional term proportion samples, citation mode samples, expression mode samples, and text structure samples under the plurality of domain labels.

[0044] A hierarchical attention network is constructed, and the hierarchical attention network is connected with the multi-label domain classifier, wherein the hierarchical attention network includes a plurality of layers of domain features, and the plurality of layers of domain features include word-level domain features, sentence-level domain features, and paragraph-level domain features; domain feature extraction is performed on the text-based feature vector according to the hierarchical attention network, and a domain feature vector is output, and the domain feature vector is calculated by the multi-label domain classifier.

[0045] Specifically, a multi-label domain classifier based on BERT is constructed, the multi-label classifier is a machine learning model capable of assigning multiple labels to each input text, and supports simultaneously recognizing the confidence distribution of the text in 15 main subject areas. The classifier architecture is: BERT encoder + full connection layer + Sigmoid activation function, outputting a 15-dimensional domain confidence vector. A special domain feature extractor is designed, including professional term density calculation for statistical analysis of the proportion of professional terms in the text, citation mode analysis for analyzing the domain distribution characteristics of the literature cited, expression mode recognition for identifying the expression mode specific to different domains, and text structure feature analysis for analyzing the structure characteristics of the text (such as IMRaD structure).

[0046] In multi-label classification, each label usually comes with a confidence score, indicating the degree or probability of the label applying to the text. The higher the confidence score, the more suitable the domain label is for the current text. Confidence training samples are samples with confidence values for each domain label during the training process, which help the multi-label domain classifier learn how to assign multiple domain labels to each text and estimate the confidence of the labels. Domain-specific term density samples measure the frequency of specific domain-specific terms in the text, domain-specific term proportion samples measure the proportion of specific domain-specific terms in all terms in the text, citation pattern samples represent the distribution characteristics of cited literature sources in the analyzed literature, expression pattern samples are the language style and term usage habits of the text, and text structure samples are the structured layout of the literature, such as the IMRaD structure (Introduction-Method-Result-Discussion structure).

[0047] The domain feature extractor first calculates the density of professional terms in the text, i.e., the frequency of domain-related terms. Specifically, a domain term library is established, such as a medical domain term library (e.g., MeSH term table); each word in the text is analyzed to determine whether it belongs to a term in the domain term library; and the term density is calculated by calculating the frequency of the term, such as the density of computer science terms being 20% of all terms in the text.

[0048] The proportion of terms related to a certain domain in the text reflects the proportion of professional terms relative to the total number of terms in the text, which is used to assess the domain orientation of the text. By matching with the domain term library, the proportion of professional terms in the text is calculated. According to the proportion, the strength of the domain to which the text belongs is inferred. If the term proportion is high, it indicates that the text is strongly related to the domain.

[0049] The sources cited in the literature often reflect the domain to which it belongs. Medical literature may cite a large number of articles from medical journals, while computer science literature may cite technical conference papers. By analyzing the sources and types of citations, the accuracy of domain judgment is further enhanced. That is, the references in the literature are analyzed to extract the source types (such as journals, conferences, books, etc.) of the citations, and it is determined whether these citations conform to the characteristics of literature in a certain domain. For example, if the literature mostly cites medical journals, it is likely to belong to the medical domain.

[0050] Expression pattern recognition helps to determine the domain of the literature by analyzing the language style, term usage, etc. of the text. Different domains have their own specific language patterns and terms. For example, medical literature often uses pathological terms and treatment methods, while computer science literature often uses algorithm descriptions and technical terms. Specifically, natural language processing techniques (such as part-of-speech tagging, named entity recognition, etc.) are used to analyze the language features in the text. Based on these language features, it is determined whether the text conforms to the expression pattern of a certain domain.

[0051] The structural features of the document are also important basis for judging the field. For example, medical literature often follows the IMRaD structure (introduction, method, result, discussion), while computer science literature may use more experimental design, algorithm description, etc. Analyze the structural layout of the text, identify the chapters, paragraphs and their layout in the text (such as introduction, method, experiment, result, etc.). According to these structural features, infer the field of the document. For example, if the document contains a lot of content about experimental methods and result analysis, it is judged as a physics document. Combine the features of professional term density, term proportion, citation pattern, expression pattern and text structure as input features for training the multi-label field classifier. Through these features, the multi-label field classifier can judge multiple field labels of the document and output the confidence of each label.

[0052] On the basis of the BERT encoder, a hierarchical attention mechanism is added to calculate the field features at the word level, sentence level and paragraph level. Word-level attention focuses on professional terms, sentence-level attention focuses on expression patterns, and paragraph-level attention focuses on topic structure. Hierarchical attention network is a neural network structure used to focus on different levels of features through multiple layers. Hierarchical attention network extracts more accurate field features by focusing on different levels of text (such as word, sentence, paragraph level). Each layer of attention mechanism focuses on relevant domain information at a specific level, thereby improving the expression ability of text field features.

