Traditional Chinese medicine ancient book mapping knowledge domain construction method, device and medium

Abstract

The invention discloses a traditional Chinese medicine ancient book mapping knowledge domain construction method and device, and a medium. The method comprises the steps of generating structured text data; constructing an entity recognition model based on a pre-training language model in the field of traditional Chinese medicine, and performing entity recognition; constructing a single-book graph structure by taking entities as nodes and taking context semantics of the entities as initial edge weights; performing representation learning on nodes in combination with a graph neural network model, fusing context features, position information and graph structure features of entities, extracting semantic relationships among the entities, and constructing a single-book knowledge graph; the method comprises the following steps: fusing entities from a plurality of single-book knowledge maps based on a cross-map attention mechanism, constructing a cross-map attention matrix, dynamically adjusting a matching threshold value of the entities according to the interaction strength of the entities and the number of common neighbors, carrying out entity alignment and edge weight fusion, and generating a global traditional Chinese medicine ancient book knowledge map. According to the invention, automatic identification of entities and relationships in traditional Chinese medicine ancient books, knowledge graph construction and multi-book knowledge ablation are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing of ancient Chinese medical books, and in particular to a method, device, and medium for constructing a knowledge graph of ancient Chinese medical books. Background Art

[0002] With the development of traditional Chinese medicine (TCM) digitization and knowledge graph technology, the automated construction of knowledge graphs for ancient TCM texts has become a key direction for the intelligent utilization and cultural inheritance of ancient texts. Entity extraction and relationship extraction are the core components of knowledge graph construction, and their technical level directly affects the quality and application value of the graphs.

[0003] In terms of entity extraction technology, patent CN119761372A proposes a method for extracting entities from ancient Chinese books based on a deep active learning strategy. By integrating active learning strategies with deep neural network models, this method improves model fitting capabilities while utilizing improved query strategies to mitigate the problem of insufficient annotated data. The use of a dual-encoder structure further optimizes the ability to capture contextual semantics and enhances entity extraction performance. However, the model structure is complex, and the design of dual encoders and hybrid loss functions significantly increases computational costs and places high demands on hardware resources. Furthermore, active learning still relies on expert annotation, which reduces the amount of annotation but does not completely eliminate human intervention.

[0004] Patent CN116720512A discloses a method for extracting entities from ancient texts based on deep learning and artificial bee colonies. This method constructs a model consisting of an input layer, an embedding layer, a BiLSTM layer, and a CRF layer. Taking character sequences as input, the bidirectional structure of the BiLSTM captures contextual information, and the output of the annotation sequence from the CRF layer allows for accurate extraction of entities from ancient texts while reducing expert input. However, this method relies on pre-trained Word2vec word vectors, which are insufficient for representing rare characters unique to ancient texts or ancient Chinese vocabulary. The introduction of the artificial bee colony algorithm adds additional optimization steps, potentially extending training time. Furthermore, the method only addresses character-level annotations, not word-level semantic associations, and has limited processing capabilities for long-range dependencies.

[0005] Patent CN118966226A proposes a method and system for extracting entity relationships from ancient texts based on deep learning. By integrating multimodal information, a self-attention mechanism, a graph neural network, and dynamic knowledge graph technology, it achieves accurate identification and extraction of entities and their relationships in ancient texts, effectively processing the complex grammatical structure of classical Chinese and capturing long-distance dependencies. It also improves knowledge management efficiency by updating the knowledge graph in real time. However, this method's multimodal processing requires simultaneous parsing of text and images, resulting in high data preprocessing complexity. For example, image feature extraction using ResNet50 takes a long time. The real-time update mechanism of the dynamic knowledge graph may face data consistency challenges. The combination of the self-attention mechanism and the graph neural network does not specify a parameter sharing strategy, which may lead to model redundancy. Furthermore, the ability to extract implicit relationships in ancient texts is not explicitly mentioned.

[0006] In summary, although there are many optimization methods for entity extraction and relationship extraction, there are still problems such as high computational cost, reliance on manual annotation, insufficient representation of rare characters, limited long-distance dependency processing, and complex multimodal preprocessing. There is an urgent need for a more efficient, less annotation-dependent, and automatic knowledge graph construction method that can comprehensively capture the semantic and structural associations of ancient books. Summary of the Invention

[0007] Based on the above deficiencies in the existing technology, the present invention provides a method, device, and medium for constructing a knowledge graph of ancient Chinese medical books, which can realize the automatic recognition of entities and relationships in ancient Chinese medical books, the construction of a knowledge graph, and the ablation of multi-book knowledge.

[0008] To solve the above technical problems, the first aspect of the present invention discloses a method for constructing a knowledge graph of ancient Chinese medical books, the method comprising: Acquire ancient Chinese medical book data and perform standardization on the ancient Chinese medical book data; Combined with the domain dictionary, word segmentation priority rules are constructed, the minimum matching length and dictionary priority are set, and a bidirectional maximum matching algorithm is used to segment the ancient Chinese medical book data to generate structured text data; An entity recognition model is constructed based on a pre-trained language model in the field of Traditional Chinese Medicine. The entity recognition model is weakly supervised through pseudo-label expansion and introduces domain adversarial training to achieve cross-language semantic transfer, outputting entities and corresponding entity annotation results. A single-book graph structure is constructed using the entities as nodes and the contextual semantics of the entities as initial edge weights. A graph neural network model is used to learn the representations of the nodes, integrating the contextual features, location information, and graph structural features of the entities. The attention mechanism is used to extract semantic relationships between entities and construct a single-book knowledge graph. Based on the cross-graph attention mechanism, entities from multiple single-book knowledge graphs are fused, and a cross-graph attention matrix with weights equal to the product of semantic similarity, structural similarity and path similarity is constructed. The matching threshold of entities is dynamically adjusted according to the entity interaction intensity and the number of common neighbors, and entity alignment and edge weight fusion are performed to generate a global knowledge graph of ancient Chinese medicine books.

