Automatic labeling and semantic analysis system for ancient Chinese literature based on deep learning

Abstract

The invention discloses an ancient Chinese literature automatic annotation and semantic analysis system based on deep learning, and belongs to the technical field of natural language processing, and the system comprises a cross-time semantic evolution tracking module, a traditional Chinese medicine knowledge fusion enhancement module, a multi-granularity cross attention disambiguation module and a self-adaptive closed-loop feedback optimization module. The method comprises the following steps: capturing a word historical evolution track through a time sequence diagram neural network, fusing a traditional Chinese medicine knowledge graph through a graph attention network, realizing word-word-sentence three-granularity collaborative disambiguation through a multi-head cross attention mechanism, forming closed-loop optimization through a confidence-driven feedback mechanism, and deeply coupling four modules to form a collaborative interaction mechanism. The labeling accuracy rate reaches 92.3%, the processing efficiency is improved by 85%, and the method is particularly suitable for automatic processing and knowledge mining of traditional Chinese medicine ancient books.

Description

Technical Field

[0001] The present invention relates to the technical fields of natural language processing and deep learning, and particularly relates to an automatic annotation and semantic analysis system for ancient Chinese literature based on deep learning, which is particularly suitable for the automated processing and knowledge mining of ancient Chinese medical literature. Background Art

[0002] Chinese civilization has a long history, and there are a vast number of ancient literature classics throughout the dynasties, which contain rich historical cultures and scientific knowledge. Especially in the field of traditional Chinese medicine, ancient books such as "Huangdi Neijing", "Treatise on Febrile Diseases", and "Compendium of Materia Medica" record a large amount of precious medical theories and clinical experiences. However, the digital processing of ancient Chinese literature faces many challenges: the phenomenon of polysemy is common, the meaning of words changes significantly over time, it is difficult to understand professional terms, and there are no punctuation and word segmentation marks. How to use modern artificial intelligence technology to achieve the automatic annotation and semantic analysis of ancient Chinese literature has become an important topic in the research of digital humanities and the modernization of traditional Chinese medicine.

[0003] Chinese Patent CN1*9829159A discloses an integrated automatic morphological analysis method for ancient Chinese texts. This method uses the Word2Vec model for pre-training of word vectors, uses the Bi-LSTM-CRF neural network model to achieve integrated processing of sentence segmentation, word segmentation, and part-of-speech tagging, and constructs a special name database for ancient books to assist named entity recognition. This method improves the automatic processing efficiency of ancient Chinese texts to a certain extent, but has the following deficiencies:

[0004] First, the Word2Vec model can only learn static word vector representations and cannot capture the dynamic evolution law of word semantics in the long history. Ancient Chinese spans thousands of years, and the same word often has different semantics in different periods. For example, "tang" refers to hot water in the pre-Qin period and specifically refers to decocted medicinal liquid in Han Dynasty medical literature. The existing method mixes the corpora of all periods for training, ignoring the temporal evolution characteristics, resulting in inaccurate semantic understanding of cross-era literature.

[0005] Second, although the existing method constructs a special name database for ancient books, it is only used as a dictionary matching tool and fails to deeply mine the semantic associations and domain knowledge between special names. Especially in the field of traditional Chinese medicine, there are complex compatibility relationships between prescriptions and medicinal materials, and there are diverse treatment correspondences between medicinal materials and diseases. These structured domain knowledge is crucial for accurately understanding the content of ancient books, but the existing method fails to effectively use this knowledge for semantic enhancement.

[0006] Third, the Bi-LSTM-CRF model uses a single-granularity sequence labeling approach, performing feature extraction and label prediction only at the word sequence level, lacking collaborative modeling of semantic information at multiple granularities, including characters, words, and sentences. Classical Chinese exhibits a prominent polysemy, and accurate word meaning disambiguation requires comprehensive consideration of multi-level information such as character word formation rules, word collocation patterns, and sentence context. A single-granularity approach is insufficient to effectively resolve ambiguity issues.

[0007] Fourth, existing methods lack an adaptive optimization mechanism based on annotation quality. After model training, the parameters remain fixed and cannot be dynamically adjusted according to the actual annotation results. For difficult samples with low annotation confidence, the system cannot automatically identify and optimize the relevant module parameters, resulting in limited improvement in overall annotation quality and difficulty in adapting to complex and ever-changing ancient texts.