[0053] The hierarchical attention network includes multiple layers of field features, including word-level field features, sentence-level field features and paragraph-level field features. Word-level attention features focus on professional terms in the text, that is, whether each word in the text belongs to the terminology of a certain field. By attention weighting each word, the importance and field features of each word are extracted. Sentence-level field features focus on expression patterns in the text, such as certain field expression methods or term collocations. Paragraph-level attention focuses on the topic structure in the text, such as whether a paragraph focuses on a specific topic of a certain field.

[0054] The layered attention network is used to extract domain features from the text basic feature vector, and output a domain feature vector. The domain information in the text basic feature vector is extracted according to different levels of features. By learning the contribution value of each word in different domains, the domain feature vector of each word is output. According to the expression mode and language characteristics of each sentence, a domain vector is generated for the sentence. By analyzing the structure and theme of the paragraph, a paragraph-level domain vector is generated. The attention mechanism of each layer adjusts the contribution of the features according to its weight. After three layers of attention mechanism processing, the word-level, sentence-level and paragraph-level domain features are combined to form the final domain feature vector, which not only contains the syntax features and lexical information of the text, but also combines the knowledge of the professional domain. The layered attention network is connected with the multi-label domain classifier. The domain feature vector extracted by the layered attention network is sent to the multi-label domain classifier to calculate the confidence of each domain label. The domain confidence represents the matching degree of a domain label and the text. The higher the confidence, the stronger the applicability of the domain label to the text.

[0055] The domain feature vector is input into the multi-label domain classifier and matched with multiple domain labels. For each label, the multi-label domain classifier calculates a score (score i ) representing the applicability of the label to the text based on the relationship between the text features and the domain label. After obtaining the original scores of multiple domain labels, the scores are normalized using the softmax function. The purpose of normalization is to convert the confidence of all labels into a probability distribution, ensuring that their sum is 1. The confidence calculation formula is confidence i = softmax(score i + smooth factor ), where score i is the original score of domain label i, smooth factor is a smoothing factor, and smooth factor= a * avg(scores), where a is a smoothing factor and avg(scores) is the average score. Generally, a smaller a value will result in a weaker smoothing effect, while a larger a value will increase the smoothing effect. Normalization is the process of adjusting the confidence scores of all domain labels so that their sum is 1. Through normalization, it is ensured that the confidence scores of each domain label are relative. Label smoothing is a regularization technique aimed at preventing the multi-label domain classifier from being overly confident in its predictions for certain domain labels. By smoothing the predicted confidence scores, extreme predictions are reduced, providing a more reliable estimate of confidence, often by adjusting the confidence scores with the addition of a smoothing factor. The normalized multiple normalized domain confidences are output as the domain confidence calculation results after normalization processing according to the multiple domain confidences.

[0056] For each domain label, its original score score i is added to a smoothing factor smooth factor , and then normalized using the softmax function to obtain the final confidence confidence i of the domain label. The softmax function converts the confidence values of all domain labels into a probability distribution, ensuring that the sum of the confidences of all labels is 1. For example, suppose after the multi-label domain classifier calculation, the following domain labels are obtained: Computer Science 2.1, Artificial Intelligence 1.8, Medicine 1.2. Set the smoothing factor a to 0.1, then avg(scores) = (2.1 + 1.8 + 1.2) / 3 = 1.7, and smooth factor = 0.1 * 1.7 = 0.17. Using the softmax function to calculate the confidence of each label: the confidence of the Computer Science label is 0.46, the confidence of the Artificial Intelligence label is 0.34, and the confidence of the Medicine label is 0.20.

[0057] A dynamic confidence threshold adjustment mechanism is established to automatically adjust the confidence threshold for domain judgment according to the length, complexity and historical processing results of the text. When the highest confidence exceeds the dynamic threshold, it is considered that the text has a clear domain inclination. The dynamic threshold adjustment mechanism automatically adjusts the confidence threshold used to determine whether the domain label is valid according to the characteristics of the input text (such as text length, complexity) and historical processing results. That is, different texts may require different confidence thresholds to determine their domain labels. For example, for shorter texts, the confidence threshold of the domain label is lowered; while for long and complex texts, the threshold may be raised to ensure the accuracy of the judgment. When the confidence of the text's domain label exceeds the dynamic threshold, it is considered that the text has a clear domain inclination. Otherwise, it is considered that the text belongs to multiple domains or is uncertain about the attribution to a certain domain.