[0009] In some embodiments, weakly supervised training is performed by pseudo-label expansion, including: Extract entity classification pairs from the domain dictionary, where the entity classification pairs are initial positive samples; Extract entity feature vectors through the pre-trained language model, calculate the cosine similarity between entity feature vectors, determine the extended entity based on the cosine similarity, and obtain an extended pseudo-label training set; the extended pseudo-label training set includes automatically annotated entities, entity segments, and relationship triples; Calculate the semantic similarity between entities and text fragments, and select text fragments with confidence levels higher than a preset confidence threshold; When at least two entities exist in multiple text segments, there is a potential relationship between the two entities, and a weakly supervised training sample is obtained; the weakly supervised training sample includes a text segment, an entity label sequence, and a relationship label.

[0010] In some implementations, domain adversarial training is introduced to achieve cross-language semantic transfer, including: Setting a gradient reversal layer in the BERT layer and the task layer of the pre-trained language model to perform gradient reversal on the ancient book feature vector and the modern Chinese medicine feature vector output by the BERT layer; The gradient-inverted ancient book feature vector and modern Chinese medicine feature vector are input into the domain discriminator to identify the source of the corpus; Based on the entity recognition cross entropy loss function and the domain adversarial loss function, a weighted combination task loss function is constructed to optimize the entity recognition model.

[0011] In some embodiments, the domain adversarial loss function is:

[0012] in, is the domain adversarial loss function; is the feature output by the BERT layer, For the distribution of ancient book data; Represents samples from the distribution of ancient book data x Find the expectation; G is the domain discriminator; The task loss function is:

[0013] in, is the domain adversarial loss function, is the cross entropy loss function for entity recognition, is the balance parameter.

[0014] In some embodiments, a single-book graph structure is constructed using the entity as a node and the contextual semantics of the entity as the initial edge weight. The nodes are represented by a graph neural network model, and the contextual features, location information, and graph structure features of the entity are integrated. The semantic relationship between entities is extracted using an attention mechanism to construct a single-book knowledge graph, including: It traverses all entities and extracts entity pairs and their context fragments based on their co-occurrence relationships in ancient texts to form an entity pair set. Encode the text of the context segment into a word vector, use a multi-head self-attention mechanism to extract the context features of the entity pair, and generate a fused semantic feature vector of the entity pair; Constructing a graph structure based on the co-occurrence relationship of entities, and using the fused semantic feature vector as the initial feature of the node, performing multi-layer representation learning on the node through a graph convolutional network to obtain node features containing global semantic information; Aggregate and classify the entity pair feature vectors output by the graph neural network, determine the relationship type between entities, and generate entity relationship triples; the entity relationship triples contain the source entity, target entity, relationship category, context fragment, confidence level, and source book; Using entities as nodes and triple relationships as edges, combined with attribute information, a dynamically updated single-book knowledge graph is constructed.

[0015] In some embodiments, entities from multiple single-book knowledge graphs are fused based on a cross-graph attention mechanism, a cross-graph attention matrix is constructed whose weights are the product of semantic similarity, structural similarity, and path similarity, and the matching threshold of entities is dynamically adjusted according to the entity interaction strength and the number of common neighbors, including: Construct a cross-graph attention matrix, where the weight of the cross-graph attention matrix is determined by the product of semantic similarity, structural similarity, and path similarity between entity pairs; Calculating an entity interaction vector based on the cross-graph attention matrix; The matching threshold is dynamically adjusted according to the interaction strength between entities, the number of common neighbors, and the path similarity.

[0016] In some embodiments, further comprising: Receive newly added single book graphs containing new entity and relationship information, as well as existing global knowledge graph data; Calculate the semantic similarity, structural similarity, and path similarity between each entity in the newly added single-book graph and the entities in the global graph, and prioritize entity alignment based on matches with high structural similarity; For entities that have completed alignment, the newly added relationship paths are automatically mapped to the corresponding nodes in the global graph, their neighbor structure and relationship path features are updated, the dynamic inheritance of entity relationships is realized, and the updated global knowledge graph is output.

[0017] In some embodiments, the entity interaction vector is calculated using the following formula:

[0018] in, For Entity The interaction vector, For entities in other graphs, For Entity and Cross-graph attention weights between ; The matching threshold is adjusted by the following formula:

[0019] in is the initial threshold, is the adjustment coefficient of entity interaction strength, is an entity pair; It is a comprehensive indicator of the number of common neighbors across graphs and path similarity.

[0020] In a second aspect, a computer device is disclosed, characterized in that it includes: a processor and a memory; wherein the memory stores a computer program, and the computer program is suitable for being loaded by the processor and executing the steps of any one of the above methods for constructing a knowledge graph of ancient Chinese medicine books.

[0021] In a third aspect, a computer storage medium is disclosed, on which a computer program is stored. When the computer program is executed by a processor, the method for constructing a knowledge graph of ancient Chinese medicine books as described in any one of the above items is implemented.

[0022] Compared with the prior art, the present invention has the following beneficial effects: The method, equipment and medium for constructing the knowledge graph of ancient Chinese medical books provided in this application can effectively improve the structural processing capabilities of ancient Chinese medical texts without manual annotation. By introducing domain dictionaries and two-way maximum matching algorithms for word segmentation, combined with domain pre-trained language models and weakly supervised training strategies, automatic recognition of key entities in ancient texts is achieved. At the same time, graph neural networks and attention mechanisms are used to fuse entity context and structural information, accurately extract semantic relationships between entities, and construct a single-book knowledge graph with semantic depth. Through the cross-graph attention mechanism, entity alignment and fusion between multiple ancient book graphs are achieved, and a unified global Chinese medicine knowledge graph is automatically constructed, which improves the consistency of knowledge organization and provides a solid foundation for subsequent intelligent retrieval and reasoning. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 A schematic diagram of the process of constructing a knowledge graph of ancient Chinese medical books provided by the present invention; Figure 2 This is a flow chart of step S31 of a method for constructing a knowledge graph of ancient Chinese medical books provided by the present invention; Figure 3 This is a flow chart of step S32 of a method for constructing a knowledge graph of ancient Chinese medical books provided by the present invention; Figure 4 This is a flow chart of step S4 of the method for constructing a knowledge graph of ancient Chinese medical books provided by the present invention; Figure 5 This is a flow chart of step S5 of the method for constructing a knowledge graph of ancient Chinese medical books provided by the present invention; Figure 6 This is a flow chart of step S6 of the method for constructing a knowledge graph of ancient Chinese medical books provided by the present invention. DETAILED DESCRIPTION

[0024] For better understanding and implementation, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0025] The terms "including" and "having" and any variations thereof in the embodiments of the present invention are intended to cover non-exclusive inclusions. For example, a process, method, system, product or apparatus that includes a series of steps or modules is not necessarily limited to those steps or modules explicitly listed, but may include other steps or modules not explicitly listed or inherent to these processes, methods, products or apparatus.