[0008] Therefore, there is an urgent need to develop a new automatic annotation and semantic analysis system for ancient Chinese literature that can capture the cross-era semantic evolution of words, deeply integrate professional knowledge from fields such as traditional Chinese medicine, achieve multi-granularity collaborative semantic understanding and word meaning disambiguation, and have adaptive closed-loop optimization capabilities based on annotation quality, so as to significantly improve the level of automated processing of ancient Chinese literature, especially ancient Chinese medicine literature. Summary of the Invention

[0009] The purpose of this invention is to provide an automatic annotation and semantic analysis system for ancient Chinese documents based on deep learning, so as to solve the technical problems existing in the prior art, such as the inability to capture semantic evolution across eras, the failure to make in-depth use of domain knowledge, the lack of multi-granularity collaborative disambiguation, and the lack of adaptive optimization mechanisms.

[0010] To achieve the above objectives, the present invention provides the following technical solution:

[0011] This invention provides an automatic annotation and semantic analysis system for ancient Chinese literature based on deep learning. The system achieves high-precision automatic annotation and semantic analysis of ancient Chinese literature, especially ancient Chinese medicine books, through the collaborative work of four deeply coupled core modules.

[0012] The system includes a cross-era semantic evolution tracking module, which constructs a temporal evolution map of ancient Chinese corpus according to historical periods, uses a temporal graph neural network to capture the historical change trajectory of word semantics, and generates an evolutionary semantic representation containing the time dimension, providing prior knowledge of historical evolution for subsequent semantic understanding.

[0013] The system includes a TCM knowledge fusion and enhancement module, which constructs a TCM knowledge graph containing entities such as prescriptions, medicinal materials, and diseases, as well as their compatibility and treatment relationships. Through a graph attention network, the structured domain knowledge in the knowledge graph is deeply fused with textual semantic features across modalities to generate an enhanced semantic representation that integrates professional knowledge.

[0014] The system includes a multi-granularity cross-attention disambiguation module, which receives the aforementioned evolved semantic representation and enhanced semantic representation, models semantic features at three granularities of character level, word level, and sentence level respectively, captures the dependencies between different granularities through the multi-head cross-attention mechanism, realizes context-aware word sense disambiguation, and generates an integrated annotation result including sentence segmentation, word segmentation, part-of-speech, and entity annotation.

[0015] The system includes an adaptive closed-loop feedback optimization module, which calculates the confidence score of the annotation result, identifies low-confidence samples, generates a feedback adjustment signal, and dynamically optimizes the parameters of the cross-era semantic evolution tracking module and the multi-granularity cross-attention disambiguation module through gradient backpropagation, forming a closed-loop optimization mechanism driven by annotation quality.

[0016] A deeply coupled collaborative relationship is formed among the four modules: the evolved semantic representation output by the cross-era semantic evolution tracking module is used as the prior knowledge input of the multi-granularity cross-attention disambiguation module to enhance its perception ability of historical semantic changes; the enhanced semantic representation generated by the traditional Chinese medicine knowledge fusion and enhancement module strengthens the feature extraction of the multi-granularity cross-attention disambiguation module through attention weighting; the annotation result of the multi-granularity cross-attention disambiguation module is fed back to the cross-era semantic evolution tracking module to optimize its evolution pattern learning; the adaptive closed-loop feedback optimization module adjusts the parameters of the cross-era semantic evolution tracking module and the multi-granularity cross-attention disambiguation module simultaneously according to the annotation quality, achieving the technical effects of deep coupling at the parameter level, closed-loop feedback at the state level, and synergistic efficiency among the four modules.

[0017] Compared with the prior art, the present invention has the following remarkable advantages:

[0018] First, through the cross-era semantic evolution tracking module, the present invention first incorporates the historical evolution of word semantics into the ancient Chinese understanding framework, and explicitly models the semantic change rules between different historical periods using a temporal graph neural network. Compared with the static modeling method of mixing corpora from all periods in the prior art, the present invention can accurately capture historical evolution phenomena such as the semantic evolution of "tang" from hot water to medicinal liquid, and the meaning transformation of "qi" from air to the core concept of traditional Chinese medicine theory, making the system's semantic understanding of cross-era literature more accurate. Experiments show that on the cross-era corpus test set, the semantic annotation accuracy of the present invention is improved by 12.6% compared with the prior art.