[0058] In addition, ensemble learning methods such as random forests, gradient boosting, etc. can be used instead of a single BERT classifier to calculate the domain confidence through the voting mechanism of multiple weak learners to obtain more accurate domain confidence. That is, each learner gives a domain confidence prediction value. The prediction results of all learners are calculated through some integration strategy (such as voting mechanism) to calculate the final domain confidence. Alternatively, use unsupervised clustering algorithms such as K-means, DBSCAN, etc. Unsupervised clustering is a machine learning method that does not require labeled data, which automatically groups data through similarity measurement. The clustering results based on text features are used to calculate the domain confidence, and the training process does not require labeled data. Alternatively, construct a keyword dictionary for each domain, each domain has its corresponding keyword dictionary, such as computer science which may include keywords such as algorithms, programming, data structures, etc. Calculate the degree of matching of keywords in the text with the keyword dictionary of each domain to calculate the domain confidence. TF-IDF (Term Frequency-Inverse Document Frequency) is a commonly used text feature extraction method to measure the importance of words in a text. TF is the frequency of a word in a document, and IDF is the rarity of a word in all documents. TF-IDF calculates the weighted value of each word to evaluate its contribution to the document. By calculating the matching degree of keywords in the text with the keyword dictionary of each domain, combined with TF-IDF weighting, the confidence of the text belonging to each domain can be obtained.

[0059] Further, the application further includes the following steps: obtaining the domain label of the first domain confidence in the domain confidence calculation result; initializing the general knowledge layer processing weight and the professional knowledge layer processing weight parameter according to the first domain confidence and the domain label, and obtaining the adaptive fusion weight parameter from the general knowledge layer processing weight and the professional knowledge layer processing weight.

[0060] A weight adjustment factor is constructed, which includes a text complexity adjustment factor, a domain professional term density adjustment factor, a context consistency adjustment factor, and a historical performance feedback adjustment factor. The adaptive fusion weight parameter is updated according to the weight adjustment factor, and the updated adaptive fusion weight parameter is output. The updated adaptive fusion weight parameter is subjected to weight smoothing processing, and the general-text high-dimensional feature vector and the professional text high-dimensional feature vector are weighted and fused with the smoothed adaptive fusion weight parameter.

[0061] A reward function is constructed, and the performance score of the smoothed adaptive fusion weight parameter is calculated according to the reward function, and the real-time performance score result is output. The real-time performance score result is used as the feedback signal of adaptive reinforcement learning to adaptively update the smoothed adaptive fusion weight parameter until the preset performance score threshold is met, and the adaptive fusion weight parameter optimal solution is output.

[0062] Specifically, among the multiple field confidences, the first field confidence refers to the highest field confidence in the field confidence calculation result. The field label is the label corresponding to the first field confidence, i.e., the field label with the strongest association with the scientific literature text. According to the first field confidence and the field label, the general knowledge layer processing weight and the professional knowledge layer processing weight are initialized. The professional knowledge layer processing weight is the maximum field confidence in the field confidence calculation result based on the field confidence calculation basis weight distribution, and the general knowledge layer processing weight is calculated by 1-the maximum field confidence in the field confidence calculation result. For example, the first field confidence is 0.85, the corresponding field label is computer science, the professional knowledge layer processing weight is 0.85, and the general knowledge layer processing weight is 0.15.

[0063] According to the calculated general knowledge layer processing weight and professional knowledge layer processing weight, the influence of general knowledge and professional knowledge is weighted in the fusion process. The weight adjustment factor is a plurality of factors used to dynamically adjust the adaptive fusion weight parameter, which is usually related to the characteristics or context information of the text, and is used to flexibly adjust the weight according to different text features and historical data, so as to accurately reflect the domain features of the text in the fusion process. The weight adjustment factor includes a text complexity adjustment factor, a domain professional term density adjustment factor, a context consistency adjustment factor, and a historical performance feedback adjustment factor. The text complexity adjustment factor is used to increase the general knowledge weight for texts with high complexity. Texts with high complexity may contain more details or more complex terms, so more general knowledge layers are needed to assist understanding, thereby dynamically increasing the weight of the general knowledge layer. The domain professional term density adjustment factor is used to increase the professional knowledge weight for texts with high professional term density. Texts with high density (i.e., texts containing a large number of domain professional terms) usually have a strong association with the domain, so the weight of the professional knowledge layer needs to be increased. The context consistency adjustment factor is used to adjust the weight distribution according to the context consistency. If the text maintains high consistency within a certain domain, the weight of the domain can be appropriately increased; otherwise, the weight of the general knowledge may need to be increased. The historical performance feedback adjustment factor is used to adjust the weight parameter according to the historical processing effect.