[0026] The embodiments of the present invention disclose a method for constructing a knowledge graph of ancient Chinese medical books. In view of the fact that ancient Chinese medical texts have no annotated data, combined with the corpus in the field of traditional Chinese medicine, automatic recognition of entities and relationships in ancient Chinese medical books, knowledge graph construction and multi-book knowledge ablation are achieved through transfer learning, weakly supervised entity extraction and dynamic graph fusion technology.

[0027] like Figure 1 As shown, the method includes: Step S1: Acquire ancient Chinese medical book data and perform normalization processing on the ancient Chinese medical book data.

[0028] To achieve a structured understanding of ancient TCM texts and subsequently construct a knowledge graph, we first acquired data on ancient TCM texts, including 700 historical TCM texts and a TCM knowledge base. These texts, provided in GBK or UTF-8 encoding, were sourced from classic texts such as the Yellow Emperor's Classic of Internal Medicine, Compendium of Materia Medica, Treatise on Cold Damage, and Qianjin Fang. The TCM knowledge base contains over 50,000 entities across eight entity categories: Chinese medicinal materials, prescriptions, symptoms, acupoints, viscera, symptoms, preparation methods, and diagnostic terms.

[0029] In the data normalization stage, the data of ancient Chinese medical books are cleaned. Regular expressions are used to filter punctuation marks, tabs, spaces and other non-language characters in the text. The text is structured through natural paragraph segmentation logic, empty paragraphs and redundant line breaks are removed, and the text format is unified to construct a standardized text set, which is recorded as ,in,( ) indicates that the i-th processed ancient book text is used for subsequent semantic enhancement processing.

[0030] Extracting entity sets from traditional Chinese medicine corpus ,in Indicates the c Class j Entities, that is, to establish a mapping relationship between each entity and its corresponding classification, such as "Atractylodes macrocephala - medicinal material", "Honey roasting - processing method", to form an entity-classification domain dictionary ,in C Classify a collection of entities.

[0031] Step S2: construct word segmentation priority rules based on the domain dictionary, set the minimum matching length and dictionary priority, and use a bidirectional maximum matching algorithm to segment the ancient Chinese medical book data to generate structured text data; To improve the performance of text in subsequent entity recognition tasks, this embodiment uses a bidirectional maximum matching (Bi-MMF) segmentation algorithm. This method combines forward maximum matching (FMM) and backward maximum matching (BMM) strategies, ensuring the accuracy of segmentation results through conflict resolution rules such as dictionary priority, word length priority, and single-word filtering. During the segmentation process, a minimum matching length of 2 characters is set to prevent non-independent characters such as "yellow" and "white" from being mistakenly identified as entities. If there are conflicts in the forward and reverse segmentation results, for example, "Atractylodes macrocephala" can be used as a medicinal material with a priority of p=1, but can also be considered a preparation method with a priority of p=3, the segmentation result with the higher priority will be selected based on the preset entity priority. Therefore, "Atractylodes macrocephala" is considered a preparation method. In other cases, such as when "white tuckahoe" conflicts with "white + tuckahoe", the longer word "white tuckahoe" is preferred as the segmentation result, preserving its complete semantic structure.

[0032] After word segmentation, each ancient book text is converted into structured text data. This structured text data includes a cleaned and segmented ancient book text dataset and a domain dictionary dataset. The ancient book text dataset includes the Bi-MMF word segmentation results, its corresponding part of speech, and candidate entity categories. The output dataset at this time contains two parts: one is the cleaned and segmented text set corresponding to each ancient book, in which each word has a semantic boundary to facilitate subsequent model recognition processing; the other is a standardized and constructed dictionary in the field of traditional Chinese medicine, which contains entity text, entity classification, and priority labels, providing a semantic benchmark for subsequent weakly supervised training and cross-book fusion.

[0033] This embodiment integrates domain knowledge and rule-based algorithms to complete the standardization, entity explicitness and semantic enhancement of the original text of ancient books without relying on manual annotation, significantly improving the model's comprehensibility and structural transferability of the traditional Chinese medicine language system in a low-resource environment.

[0034] Step S3: constructing an entity recognition model based on a pre-trained language model in the field of traditional Chinese medicine. The entity recognition model is weakly supervised trained by pseudo-label expansion, and domain adversarial training is introduced to achieve cross-language semantic transfer, and the entity and the corresponding entity labeling results are output; In order to realize the automatic recognition of entities in ancient Chinese medical books, this application constructs an entity recognition model based on the pre-trained language model in the field of traditional Chinese medicine. In view of the problems of complex language structure and lack of annotated data in ancient Chinese medical books, this application integrates technical strategies such as semantic representation learning, pseudo-label expansion and domain adversarial training, and effectively realizes the ability to automatically extract core Chinese medical entities such as "medicinal materials", "diseases" and "prescriptions" from unannotated ancient book texts.

[0035] The pre-trained language model uses the BERT-BiLSTM-CRF structure as the basis for the entity recognition model. The BERT layer is initialized with a language model pre-trained on TCM corpus to enhance the encoding capability of mixed texts of ancient Chinese and TCM terms. t Input the BERT layer to generate the contextual semantic representation of each Chinese character. That is,

[0036] in, It is the word-level feature vector sequence output by the BERT model, with the dimension of sequence length × hidden layer dimension; each element Indicates the first t The context-fused semantic feature vector of each word contains the contextual information.