[0019] Second, through the traditional Chinese medicine knowledge integration enhancement module, the present invention deeply integrates the structured professional knowledge in the field of traditional Chinese medicine into the semantic understanding process. Different from the existing technology which only performs simple dictionary matching, the present invention constructs a knowledge graph containing entity association relationships, and learns semantic relationships such as compatibility and treatment between entities through a graph attention network, realizing the cross-modal deep integration of knowledge and text. This enables the system to accurately recognize complex prescription names such as "Guizhi Decoction", understand the compatibility relationships of medicinal materials such as "Ramulus Cinnamomi, Paeoniae Radix Alba, Glycyrrhizae Radix et Rhizoma", and grasp the theoretical connotations of traditional Chinese medicine for diseases such as "Taiyang disease". On the special test set of traditional Chinese medicine ancient books, the F1 score of the present invention for the recognition of prescription entities reaches 94.7%, an increase of 18.3% compared with the existing technology.

[0020] Third, through the multi-granularity cross-attention disambiguation module, the present invention collaboratively models semantic features at three granularities of character, word, and sentence, overcoming the limitations of single-granularity processing in the existing technology. Character-level attention captures the word formation rules of single characters, word-level attention extracts word collocation patterns, and sentence-level attention models the global context. The features of the three granularities are deeply integrated through the cross-attention mechanism to achieve precise disambiguation of polysemous phenomena such as "fa" (hair / growth / dispersion). Experiments show that the accuracy of word sense disambiguation of the present invention reaches 92.3%, an increase of 9.8% compared with the existing technology, and it has obvious advantages especially in the processing of highly polysemous words.

[0021] Fourth, through the adaptive closed-loop feedback optimization module, the present invention establishes a parameter dynamic adjustment mechanism based on annotation quality. The system can automatically identify low-confidence annotation samples, generate targeted feedback signals, and simultaneously optimize the parameters of the two core modules of evolution tracking and disambiguation annotation through gradient backpropagation, forming a complete closed loop of "annotation → evaluation → feedback → optimization". This adaptive mechanism enables the system to continuously self-optimize when processing difficult samples, and the annotation performance continues to improve. Experiments show that after closed-loop optimization, the accuracy of the system on the difficult example test set increases by 7.4%, and the overall annotation efficiency increases by 85%.

[0022] Fifth, the deep coupling and closed-loop collaborative mechanism formed among the four modules realizes a synergistic effect of 1 + 1 > 2. The evolving semantic representation provides historical priors for disambiguation, the knowledge graph embedding injects domain wisdom into disambiguation, the disambiguation results feedback to optimize the evolution modeling, and the quality assessment drives global parameter adjustment. The four modules promote each other and have a superimposed synergistic effect. Compared with the existing technology where each module works independently, the module collaboration of the present invention significantly improves the overall performance of the system, reaching an accuracy of 92.3% in the comprehensive annotation task of classical Chinese, an increase of 12.6 percentage points compared with 79.7% of the existing technology, and the annotation efficiency increases by 85%, providing a strong technical support for the digital processing and knowledge mining of traditional Chinese medicine ancient books. Brief Description of the Drawings

[0023] Figure 1It is a schematic diagram of the overall architecture of the system of the present invention. Detailed implementation manners

[0024] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0025] As Figure 1 shown, the automatic annotation and semantic analysis system for ancient Chinese literature based on deep learning provided by the present invention includes a cross-era semantic evolution tracking module 1, a traditional Chinese medicine knowledge fusion enhancement module 2, a multi-granularity cross-attention disambiguation module 3, and an adaptive closed-loop feedback optimization module 4. The four modules work together through a deeply coupled data flow and parameter sharing mechanism to form a complete closed-loop optimization system.

[0026] The cross-era semantic evolution tracking module 1 is responsible for capturing the evolution law of the semantics of ancient Chinese words in the long history and generating an evolution semantic representation including the time dimension. This module includes a time-series corpus construction unit, an evolution graph generation unit, and a time-series graph convolution unit.