[0064] According to the weight adjustment factor, the adaptive fusion weight parameter is updated, and the general knowledge layer processing weight and the professional knowledge layer processing weight parameter are adjusted to obtain the updated adaptive fusion weight parameter. After each weight update, in order to avoid excessive fluctuations, a weight smoothing mechanism is applied. A smoothing coefficient (λ) is used to balance the influence of the current weight and the historical weight. The smoothing coefficient usually takes a value between 0.1 and 0.3, a lower value focuses more on the current weight, and a higher value retains more historical weight. The weight smoothing formula is: smooth = λ × αprevious + (1 - λ) x a current , β smooth = λ x β previous + (1 - λ) x β current where λ is a smoothing coefficient, usually taking values from 0.1 to 0.3; a previous and β previous are the weights of the previous round, a current and β current are the weights of the current round. At the same time, weight constraints are set, in order to ensure the rationality of the fusion process, the weights of the general knowledge layer and the professional knowledge layer must be 1, that is, a smooth + β smooth = 1 (weight normalization constraint). In order to avoid too small weight of a layer, set the minimum weight limit, a smooth ≥ 0.1, β smooth ≥ 0.1 (minimum weight constraint). In order to prevent excessive fluctuations in the weight update process, set the limit of weight change amplitude: | a smooth - a previous | ≤ 0.2 (change amplitude constraint). After smoothing and weight constraints, the updated adaptive fusion weight parameters are output, which are used for weighted fusion of general-text high-dimensional feature vectors and professional text high-dimensional feature vectors. For example, a literature about the application of deep learning in medical imaging involves two fields of computer science and medicine. After calculation by the multi-label field classifier, the output confidence is: computer science 0.85, artificial intelligence 0.80, medicine 0.70. Assuming that the text is a complex literature in the field of deep learning, increase the weight of the general knowledge layer, such as to 0.2; the literature contains a large number of medical terms, increase the weight of the professional knowledge layer to 0.75; the medical related content in the literature has high consistency, and the weight of the medical field is increased according to the consistency adjustment; according to the historical data feedback, increase the weight of the professional knowledge layer, such as 0.7. Through these adjustments, the final adaptive fusion weight parameters are: general knowledge layer processing weight 0.2, professional knowledge layer processing weight 0.8.

[0065] By dynamically adjusting the weights of general-text high-dimensional feature vectors and professional text high-dimensional feature vectors, it is flexible to adapt to the characteristics of different field texts. Through the weight adjustment factor and the weight smoothing mechanism, general knowledge and professional knowledge are fused, thereby improving the understanding ability of the text.

[0066] The reward function is used to measure the contribution of the current policy (or current weight distribution) to the task goal, and gives a reward value or penalty value according to the performance of the model in a specific task (such as accuracy, recall, etc.). The design of the reward function directly affects the optimization direction in the learning process. The design of the reward function is the core of reinforcement learning. According to the performance of the downstream task (such as text classification accuracy, recall, F1-score, etc.), the reward function is designed to evaluate the pros and cons of the current weight distribution. The goal of the reward function is to give feedback signals according to the effect of the current weight distribution (for example, the accuracy of text classification). Higher accuracy and F1-score will bring higher rewards, and vice versa.

[0067] After each weight update, the performance of the downstream task (such as text classification) is evaluated in real time according to the current weights. The real-time performance score results reflect the current running effect of the model, helping to judge whether the current weights need to be further optimized. Through the reward function, the performance of the task is quantified, and the real-time performance score is obtained. For example, output a score between 0 and 1, indicating the performance of the current model in the task.

[0068] The policy network is a neural network responsible for learning the optimal weight distribution strategy. The policy network will adjust the model parameters according to the historical performance feedback (including the score given by the reward function) to optimize the distribution of general knowledge layer processing weights and professional knowledge layer processing weights. During the training of the policy network, the neural network continuously adjusts the weight distribution strategy through multiple updates and optimizations, so as to achieve the optimal task performance in the long-term learning. During the training process, the policy network uses gradient descent algorithm (such as Adam optimizer) to minimize the loss function, so that the model can automatically adjust the weight distribution according to the needs of the task.

[0069] Online learning allows the strategy to be updated according to new data and performance feedback in real-time operation. After obtaining the real-time performance score results each time, the policy network will adaptively update according to these feedback information. For example, if the current adaptive fusion weight causes the model's accuracy in the classification task to decrease, adjust the weight strategy according to the real-time score feedback (such as accuracy reduction) to gradually optimize the fusion result. After each real-time update, the policy network will adjust the weights of the general knowledge layer and the professional knowledge layer appropriately to ensure that it can adapt to the features of different texts.