[0037] The word-level feature vector sequence output by the BERT model is input to the BiLSTM layer:

[0038] The contextual dependencies of sentences are captured in both the forward and backward directions, and then concatenated to form a representation that integrates global semantics. ( is the hidden state of the forward LSTM at the tth time step, capturing the left-to-right context information; is the hidden state of the previous time step. middle is the hidden state of the backward LSTM at the tth time step, capturing the right-to-left context information; is the hidden state of the next time step. The splicing feature is the feature sequence that integrates the previous and next context information. , the dimension is sequence length × (hidden layer dimension × 2), and the bidirectional context information is integrated to alleviate the entity boundary ambiguity caused by the omission of subject and object in ancient books.

[0039] Feature Sequence It is fed into the CRF layer, which uses the conditional random field to model the possible entity tag sequence, and then uses the Viterbi algorithm:

[0040] Output the optimal label sequence to predict the label of each word, such as B-medicine, I-symptom, O (non-entity), etc. is the predicted entity label sequence, such as [B-herbal medicine, I-herbal medicine, O, B-disease,...], is the label of the t-th word, such as B-herbal medicine indicates the start of the entity, is the tag of the previous character, used to model the tag transition probability. For example, after B - medicinal material, only I - medicine or O can follow. is the tag transition probability calculated by the CRF model. The optimal path is solved through the Viterbi algorithm to obtain the entity tag sequence. Mark whether each character belongs to an entity and the entity annotation result.

[0041] Furthermore, due to the lack of sufficient manually annotated data in traditional Chinese medicine ancient books, a weak - supervision pseudo - label expansion mechanism is adopted in the model training stage. Weak - supervision training is carried out through the pseudo - label expansion method. As Figure 2 shown, it includes: Step S311: Extract entity - classification pairs from the domain dictionary. The entity - classification pairs are initial positive samples. Using the seed entities extracted from the traditional Chinese medicine domain dictionary, such as "Rehmannia glutinosa - medicinal material" marked as B - medicinal material, "cough - disease", initialize the pseudo - label set. These entities are regarded as initial positive samples and are directly used to label the corresponding segments appearing in the training text.

[0042] Step S312: Extract entity feature vectors through the pre - trained language model, calculate the cosine similarity between entity feature vectors, and determine the extended entities based on the cosine similarity to obtain an extended pseudo - label training set; the extended pseudo - label training set includes automatically annotated entities, the segments where the entities are located, and relation triples.

[0043] To expand the training data coverage, in this embodiment, a semantic similarity strategy is introduced. By calculating the cosine similarity between the seed entity vector and the unannotated segment, entity segments with similar semantics are automatically identified. The cosine similarity is calculated by the following formula:

[0044] If the cosine similarity is not less than the cosine similarity threshold , , then is marked as an entity of the same type as e. In some embodiments, the cosine similarity threshold can be set to 0.7 and can be adjusted according to the actual situation. For example, "raw rehmannia" and "Rehmannia glutinosa" are semantically similar. After meeting the similarity threshold, "raw rehmannia" is also marked as a medicinal - material - type entity. Thus, pseudo - labels are automatically generated. This extended training set not only includes automatically annotated entities, but also the context segments where they are located and the potential relation triples inferred from co - occurrence information.

[0045] Step S313: Calculate the semantic similarity between the entity and the text segment, and select the text segments with a confidence level higher than the pre - set confidence threshold; In addition, the semantic similarity between the seed entity and the context fragment is calculated through the Sentence-BERT model, the entity combinations implied in the context are identified, and the text fragments with a confidence level higher than the preset confidence threshold are selected as additional training samples. For example, if the similarity between the context containing "raw rehmannia" and "yellow" and "radix rehmanniae" is greater than 0.8, then this fragment is adopted for training data expansion and labeled as B-herb and I-herb.

[0046] Step S314: When there are at least two of the said entities in multiple text fragments, there is a potential relationship between the two said entities, and weak supervision training samples are obtained; the weak supervision training samples include text fragments, entity label sequences, and relationship labels.

[0047] When it is detected that there are at least two of the said entities in multiple text fragments, there is a potential relationship between the two said entities. For example, when "radix rehmanniae" and "blood deficiency syndrome" frequently co-occur, a "treatment" relationship candidate can be generated. Finally, weak supervision training samples are constructed, and the weak supervision training samples include text fragments, entity label sequences, and relationship labels for model fine-tuning training.

[0048] Furthermore, to address the differences in the语体 structure between the ancient Chinese language style and modern Chinese medicine corpora, domain adversarial training is introduced to achieve cross语体 semantic transfer and improve the cross语体 generalization ability of the model. As Figure 3 shown, specifically including: Step S321: Set a gradient reversal layer in the BERT layer and the task layer of the pre-trained language model, and perform gradient reversal on the ancient Chinese feature vector and the modern Chinese medicine feature vector output by the BERT layer.

[0049] During the training process, the entity recognition model simultaneously receives corpus samples from ancient Chinese and modern Chinese medicine literature and extracts their character-level features respectively. Set a gradient reversal layer (GRL) in the BERT layer and the task layer of the pre-trained language model, and perform gradient reversal on the ancient Chinese feature vector and the modern Chinese medicine feature vector output by the BERT layer , making the model unable to distinguish the two corpus sources. For example, the association pattern of "herb-disease".

[0050] Step S322: Input the ancient Chinese feature vector and the modern Chinese medicine feature vector after gradient reversal into a domain discriminator for discriminating the corpus source; Input the ancient Chinese feature vector and the modern Chinese medicine feature vector after gradient reversal into a domain discriminator for discriminating the corpus source. The domain discriminator attempts to distinguish the sources, but due to the feature reversal, the training objective makes the entity recognition model confuse the two, so as to learn a more general entity semantic representation.

[0051] Step S323: Based on the entity recognition cross-entropy loss function and the domain adversarial loss function, construct a weighted combined task loss function to optimize the entity recognition model.

[0052] The domain adversarial loss function is:

[0053] where is the feature output by the BERT layer, is the ancient book data distribution; denotes the expectation of samples from the ancient book data distribution x ; G is the domain discriminator, which forces the model to confuse domain differences by maximizing the domain discrimination error.

[0054] The task loss function is , is the entity recognition cross-entropy loss function, which optimizes label prediction. is the domain adversarial loss function, which optimizes the cross-domain generalization ability. is the balance parameter, usually 0.1.