[0027] The time-series corpus construction unit divides the collected ancient Chinese corpus into six time periods according to historical periods: pre-Qin (from the 21st century BC to 221 BC), Qin and Han (from 221 BC to 220 AD), Wei, Jin, Southern and Northern Dynasties (from 220 AD to 589 AD), Sui, Tang and Five Dynasties (from 589 AD to 960 AD), Song, Liao, Jin and Yuan (from 960 AD to 1368 AD), Ming and Qing (from 1368 AD to 1911 AD). This division not only takes into account the historical facts of dynasty changes but also takes into account the continuity characteristics of language evolution. For each word, the system extracts all the occurrence contexts within each time period and constructs a time-series semantic node sequence. For example, for the word "tang", it mainly appeared in daily life contexts such as "bathing soup" in the pre-Qin period, began to appear in medical contexts such as "Guizhi Decoction" and "Mahuang Decoction" in the Qin and Han periods, and dominated in medical literature in the Ming and Qing periods.

[0028] The evolution graph generation unit constructs a cross-era word evolution graph based on the time-series semantic node sequence. This graph adopts a directed graph structure, where nodes represent the semantic states of words in specific historical periods, and edges represent the semantic evolution relationships between adjacent time periods. Specifically, for the word w, its semantic node in the time period is represented as , and the evolution edge from the time period to is represented as The construction process of the evolution graph fully considers the principle of semantic continuity, that is, the semantic states of adjacent periods should have a high correlation, while also allowing for significant semantic changes. In actual implementation, the system first initializes the embedding vector of the word semantics for each period, and then calculates the semantic similarity between periods by analyzing the distribution of co-occurring words, syntactic patterns and contextual features of the word in different periods, and constructs the evolution relationship edges.

[0029] The temporal graph convolutional unit applies a temporal graph convolutional network to extract features from the constructed evolutionary graph. This unit employs a gated temporal graph convolutional network, introducing a gating mechanism when aggregating features from temporal neighborhood nodes. For word w within a time period... nodes Its updated representation Calculated using the following innovative algorithm: ,

[0030] in, For the word w in the time period The hidden state vector, This is the hidden state vector from the previous time period. This is the input feature vector for the current time period. Here is the forget gate weight matrix. To update the gate weight matrix, The candidate state weight matrix, It is the Sigmoid activation function. This is an element-wise multiplication operation. The hyperbolic tangent activation function is used. This represents a vector concatenation operation. The first term... Forgetting gates control how much historical semantic information is retained from the previous period; the second item... The update gate determines the proportion of semantic features absorbed in the current period. This gating mechanism enables the model to adaptively balance the inheritance of historical semantics with the updating of current semantics, accurately capturing the continuity and abrupt changes in semantic evolution.

[0031] In practical implementation, the hidden state vector and input feature vector The dimensions are all set to 256. The output range of the forget gate and update gate is... , when the output of the forgetting gate is close to 1, it indicates that the semantics of the word remain stable in adjacent periods and historical information needs to be fully retained; when the output of the update gate is close to 1, it indicates that significant semantic changes occur in the current period and semantic representations need to be greatly updated. For example, for the word "tang" (decoction), during the transition from the pre-Qin period to the Qin and Han dynasties, the output of the forgetting gate was approximately 0.6 and the output of the update gate was approximately 0.7, indicating that the word retained part of its original meaning of "hot water" while absorbing the new meaning of "medicinal liquid".

[0032] The temporal graph convolutional network gradually aggregates temporal neighborhood information over a larger range through multiple layers (preferably 3 layers), and finally generates an evolving semantic representation vector that integrates the complete historical evolution trajectory of each word. This vector not only contains the current semantics of the word but also implies its semantic change patterns from ancient times to the present, providing rich prior knowledge of historical evolution for subsequent semantic understanding and word sense disambiguation.

[0033] The traditional Chinese medicine knowledge fusion enhancement module 2 is responsible for integrating structured professional knowledge in the field of traditional Chinese medicine into the text semantic understanding process. This module includes a knowledge graph construction unit, an entity linking unit, and a cross-modal fusion unit.