[0070] A preset performance score threshold, such as 0.9, is set, which means that when the real-time performance score reaches or exceeds the threshold, it is considered that the current weight distribution strategy is the optimal solution. When the performance score reaches this preset threshold, stop adaptive updating and output the optimal adaptive fusion weight parameter. After multiple rounds of optimization and feedback, the optimal solution of the adaptive fusion weight parameter is obtained, which is the optimal weight distribution strategy based on reinforcement learning, which can ensure dynamic adjustment of the fusion mode of general knowledge and professional knowledge when processing different texts.

[0071] The concept mapping relationship between different domain ontologies is learned using graph neural network technology to realize automatic association and semantic enhancement of cross-domain knowledge. The general knowledge layer and 15 professional knowledge layers are constructed into a unified multi-domain knowledge graph: node types: general concept nodes, professional concept nodes, bridge concept nodes, edge types: intra-layer relationship edges, cross-layer mapping edges, similarity relationship edges, attribute information: concept description, domain identifier, confidence score. Heterogeneous graph neural networks are used to process multiple types of nodes and relationships, and the network hierarchy includes an input layer: node feature embedding (512-dimensional vector), a hidden layer: 3 layers of GraphSAGE convolutional layers, each layer with 256 dimensions, an attention layer: multi-head attention mechanism (8 attention heads), and an output layer: final concept representation vector (512-dimensional).

[0072] A special cross-domain mapping learning algorithm is designed to calculate the similarity score of cross-domain concepts based on the semantic description, context distribution, and co-occurrence pattern of concepts. The GNN model is trained through contrastive learning to learn the mapping relationship of cross-domain concepts: Loss = Σ[max(0, margin + sim(a, n) - sim(a, p))], where a is the anchor concept, p is the positive sample concept, and n is the negative sample concept. The semantic information of cross-domain is propagated and aggregated in the knowledge graph using the message propagation mechanism on the graph. A dynamic mapping weight learning mechanism is established: dynamically adjust the weight of different mapping relationships according to the text content, automatically find the most relevant cross-domain concepts through the attention mechanism, and support fast integration and mapping relationship learning of new domain ontologies.

[0073] An automatic evaluation mechanism for mapping quality is designed, which evaluates the mapping accuracy based on the gold standard annotated by humans, optimizes the mapping quality through the performance feedback of downstream tasks, and establishes a confidence score mechanism for mapping relationships.

[0074] The semantic information of the double-level knowledge is fused into a unified representation vector, which not only retains the universality of general knowledge, but also enhances the depth of professional knowledge. Multi-level semantic features are extracted from the general knowledge layer and the professional knowledge layer respectively. From the general knowledge layer, basic semantic features (word semantic representation based on Word2Vec, GloVe, etc.), relationship semantic features (relationship semantic representation based on knowledge graph embedding), and context semantic features (context semantic representation based on BERT model, etc.) are extracted. From the professional knowledge layer, domain term features (domain-specific semantic representation of professional terms), concept hierarchy features (concept semantic representation based on ontology hierarchy), and professional relationship features (semantic representation of domain-specific relationships) are extracted. A multi-head attention mechanism is designed to realize weighted fusion of features, and the calculation formula of the multi-head attention mechanism is: wherein Q is a query vector, K is a key vector, and V is a value vector. Multi-head attention fusion: MultiHead(Q,K,V)=Concat(head1,...,head h )W O , wherein head i =Attention(QW i Q ,KW i K ,VW i V ) is realized. The hierarchical semantic fusion from the word level to the document level is realized. The word-level fusion is used to fuse the general semantics and professional semantics of words to form enhanced word representations; the phrase-level fusion is used to fuse phrase semantics based on dependency relationships and grammatical structures; the sentence-level fusion is used to integrate the fused semantics of all words and phrases in a sentence; and the document-level fusion is used to aggregate the hierarchical semantic information of the entire document. A dynamic fusion strategy optimization mechanism is established, including a strategy network (training a small neural network to learn the optimal fusion strategy), reinforcement learning optimization (optimizing fusion parameters through reinforcement learning), and multi-task learning (jointly optimizing the fusion strategy on multiple downstream tasks). A quality control mechanism for the fusion results is established, including semantic consistency verification (ensuring that the semantic representation after fusion remains consistent), information retention degree evaluation (evaluating the information retention degree in the fusion process), and performance gain analysis (analyzing the performance improvement of the downstream tasks due to fusion).