[0055] Based on the entity recognition cross-entropy loss function and the domain adversarial loss function, construct a weighted combined task loss function to optimize the entity recognition model, which can improve the recognition accuracy while ensuring that the model has strong cross-domain transfer ability.

[0056] Step S4: Using the entity as a node and the context semantics of the entity as the initial edge weight, construct a single-book graph structure; combine the graph neural network model to perform representation learning on the nodes, fuse the context features, position information and graph structure features of the entity, and use the attention mechanism to extract the semantic relationship between entities to construct a single-book knowledge graph.

[0057] After entity recognition, further automatically extract the semantic relationship between entities from the recognized entities and construct a single-book knowledge graph containing entity nodes and relationship edges, as Figure 4 shown, specifically including: Step S41: Traverse all entities, and based on the co-occurrence relationship of entities in the ancient book text, extract entity pairs and their context fragments to form an entity pair set; Traverse the entities marked in the ancient book text, and generate an entity pair set by counting the co-occurrence relationship between entities in the local context. For example, in "Compendium of Materia Medica", if "Rehmannia glutinosa" and "blood deficiency syndrome" appear in the same text segment, the system records this entity pair (e 1 ,e 2 ) , and extract its context fragment from the original text, with a default window of 50 words before and after, to form a context sample, for example [(e1 ,e 2 ,context),...] .

[0058] Step S42: Encode the text of the context segment into a word vector, use a multi-head self-attention mechanism to extract the context features of the entity pair, and generate a fused semantic feature vector of the entity pair; To capture the semantic associations between entity pairs in context, a multi-head self-attention mechanism is introduced after word vector embedding. By constructing three sets of vectors: query (Q), key (K), and value (V), attention scores are calculated in parallel on multiple sets of attention heads to model the contextual semantic dependencies between entity pairs. The formula for calculating the attention score is:

[0059] is a learnable parameter, Used for scaling.

[0060] The outputs of all attention heads are concatenated and linearly mapped to obtain the fused semantic feature vector Z of the generated entity pair, which represents the strength and direction of the semantic relationship between the entity pair in its text context.

[0061] Step S43: constructing a graph structure based on the co-occurrence relationship of the entities, and using the fused semantic feature vector as the initial feature of the node, performing multi-layer representation learning on the node through a graph convolutional network to obtain node features containing global semantic information; Based on the above entity pairs and their fused semantic feature vectors Z, a graph structure is constructed to infer global relationships. All identified entities are used as nodes of the graph, and the co-occurrence relationships between different entities are constructed as edges. The initial edge weights are derived from their attention semantic features. The entity pair fused semantic feature vector Z is used as the initial node feature of the graph neural network (GCN), and the adjacency matrix A is constructed in combination with the entity co-occurrence frequency. Self-loops are added to obtain , forming a structural connection diagram between entities.

[0062] In graph neural networks, the system normalizes the adjacency matrix:

[0063] in, is the degree matrix.

[0064] Perform graph convolution operations layer by layer:

[0065] For the l Layer features, are learnable weights, is the activation function.

[0066] Gradually aggregate the features of neighbor nodes to the target node, and finally form node features containing global graph structure information .

[0067] Step S44: Aggregate and classify the entity pair feature vectors output by the graph neural network, determine the relationship type between entities, and generate entity relationship triples; the entity relationship triples include the source entity, target entity, relationship category, context fragment, confidence level, and source book; Using the entity pair features output by the GCN as input, the relationships between the entity pairs are aggregated and classified. By concatenating or weighted pooling the final representation vectors of the two nodes, the vectors are fed into a fully connected layer and then fed into a softmax classifier to predict the specific semantic relationship type between the entity pairs, such as "treatment," "composition," or "main indications." Ultimately, the results are output as triples, each containing: source entity, target entity, relationship type, confidence score, context snippet, and the source text (e.g., "shengdihuang," "blood deficiency syndrome," "treatment," 0.92, context, Compendium of Materia Medica").

[0068] Step S45: Using entities as nodes and triple relationships as edges, combined with attribute information, a dynamically updated single-book knowledge graph is constructed.

[0069] Entities are nodes, and the semantic relationships between entities are edges. The graph structure has the characteristics of clear semantics, rich attributes, and clear structure. At the node level, the system supports unified modeling of multiple types of entity nodes. Node types include but are not limited to TCM semantic categories such as "medicinal materials," "diseases," "prescriptions," "processing methods," and "symptoms." Each node not only contains basic attributes such as the entity's text content and the category it belongs to, but also includes statistical information such as a list of source books and the frequency of occurrence, which is used to analyze the breadth and importance of the entity in ancient books.

[0070] To support subsequent cross-book fusion and similarity calculation, each node stores feature vectors obtained from graph neural network reasoning. These vectors can be regarded as semantic embeddings of entities and are the core basis for structural alignment and graph aggregation. The storage format of the node is:

[0071] In the construction of edges, based on the relationship triples, each group (entity A, entity B, relationship R) is mapped to a directed edge in the graph. The edge type is a predefined relationship category, such as "treatment", "composition", "contraindications", "main indications", etc., which comes from the model prediction results or remote supervision assumptions. The attributes of each edge include the relationship type , co-occurrence count , semantic confidence and its weight The edge weights are calculated using a weighting function that balances statistical frequency with model confidence:

[0072] middle, To balance statistical frequency and model confidence, ; is the number of co-occurrences, Is the relationship type.

[0073] The edge weights are calculated so that edges that appear frequently and are predicted to be reliable by the model have higher graph weights, thus having greater structural influence in subsequent graph operations. In this embodiment, the edge storage format is:

[0074] The overall data structure of the graph is organized and stored in a standard graph format. Nodes, edges, and their attribute information are generally stored in a graph database such as Neo4j to facilitate graph query and visualization. Alternatively, it can be exported to a structured file format such as GraphML or JSON to facilitate integration with external systems or offline analysis and processing.