[0034] The knowledge graph construction unit extracts entities and relationships from classic traditional Chinese medicine literature such as "Huangdi Neijing", "Shanghan Lun", "Jingui Yaolue", "Bencao Gangmu", and "Wenbing Tiaobian" to construct a traditional Chinese medicine knowledge graph. This graph contains five types of core entities: formula entities (such as "Guizhi Tang", "Xiaochaihu Tang"), medicinal material entities (such as "Guizhi", "Baishao", "Gancao"), disease entities (such as "Taiyang disease", "Shaoyang disease"), acupoint entities (such as "Hegu Point", "Zusanli Point"), and meridian entities (such as "Hand Taiyin Lung Meridian", "Foot Yangming Stomach Meridian"). Seven relationships are defined between entities: treatment relationship (formula → disease), compatibility relationship (medicinal material → medicinal material), composition relationship (formula → medicinal material), meridian tropism relationship (medicinal material → meridian), main treatment relationship (acupoint → disease), circulation relationship (meridian → human body part), and efficacy relationship (medicinal material → efficacy). In specific implementation, the knowledge graph contains 21,647 entity nodes and 53,829 relationship edges, forming a rich knowledge network in the field of traditional Chinese medicine.

[0035] The entity extraction process adopts a deep learning-based named entity recognition method and combines a traditional Chinese medicine domain dictionary for semi-automatic extraction. For formula entities, the system recognizes typical patterns such as "XX Tang", "XX San", "XX Wan", etc.; for medicinal material entities, the system refers to the names of medicinal materials included in the "Pharmacopoeia of the People's Republic of China"; for disease entities, the system extracts disease name expressions such as "XX disease", "XX syndrome", etc. Relationship extraction adopts a method based on dependency syntax analysis and pattern matching, such as recognizing the treatment relationship in "XX Tang treats XX disease" and the compatibility relationship in "XX is compatible with XX", etc.

[0036] The entity linking unit is responsible for matching candidate entities identified in the text to be labeled with standard entities in the knowledge graph. This unit first uses a named entity recognition model to detect candidate entities in the input text, then calculates the cosine similarity between the text representation of the candidate entity and the embedding vectors of each entity in the knowledge graph, selecting the entity with the highest similarity exceeding a threshold (preferably 0.75) as the link result. For ambiguous cases, the system comprehensively considers the context and the entity's connection relationships in the knowledge graph for disambiguation. For example, for the term "white tiger," the system determines whether it refers to "white tiger soup" (a prescription) or "white tiger" (an animal) based on the context, and determines the correct link by analyzing whether surrounding elements include medicinal materials such as "gypsum" or "anemarrhena" or symptoms such as "fever" or "high fever."

[0037] The cross-modal fusion unit employs a graph attention network to integrate structured knowledge from the knowledge graph into the text semantic representation. This unit first learns an embedding vector for each entity in the knowledge graph, and then computes the attention weights between entity nodes. For each entity node... Its updated embedding representation Calculated using the following innovative algorithm:

[0038] ,

[0039] ,

[0040] ,

[0041] in, For entities Updated embedding vector, and Entities and The initial embedding vector, For entities The set of neighbor nodes in a knowledge graph These are the normalized attention weight coefficients. For learnable graph convolution weight matrices, For activation function, To score attention, For attention parameter vectors, For entities and Embedded vectors of the relationship between them This represents a vector concatenation operation. The Leaky ReLU activation function is used, with its negative half-axis slope set to 0.2. The attention score calculation considers both node feature similarity and relationship type information, resulting in higher attention weights for semantically related entity pairs with important relationships.

[0042] In the specific implementation, the entity embedding vector dimension is set to 128, and the graph convolution weight matrix... Dimensions For TCM formula entities, the system assigns them a high base weight (preferably 1.2 times) when calculating attention weights, as formulas are often the core carriers of TCM knowledge. For pairs of medicinal materials with compatibility relationships, the system dynamically adjusts the attention weights according to the compatibility level (principal, assistant, adjuvant, and guide): the weight coefficient for the principal herb is 1.5, for the assistant herb 1.2, for the adjuvant herb 1.0, and for the guide herb 0.8. This weighting strategy allows the knowledge graph to highlight the core concepts and key relationships in TCM theory.

[0043] Cross-modal fusion combines knowledge graph entity embeddings with text features through an attention-weighted mechanism. For each word's representation vector in the text... The system calculates its embedding with the linked entity. Attention score, generating enhanced representation after fusion. The fusion process employs residual connections and layer normalization to ensure training stability. Ultimately, the enhanced semantic representation output by this module incorporates both the linguistic features of the original text and the structured domain knowledge of the traditional Chinese medicine knowledge graph, providing domain-specific intelligence support for subsequent semantic disambiguation.