[0075] In summary, the double-level adaptive fusion method for scientific literature knowledge provided by the present application has the following beneficial effects:

[0076] By constructing a double-level adaptive fusion architecture, the double-level adaptive fusion architecture includes a general knowledge layer processing module, a professional knowledge layer processing module, an adaptive fusion module and an application output module; the double-level adaptive fusion architecture receives scientific literature text and extracts a text basic feature vector; the general knowledge layer processing module maps the text basic feature vector to a general high-dimensional feature vector, outputs a general-text high-dimensional feature vector, and the professional knowledge layer processing module maps the text basic feature vector to a professional high-dimensional feature vector, outputs a professional text high-dimensional feature vector; the adaptive fusion module is used for domain confidence calculation on the text basic feature vector, and the adaptive fusion weight parameter is configured according to the domain confidence calculation result, the general-text high-dimensional feature vector and the professional text high-dimensional feature vector are weighted and fused with the adaptive fusion weight parameter, and a text fusion feature vector is output, and the application output module executes the processing of the scientific literature with the fusion feature vector. That is, by constructing a double-level adaptive fusion architecture, containing a general knowledge layer processing module, a professional knowledge layer processing module, an adaptive fusion module and an application output module, the text basic feature vector is mapped to a general high-dimensional feature vector and a professional high-dimensional feature vector, the fusion weight is dynamically calculated according to the domain confidence calculation result, the contribution degree of general and professional knowledge is adjusted based on the domain feature, the adaptability of general model in professional field is enhanced, at the same time the cross-domain adaptability of domain special model is guaranteed, the adaptability of knowledge fusion is improved, and the accuracy of knowledge fusion is improved.

[0077] In the second embodiment, based on the same inventive concept as the double-level adaptive fusion method of scientific literature knowledge in the foregoing first embodiment, the present application also provides a double-level adaptive fusion system of scientific literature knowledge, please refer to the attached Figure 3 , the double-level adaptive fusion system of scientific literature knowledge includes:

[0078] The architecture organization unit 11 is configured to construct a double-level adaptive fusion architecture, which comprises a general knowledge layer processing module, a professional knowledge layer processing module, an adaptive fusion module and an application output module; the vector extraction unit 12 is configured to receive a scientific literature text by the double-level adaptive fusion architecture and extract a text basic feature vector; the feature mapping unit 13 is configured to perform general high-dimensional feature vector mapping on the text basic feature vector by the general knowledge layer processing module, output a general-text high-dimensional feature vector, and perform professional high-dimensional feature vector mapping on the text basic feature vector by the professional knowledge layer processing module, output a professional text high-dimensional feature vector; the text fusion unit 14 is configured to perform domain confidence calculation on the text basic feature vector by the adaptive fusion module, configure an adaptive fusion weight parameter according to the domain confidence calculation result, perform weighted fusion on the general-text high-dimensional feature vector and the professional text high-dimensional feature vector by the adaptive fusion weight parameter, output a text fusion feature vector, and perform processing of the scientific literature by the application output module with the fusion feature vector.

[0079] Further, the feature mapping unit 13 in the double-level adaptive fusion system for scientific literature knowledge is further configured to: wherein the general knowledge layer processing module comprises a general knowledge graph, the general knowledge graph comprises general concept nodes and semantic relationship acquisition defined based on general knowledge samples; the general concept nodes are matched with the text basic feature vector by using the general knowledge graph, matched concept nodes are acquired, and concept nodes are expanded according to the semantic relationship acquisition; the text basic feature vector is mapped into a general high-dimensional feature vector in a high-dimensional vector space by the matched concept nodes and the expanded concept nodes, and a general-text high-dimensional feature vector is output.

[0080] Further, the feature mapping unit 13 in the double-level adaptive fusion system for scientific literature knowledge is further configured to: wherein the professional knowledge layer processing module comprises multi-domain knowledge samples; a multi-source text domain feature vector of the text basic feature vector is extracted, the multi-source text domain feature vector is processed by professional text conversion according to the multi-domain knowledge samples, a standardized professional text feature vector is obtained, and the standardized professional text feature vector is mapped into a professional high-dimensional feature vector in a high-dimensional vector space, and a professional text high-dimensional feature vector is output.

[0081] Further, the text fusion unit 14 in the two-level adaptive fusion system for scientific literature knowledge is further configured to train a multi-label field classifier, wherein the multi-label field classifier is trained by confidence training samples of a plurality of field labels; input the text basic feature vector into the multi-label field classifier to perform confidence calculation with the plurality of field labels respectively, and output a plurality of field confidences corresponding to the plurality of field labels; and perform normalization processing on the plurality of field confidences, and output the normalized field confidences as field confidence calculation results.

[0082] Further, the text fusion unit 14 in the two-level adaptive fusion system for scientific literature knowledge is further configured to train the multi-label field classifier by confidence training samples of a plurality of field labels, wherein the confidence training samples include corresponding field professional term density samples, field professional term proportion samples, citation mode samples, expression mode samples, and text structure samples under the plurality of field labels.