[0075] Step S5: Based on the cross-graph attention mechanism, entities from multiple single-book knowledge graphs are integrated to construct a cross-graph attention matrix whose weights are the product of semantic similarity, structural similarity and path similarity. The matching threshold of the entity is dynamically adjusted according to the entity interaction intensity and the number of common neighbors, and entity alignment and edge weight fusion are performed to generate a global knowledge graph of ancient Chinese medicine books.

[0076] To address the fragmentation of cross-book knowledge in the knowledge graph of ancient Chinese medicine, where one medicine has multiple names or the same disease has different treatments, we perform structural fusion based on multiple single-book graphs. By building a cross-graph attention mechanism, we achieve entity alignment and relationship weight fusion, and ultimately generate a unified global knowledge graph of ancient Chinese medicine. Specifically: To achieve semantic alignment of entities across graphs, a semantic embedding model for domain entities was constructed using Traditional Chinese Medicine corpus. This process is based on the Sentence-BERT architecture and introduces a triplet loss function for training. The triplet loss function is:

[0077] Where d is the cosine distance, They are anchor, positive and negative entities respectively.

[0078] Training samples are derived from the domain dictionary. We extract similar entities, such as "ginseng-white ginseng," as positive pairs, and randomly combine entities of different categories, such as "ginseng-cough," as negative pairs. This generates a set of training samples with the structure (anchor, positive, negative). In the semantic embedding space, we narrow the distance between synonymous entities while distinguishing entities of different categories as much as possible. This yields a trained Sentence-BERT model and an entity fusion semantic feature vector mapping table.

[0079] Entities in multiple single-book graphs are automatically aligned. The first-order neighbor information and path pattern of the entity are extracted from each graph, such as "Ginseng → Treatment → Blood Deficiency" or "White Ginseng → Main Indications → Qi Deficiency", to construct a graph structure feature representation. Then, the system calculates the similarity of entity pairs in three dimensions: the first is semantic similarity, measured by the cosine distance in the Sentence-BERT embedding space; the second is structural similarity, represented by the overlap and connection density of entity neighbor nodes; the third is path similarity, obtained by comparing the relationship path patterns of the two entities and calculating the maximum structural matching score.

[0080] To address the graph interaction flaws of the minimum distance matching method, we achieve precise alignment through multi-dimensional graph structure fusion, automatically aligning entities in multiple single-book graphs. First, we extract the first-order neighbor information and path patterns of entities from each single-book knowledge graph, such as "ginseng → treatment → blood deficiency" or "white ginseng → main indications → qi deficiency", to construct a graph structure feature representation.

[0081] Calculate the semantic similarity, structural similarity, and path similarity of entity pairs respectively. Semantic similarity is measured by the cosine distance in the Sentence-BERT embedding space, specifically:

[0082] Structural similarity is represented by the overlap and connection density of entity neighbor nodes:

[0083] Path similarity is obtained by comparing the relationship path patterns of two entities and calculating the maximum structural matching score:

[0084] like Figure 5 As shown, step S5 further includes: Step S51: construct a cross-graph attention matrix, where the weight of the cross-graph attention matrix is determined by the product of semantic similarity, structural similarity and path similarity between entity pairs.

[0085] The weight is the product of the three-dimensional similarity mentioned above:

[0086] is the semantic similarity, is the structural similarity, is the path similarity. The cross-graph attention matrix represents the strength of structural semantic fusion between each pair of entities from different graphs.

[0087] Step S52: Calculate the entity interaction vector based on the cross-graph attention matrix.

[0088] Based on the cross-graph attention matrix, the interaction vector of each target entity is calculated, that is, the semantic features of neighboring entities fused from other graphs, to achieve cross-graph information aggregation and enhance the global consistency of entity representation.

[0089] The entity interaction vector is: , For Entity The interaction vector, For entities in other graphs, For Entity and The cross-graph attention weights between them.

[0090] Step S53: Dynamically adjust the matching threshold according to the interaction intensity between entities, the number of common neighbors, and the path similarity.

[0091] To ensure the accuracy and adaptability of entity matching, a dynamic interaction threshold adjustment mechanism is introduced. The matching threshold for whether to merge entities is dynamically updated based on the comprehensive score of the entity interaction strength, number of common neighbors, and path structure overlap of each pair of entities in the attention matrix.

[0092] The matching threshold is adjusted by the following formula:

[0093] in is the initial threshold, is an entity pair; It is a comprehensive indicator of the number of common neighbors across graphs and path similarity.

[0094] For example, when two entities have high scores across three dimensions and have similar adjacency structures in multiple graphs, such as "Ginseng → Treatment → Blood Deficiency" and "White Ginseng → Indications → Qi Deficiency," the matching threshold can be appropriately lowered to merge them into the same entity node. Conversely, if the structures differ significantly, even if they have similar semantics, they may be retained as separate entities to prevent incorrect merging.

[0095] After entity alignment is complete, automated graph fusion is performed based on the alignment mapping table M and the original graph data. This includes node attribute merging and edge weight fusion. Node attribute merging includes retaining multiple names as aliases, merging source books, and accumulating occurrence frequencies. Edge weight fusion involves proportionally weighting edge weights in a single book graph with cross-graph attention weights to obtain unified edge attributes. For example, the path "Treatment - Blood Deficiency" is retained.

[0096] Edge weight calculation is to combine the single-book edge weight and the cross-graph interaction weight. For example,

[0097] is the single-book edge weight, is the cross-graph attention weight.

[0098] Ultimately, the generated global knowledge graph nodes contain cross-graph structural features, and the edges contain interaction weights, which retains the semantic structure between entities in the original graph while unifying name expressions, semantic boundaries, and cross-document semantic mapping relationships.

[0099] What's more, when new ancient books are added, entities can be automatically extracted and automatically identified, compared and integrated to ensure the sustainable evolution of the knowledge graph structure and maintain consistency in cross-graph interactions.

[0100] like Figure 6 As shown, specifically including: Step S61: receiving a newly added single book graph containing new entity and relationship information, and existing global knowledge graph data; Step S62: Calculate the semantic similarity, structural similarity, and path similarity between each entity in the newly added single-book graph and the entities in the global graph, and prioritize entity alignment based on matching items with high structural similarity; Step S63: For entities that have completed alignment, automatically map the newly added relationship paths to the corresponding nodes in the global graph, update their neighbor structure and relationship path features, realize the dynamic inheritance of entity relationships, and output the updated global knowledge graph.