[0044] The multi-granularity cross-attention disambiguation module 3 is the core annotation engine of the system. It is responsible for collaboratively modeling semantic features at the character, word, and sentence levels to achieve accurate word meaning disambiguation and integrated annotation. This module includes character-level attention units, word-level attention units, sentence-level attention units, and multi-granularity fusion units.

[0045] Character-level attention units compute self-attention distributions at the character granularity, capturing the positional features and word formation patterns of individual characters within words. For the input text sequence... ,in Indicates the first For each character, word-level attention first maps each character to a query vector. Key vector Sum value vector Attention weights between characters The calculation method is as follows:

[0046] ,

[0047] in, Divide the character embedding dimension (set to 64) by This is a scaling factor to prevent the gradient of the Softmax function from vanishing due to excessively large dot product values. Word-level attention output Obtained by weighted summation:

[0048] ,

[0049] Character-level attention can capture character-level semantic information such as the association between "ji" and "bing" in "disease" and the word formation patterns of "fa" and "san" in "fasan". This is crucial for understanding common compound words and semantic extension phenomena in ancient Chinese.

[0050] The word-level attention unit calculates cross-word attention weights at the word granularity, modeling the syntactic dependency relationships and semantic collocation patterns between words. The system first performs word segmentation preprocessing to split the character sequence into a word sequence. . Then, combining the evolved semantic representation output by the cross-era semantic evolution tracking module 1 and the enhanced semantic representation output by the traditional Chinese medicine knowledge fusion and enhancement module 2, a comprehensive representation vector is constructed for each word. The word-level attention calculates the semantic dependency relationships between words:

[0051] ,

[0052] ,

[0053] where 、 、 are the query, key, and value vectors of words and respectively, is the word embedding dimension (set to 256), is the attention weight of word to word , is the word-level attention output of word . The word-level attention particularly focuses on syntactic collocation relationships such as verb-object, subject-predicate, and modifier-head, as well as the co-occurrence patterns of common compatibility word pairs such as "ramulus cinnamomi" and "radix paeoniae alba".

[0054] The sentence-level attention unit calculates global context attention at the sentence granularity, extracting long-distance semantic dependency and discourse coherence features. This unit adopts the Transformer encoder structure to model the global semantic representation of the sentence through multiple layers of self-attention mechanisms. The calculation of sentence-level attention follows the standard multi-head self-attention mechanism:

[0055] ,

[0056] ,

[0057] ,

[0058] where 、 , are the query matrix, the key matrix, and the value matrix respectively. , , is the projection matrix of the -th attention head. is the output projection matrix. is the number of attention heads (set to 8). is the dimension of each head (set to 32). represents concatenating the outputs of multiple heads. The multi-head mechanism enables the model to capture information from different semantic subspaces and enhances the representation ability. Sentence-level attention can model long-distance dependencies spanning multiple words, such as understanding the semantic association between "patient" and "recovery" in "The patient... took Guizhi Decoction... and recovered".

[0059] The multi-granularity fusion unit is responsible for adaptively fusing the attention features at the character level, word level, and sentence level. This unit adopts a hierarchical gating mechanism. First, it performs the first-layer fusion between the character level and the word level, and then combines the fusion result with the sentence-level features for the second-layer fusion. The gating weights are calculated as follows:

[0060] ,

[0061] ,

[0062] where and are the attention outputs at the character level and the word level respectively, is the gating weight matrix, is the bias vector, represents vector concatenation, limits the gating value to the interval using the Sigmoid function, is the result of the first-layer fusion. The value of the gating weight is determined by the characteristics of the current semantic disambiguation task: for cases where characters are similar in form and easily confused (such as "tang" and "chang"), the weight of the character-level features is higher; for cases of word collocation ambiguity (such as "sweating" and "fever"), the weight of the word-level features is higher.

[0063] The second-layer fusion combines the first-layer result with the sentence-level feature :

[0064] ,

[0065] ,

[0066] [[ID= For sentence-level attention output, and These are the second-level gating parameters. This results in the final multi-granularity fusion feature. This fusion feature integrates semantic information from three levels: character word formation, word collocation, and global context. It can effectively handle the polysemy of words in Classical Chinese and achieve accurate word meaning disambiguation.