[0083] Further, the text fusion unit 14 in the two-level adaptive fusion system for scientific literature knowledge is further configured to construct a hierarchical attention network, and connect the hierarchical attention network with the multi-label field classifier, wherein the hierarchical attention network includes a plurality of layers of field features, and the plurality of layers of field features include word-level field features, sentence-level field features, and paragraph-level field features; perform field feature extraction on the text basic feature vector according to the hierarchical attention network, and output a field feature vector; and perform calculation on the field feature vector by the multi-label field classifier.

[0084] Further, the text fusion unit 14 in the two-level adaptive fusion system for scientific literature knowledge is further configured to obtain a field label of a first field confidence in the field confidence calculation results; initialize general knowledge layer processing weight and professional knowledge layer processing weight parameters according to the first field confidence and the field label; and obtain adaptive fusion weight parameters from the general knowledge layer processing weight and the professional knowledge layer processing weight.

[0085] Further, the text fusion unit 14 in the two-level adaptive fusion system for scientific literature knowledge is further configured to: construct a weight adjustment factor, the weight adjustment factor including a text complexity adjustment factor, a domain-specific term density adjustment factor, a context consistency adjustment factor, and a historical performance feedback adjustment factor; update the adaptive fusion weight parameter according to the weight adjustment factor, output an updated adaptive fusion weight parameter; and perform weight smoothing processing on the updated adaptive fusion weight parameter, and perform weighted fusion on the general-text high-dimensional feature vector and the professional-text high-dimensional feature vector with the smoothed adaptive fusion weight parameter.

[0086] Further, the text fusion unit 14 in the two-level adaptive fusion system for scientific literature knowledge is further configured to: construct a reward function, perform performance scoring on the smoothed adaptive fusion weight parameter according to the reward function, and output a real-time performance scoring result; and perform adaptive updating on the smoothed adaptive fusion weight parameter with the real-time performance scoring result as a feedback signal of adaptive reinforcement learning until a preset performance scoring threshold is met, and output an adaptive fusion weight parameter optimal solution.

[0087] The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The foregoing Figure 1 The two-level adaptive fusion method for scientific literature knowledge and the specific examples in Embodiment One are also applicable to the two-level adaptive fusion system for scientific literature knowledge in this embodiment. Based on the foregoing detailed description of the two-level adaptive fusion method for scientific literature knowledge, those skilled in the art can clearly understand the two-level adaptive fusion system for scientific literature knowledge in this embodiment. Therefore, for the sake of brevity of the specification, no further detailed description is given here.

[0088] The above description of the disclosed embodiments enables a person skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

[0089] Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the present application and its equivalents, the present application is intended to include these modifications and variations.

Claims

1. A two-level adaptive fusion method for scientific and technological literature knowledge, characterized in that, include: A two-level adaptive fusion architecture is constructed, which includes a general knowledge layer processing module, a professional knowledge layer processing module, an adaptive fusion module, and an application output module. The dual-level adaptive fusion architecture receives scientific and technological literature text and extracts the basic feature vector of the text. The general knowledge layer processing module performs general high-dimensional feature vector mapping on the basic text feature vector and outputs a general-text high-dimensional feature vector. The professional knowledge layer processing module performs professional high-dimensional feature vector mapping on the basic text feature vector and outputs a professional text high-dimensional feature vector. The adaptive fusion module is used to calculate the domain confidence of the text basic feature vector, configure adaptive fusion weight parameters according to the domain confidence calculation result, and use the adaptive fusion weight parameters to perform weighted fusion of the general-text high-dimensional feature vector and the professional text high-dimensional feature vector to output the text fusion feature vector. The application output module uses the fusion feature vector to process the scientific and technological documents.

2. The two-level adaptive fusion method for scientific and technological literature knowledge as described in claim 1, characterized in that, The general knowledge layer processing module performs a general high-dimensional feature vector mapping on the text basic feature vectors, and outputs a general-text high-dimensional feature vector. include: The general knowledge layer processing module includes a general knowledge graph, which includes general concept nodes and semantic relationship acquisition based on general knowledge samples. The general knowledge graph is used to perform general concept node matching on the basic feature vector of the text to obtain matching concept nodes, and extended concept nodes are obtained according to the semantic relationship. In the high-dimensional vector space, the matching concept node and the extended concept node perform a general high-dimensional feature vector mapping on the basic text feature vector, and output a general-text high-dimensional feature vector.

3. The two-level adaptive fusion method for scientific and technological literature knowledge as described in claim 1, characterized in that, The professional knowledge layer processing module performs professional high-dimensional feature vector mapping on the basic text feature vectors, and outputs professional text high-dimensional feature vectors. include: The professional knowledge layer processing module includes multi-domain knowledge samples; Extract the multi-source text domain feature vectors from the basic text feature vectors, and perform professional text conversion processing on the multi-domain knowledge samples to obtain standardized professional text feature vectors. In a high-dimensional vector space, the standardized professional text feature vectors are mapped to professional high-dimensional feature vectors, and the professional text high-dimensional feature vectors are output.