[0101] For example, when the newly added ancient book "Qianjin Fang" is input, the system first cleans, segments, recognizes entities and extracts relationships through the aforementioned steps S1 to S4 to construct a single-book knowledge graph corresponding to the ancient book, which contains entity nodes such as "Silla ginseng" and "Korea" and the semantic relationships between them, such as "place of origin".

[0102] To correctly integrate this single-book graph into the global graph, entity alignment was performed on each entity within it. Taking "Silla ginseng" as an example, its semantic embedding vector was first extracted, and semantic similarity was calculated using the embedding vectors of all medicinal herb entities in the global graph. Structural features were also extracted, such as the "origin" relationship path between this entity and "North Korea." This was then structurally compared with the relationship path of the "ginseng" entity in the global graph, revealing that both entities share the "→origin→North Korea" pattern.

[0103] At this time, the system calculates the matching score based on the semantic similarity, structural similarity and path similarity. If the score exceeds the set threshold, it is determined that "Silla ginseng" and "ginseng" are the same entity. Based on this, "Silla ginseng" is used as one of the aliases of "ginseng", its node attributes are merged, and "Qianjin Fang" is retained as a newly added source book, and the frequency of entity occurrence is accumulated. In this embodiment, nodes with high structural similarity are matched first to improve the accuracy and robustness of entity alignment and avoid the problem of false matching that may be caused by relying solely on semantic similarity. By comparing the graph topology features such as neighbor relationships, connection patterns and path structures of entities in the atlas, it is possible to more accurately determine whether two entities have the same semantic role in different atlases. For example, if two medicinal materials are connected to similar diseases or prescriptions in their respective ancient books through relationships such as "treatment" and "composition", they may be the same entity even if their names are different. Giving priority to structural similarity helps to maintain the consistency of the overall semantic structure of the atlas, thereby improving the reliability and expression quality of cross-book knowledge fusion.

[0104] When updating node attributes, the mapping path is also updated. For example, the relationship path "Silla ginseng → origin → North Korea" is mapped to "Ginseng → origin → North Korea". The edge weight of this edge is updated according to the fusion strategy, for example, the weight is set to the weighted average of the original edge weight and the attention score of the newly added path.

[0105] If the new entity does not have a similar node in the global graph, such as the name of a newly added acupuncture point or the name of a rare medicinal material, the entity will be automatically inserted into the graph as a new node, its attributes will be initialized, and all its original edges will be connected.

[0106] After the update is complete, an incremental alignment report is output, recording the handling of each new entity in this round of import, including: whether to merge, the merge target, the merge method, the number of newly added relationship paths, and the edge weight adjustment records. This report is available for user review and facilitates subsequent model feedback and optimization.

[0107] The method for constructing a knowledge graph of ancient Chinese medical books provided in this application can effectively improve the structural processing capabilities of ancient Chinese medical texts without manual annotation. By introducing a domain dictionary and a two-way maximum matching algorithm for word segmentation, combined with a domain pre-trained language model and a weakly supervised training strategy, automatic recognition of key entities in ancient texts is achieved. At the same time, graph neural networks and attention mechanisms are used to fuse entity context and structural information, accurately extract semantic relationships between entities, and construct a single-book knowledge graph with semantic depth. Through the cross-graph attention mechanism, entity alignment and fusion between multiple ancient book graphs are achieved, and a unified global Chinese medicine knowledge graph is automatically constructed, which improves the consistency of knowledge organization and provides a solid foundation for subsequent intelligent retrieval and reasoning.

[0108] Based on the same inventive concept, the present invention also provides a computer device, comprising: a processor and a memory; wherein the memory stores a computer program, and the computer program is suitable for being loaded by the processor and executing the steps of the above-mentioned method for constructing a knowledge graph of ancient Chinese medicine books.

[0109] The processing method of the computer device can refer to the description of the above method and will not be repeated here.

[0110] An embodiment of the present application further provides a non-transitory machine-readable storage medium having an executable program stored thereon. When the executable program is executed by a microprocessor, the processor executes the method provided in the above embodiment.

[0111] An embodiment of the present invention discloses a computer-readable storage medium storing a computer program for electronic data exchange, wherein the computer program enables a computer to execute the described method.

[0112] An embodiment of the present invention discloses a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program, and the computer program is operable to cause a computer to execute the described method.

[0113] The embodiments described above are merely illustrative, wherein the modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules, i.e., they may be located in one place or distributed across multiple network modules. Some or all of the modules may be selected based on actual needs to achieve the objectives of the embodiments. Persons of ordinary skill in the art will be able to understand and implement the embodiments without inventive effort.

[0114] Through the detailed description of the above embodiments, those skilled in the art will clearly understand that each embodiment can be implemented by means of software plus a necessary general hardware platform, or of course, by means of hardware. Based on this understanding, the above technical solution, in essence, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, including a read-only memory (ROM), a random access memory (RAM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), a one-time programmable read-only memory (OTPROM), an electronically erasable programmable read-only memory (EEPROM), a compact disc read-only memory (CD-ROM), or other optical disc storage, magnetic disk storage, magnetic tape storage, or any other computer-readable medium capable of carrying or storing data.

[0115] Finally, it should be noted that the embodiments disclosed in the present invention are only preferred embodiments of the present invention, which are only used to illustrate the technical solutions of the present invention, rather than to limit them. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments can still be modified, or some of the technical features therein can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for constructing a knowledge graph of ancient Chinese medical books, characterized in that: The method comprises: Acquire ancient Chinese medical book data and perform standardization on the ancient Chinese medical book data; Combined with the domain dictionary, word segmentation priority rules are constructed, the minimum matching length and dictionary priority are set, and a bidirectional maximum matching algorithm is used to segment the ancient Chinese medical book data to generate structured text data; An entity recognition model is constructed based on a pre-trained language model in the field of Traditional Chinese Medicine. The entity recognition model is weakly supervised through pseudo-label expansion and introduces domain adversarial training to achieve cross-language semantic transfer, outputting entities and corresponding entity annotation results. A single-book graph structure is constructed using the entities as nodes and the contextual semantics of the entities as initial edge weights. A graph neural network model is used to learn the representations of the nodes, integrating the contextual features, location information, and graph structural features of the entities. The attention mechanism is used to extract semantic relationships between entities and construct a single-book knowledge graph. Based on the cross-graph attention mechanism, entities from multiple single-book knowledge graphs are fused, and a cross-graph attention matrix with weights equal to the product of semantic similarity, structural similarity and path similarity is constructed. The matching threshold of entities is dynamically adjusted according to the entity interaction intensity and the number of common neighbors, and entity alignment and edge weight fusion are performed to generate a global knowledge graph of ancient Chinese medicine books.