4. The two-level adaptive fusion method for scientific and technological literature knowledge as described in claim 1, characterized in that, The adaptive fusion module is used to calculate the domain confidence of the text's basic feature vector, including: A multi-label domain classifier is trained, wherein the multi-label domain classifier is obtained by training through confidence training samples of multiple domain labels; The text basic feature vector is input into the multi-label domain classifier and confidence is calculated for each of the multiple domain labels. The output is the multiple domain confidence based on the multiple domain labels. The confidence scores of the multiple domains are normalized, and the normalized domain confidence scores are output as the domain confidence score calculation results.

5. The two-level adaptive fusion method for scientific and technological literature knowledge as described in claim 4, characterized in that, The multi-label domain classifier is trained using confidence training samples of multiple domain labels. The confidence training samples include domain terminology density samples, domain terminology percentage samples, citation pattern samples, expression pattern samples, and text structure samples corresponding to the multiple domain labels.

6. The two-level adaptive fusion method for scientific and technological literature knowledge as described in claim 4, characterized in that, Before inputting the text basic feature vector into the multi-label domain classifier, the following steps are also included: A hierarchical attention network is constructed and connected to the multi-label domain classifier. The hierarchical attention network includes multi-layer domain features, which include word-level domain features, sentence-level domain features, and paragraph-level domain features. The hierarchical attention network extracts domain features from the text's basic feature vector, outputting a domain feature vector, which is then calculated by the multi-label domain classifier.

7. The two-level adaptive fusion method for scientific and technological literature knowledge as described in claim 1, characterized in that, Configure adaptive fusion weight parameters based on the domain confidence calculation results, including: Obtain the domain label of the first domain confidence score from the domain confidence score calculation results; Based on the confidence level of the first domain and the domain label, initialize the general knowledge layer processing weight and the professional knowledge layer processing weight parameters, and obtain the adaptive fusion weight parameters from the general knowledge layer processing weight and the professional knowledge layer processing weight.

8. The two-level adaptive fusion method for scientific and technological literature knowledge as described in claim 7, characterized in that, After initializing the processing weights for the general knowledge layer and the professional knowledge layer, the following steps are also included: Construct weight adjustment factors, which include text complexity adjustment factors, domain terminology density adjustment factors, context consistency adjustment factors, and historical performance feedback adjustment factors; The adaptive fusion weight parameters are updated according to the weight adjustment factor, and the updated adaptive fusion weight parameters are output. The updated adaptive fusion weight parameters are subjected to weight smoothing, and the smoothed adaptive fusion weight parameters are used to perform weighted fusion of the general-text high-dimensional feature vector and the professional text high-dimensional feature vector.

9. The two-level adaptive fusion method for scientific and technological literature knowledge as described in claim 8, characterized in that, The output of the smoothed adaptive fusion weight parameters also includes: Construct a reward function, score the smoothed adaptive fusion weight parameters according to the reward function, and output the real-time performance score result. The real-time performance score is used as a feedback signal for adaptive reinforcement learning to adaptively update the smoothed adaptive fusion weight parameters until a preset performance score threshold is met, and then the optimal solution for the adaptive fusion weight parameters is output.

10. A two-level adaptive fusion system for scientific and technological literature knowledge, characterized in that, The steps for implementing the two-level adaptive fusion method for scientific and technological literature knowledge according to any one of claims 1 to 9, wherein the two-level adaptive fusion system for scientific and technological literature knowledge comprises: An architectural organization unit is used to construct a two-level adaptive fusion architecture, which includes a general knowledge layer processing module, a professional knowledge layer processing module, an adaptive fusion module, and an application output module. A vector extraction unit is used in the dual-level adaptive fusion architecture to receive scientific and technological literature text and extract the basic feature vector of the text. The feature mapping unit is used by the general knowledge layer processing module to perform general high-dimensional feature vector mapping on the basic text feature vector and output a general-text high-dimensional feature vector. The professional knowledge layer processing module performs professional high-dimensional feature vector mapping on the basic text feature vector and outputs a professional text high-dimensional feature vector. The text fusion unit is used by the adaptive fusion module to calculate the domain confidence of the text basic feature vector, configure adaptive fusion weight parameters according to the domain confidence calculation result, and perform weighted fusion of the general-text high-dimensional feature vector and the professional text high-dimensional feature vector with the adaptive fusion weight parameters to output the text fusion feature vector. The application output module uses the fusion feature vector to perform the processing of the scientific and technological documents.

Images