2. The method for constructing a knowledge graph of ancient Chinese medical books according to claim 1, characterized in that: Weakly supervised training is performed through pseudo-label extension, including: Extract entity classification pairs from the domain dictionary, where the entity classification pairs are initial positive samples; Extract entity feature vectors through the pre-trained language model, calculate the cosine similarity between entity feature vectors, determine the extended entity based on the cosine similarity, and obtain an extended pseudo-label training set; the extended pseudo-label training set includes automatically annotated entities, entity segments, and relationship triples; Calculate the semantic similarity between entities and text fragments, and select text fragments with confidence levels higher than a preset confidence threshold; When at least two entities exist in multiple text segments, there is a potential relationship between the two entities, and a weakly supervised training sample is obtained; the weakly supervised training sample includes a text segment, an entity label sequence, and a relationship label.

3. The method for constructing a knowledge graph of ancient Chinese medical books according to claim 2, characterized in that: Domain adversarial training is introduced to achieve cross-language semantic transfer, including: Setting a gradient reversal layer in the BERT layer and the task layer of the pre-trained language model to perform gradient reversal on the ancient book feature vector and the modern Chinese medicine feature vector output by the BERT layer; Inputting the gradient-inverted ancient book feature vector and modern Chinese medicine feature vector into a domain discriminator; the domain discriminator is used to discriminate the source of the corpus; Based on the entity recognition cross entropy loss function and the domain adversarial loss function, a weighted combination task loss function is constructed to optimize the entity recognition model.

4. The method for constructing a knowledge graph of ancient Chinese medical books according to claim 3, characterized in that: The domain adversarial loss function is: in, is the domain adversarial loss function; is the feature output by the BERT layer, The data distribution of ancient Chinese medical books; Represents the sample distribution from ancient Chinese medical books x Find the expectation; G is the domain discriminator; The task loss function is: in, is the domain adversarial loss function, is the cross entropy loss function for entity recognition, is the balance parameter.

5. The method for constructing a knowledge graph of ancient Chinese medical books according to claim 3, characterized in that: The single-book graph structure is constructed using the entity as a node and the contextual semantics of the entity as the initial edge weight. The node representation learning is performed on the graph neural network model, and the contextual features, location information and graph structure features of the entity are integrated. The semantic relationship between entities is extracted using the attention mechanism to construct a single-book knowledge graph, including: It traverses all entities and extracts entity pairs and their context fragments based on their co-occurrence relationships in ancient texts to form an entity pair set. Encode the text of the context segment into a word vector, use a multi-head self-attention mechanism to extract the context features of the entity pair, and generate a fused semantic feature vector of the entity pair; Constructing a graph structure based on the co-occurrence relationship of entities, and using the fused semantic feature vector as the initial feature of the node, performing multi-layer representation learning on the node through a graph convolutional network to obtain node features containing global semantic information; Aggregate and classify the entity pair feature vectors output by the graph neural network, determine the relationship type between entities, and generate entity relationship triples; the entity relationship triples contain the source entity, target entity, relationship category, context fragment, confidence level, and source book; Using entities as nodes and triple relationships as edges, combined with attribute information, a dynamically updated single-book knowledge graph is constructed.

6. The method for constructing a knowledge graph of ancient Chinese medical books according to claim 2, characterized in that: Based on the cross-graph attention mechanism, entities from multiple single-book knowledge graphs are integrated to construct a cross-graph attention matrix whose weights are the product of semantic similarity, structural similarity, and path similarity. The matching threshold of entities is dynamically adjusted according to the entity interaction strength and the number of common neighbors, including: Construct a cross-graph attention matrix, where the weight of the cross-graph attention matrix is determined by the product of semantic similarity, structural similarity, and path similarity between entity pairs; Calculating an entity interaction vector based on the cross-graph attention matrix; The matching threshold is dynamically adjusted according to the interaction strength between entities, the number of common neighbors, and the path similarity.

7. The method for constructing a knowledge graph of ancient Chinese medical books according to claim 6, characterized in that: Also includes: Receive newly added single book graphs containing new entity and relationship information, as well as existing global knowledge graph data; Calculate the semantic similarity, structural similarity, and path similarity between each entity in the newly added single-book graph and the entities in the global graph, and prioritize entity alignment based on matches with high structural similarity; For entities that have completed alignment, the newly added relationship paths are automatically mapped to the corresponding nodes in the global graph, their neighbor structure and relationship path features are updated, the dynamic inheritance of entity relationships is realized, and the updated global knowledge graph is output.

8. The method for constructing a knowledge graph of ancient Chinese medical books according to claim 6, characterized in that: The entity interaction vector is calculated by the following formula: in, For Entity The interaction vector, For entities in other graphs, For Entity and Cross-graph attention weights between ; The matching threshold is adjusted by the following formula: in is the initial threshold, is the adjustment coefficient of entity interaction strength, is an entity pair; It is a comprehensive indicator of the number of common neighbors across graphs and path similarity.

9. A computer device, characterized in that: include: A processor and a memory; wherein the memory stores a computer program, and the computer program is suitable for being loaded by the processor and executing the steps of the method for constructing a knowledge graph of ancient Chinese medicine books as claimed in any one of claims 1 to 8.

10. A computer storage medium, characterized in that A computer program is stored thereon, and when the computer program is executed by a processor, the steps of the method for constructing a knowledge graph of ancient Chinese medicine books as claimed in any one of claims 1 to 8 are implemented.

Images