[0067] Based on fusion features The system outputs integrated annotation results, including sentence segmentation markers, word boundary markers, part-of-speech tags, and named entity tags, through a classification layer. Sentence segmentation markers use a binary classification (sentence-in-sentence / sentence-end), word boundary markers use the BIES annotation system (word-beginning / word-middle / word-end / single-character words), part-of-speech tagging uses the standard of Peking University's "Modern Chinese Grammar Information Dictionary," and named entity tagging expands to include fine-grained entity types such as prescription names, medicinal material names, disease names, and acupoint names for the field of traditional Chinese medicine.

[0068] The adaptive closed-loop feedback optimization module 4 is responsible for evaluating annotation quality and driving the dynamic optimization of system parameters. This module includes a confidence calculation unit, a quality assessment unit, a feedback signal generation unit, and a parameter adaptive adjustment unit.

[0069] The confidence calculation unit calculates the labeled confidence based on the probability distribution output by the multi-granularity cross-attention disambiguation module 3. For the... The probability distribution of the classification layer output at each of the unlabeled locations is as follows: ,in The total number of label categories (for comprehensive annotation tasks, Typically 100-200). Confidence levels are calculated using an entropy-based uncertainty measure:

[0070] ,

[0071] ,

[0072] in, The information entropy of the probability distribution. The normalized confidence score has a range of values. Information entropy A lower entropy value indicates a more concentrated probability distribution, a more certain prediction outcome, and higher confidence. Conversely, a higher entropy value indicates a more uniform probability distribution, less model uncertainty, and lower confidence. For example, when the predicted probability distribution at a certain labeled location is... At that time, the information entropy was approximately 0.32, and the confidence level was as high as 0.92; when the predicted probability distribution was... At that time, the information entropy was approximately 1.35, and the confidence level was only 0.31.

[0073] The quality assessment unit compares the calculated confidence score with a dynamically adjusted confidence threshold to identify low-confidence samples that need optimization. The confidence threshold is set using an adaptive strategy: the initial threshold is set to 0.75, and it is dynamically adjusted based on overall annotation performance as training progresses. If the average accuracy in a given iteration exceeds 90%, the threshold is increased by 0.02; if the average accuracy is below 85%, the threshold is decreased by 0.02. This adaptive threshold mechanism allows the system to flexibly adjust its optimization objectives based on the current performance level. When the confidence score at a certain annotation location falls below the threshold, that location is marked as a low-confidence sample, triggering subsequent feedback optimization processes.

[0074] The feedback signal generation unit generates a feedback signal containing the gradient adjustment direction and magnitude for the identified low-confidence samples. This unit employs a feedback mechanism based on the loss function gradient to calculate the loss value corresponding to the low-confidence samples. and its model parameters gradient The loss function used is cross-entropy loss:

[0075] ,

[0076] in, For the first The position The actual label of the class (0 or 1). This represents the probability predicted by the model. For low-confidence samples, the system increases their loss weight. To enhance and optimize the effect:

[0077] ,

[0078] in, This is the weighting amplification factor (preferably 2.0). The weighted loss is... .gradient It indicates the direction of parameter optimization, the magnitude of the gradient This reflects the required adjustment range. The feedback signals comprehensively encode this information, forming parameter adjustment instructions for the cross-generational semantic evolution tracking module 1 and the multi-granularity cross-attention disambiguation module 3.

[0079] The parameter adaptive adjustment unit updates the model parameters based on the feedback signal through gradient backpropagation. This unit uses the Adam optimization algorithm for parameter updates, with a warmup learning rate: linearly increasing to a peak of 0.0001 for the first 1000 steps, then gradually decreasing using a cosine annealing strategy. The parameter update formula is:

[0080] ,

[0081] ,

[0082] ,

[0083] ,

[0084] in, For the first The model parameters of the step, For the first The learning rate of the step. and These are the first-order moment estimate and the second-order moment estimate of the gradient, respectively. and The momentum hyperparameters are set to 0.9 and 0.999 respectively. For numerical stability terms (set to) ), and This is the moment estimate after bias correction. The Adam algorithm combines momentum methods and adaptive learning rates to stably and efficiently optimize deep neural network parameters.

[0085] The adaptive parameter adjustment simultaneously affects the parameters of the temporal graph convolutional layer in the cross-era semantic evolution tracking module 1 and the attention layer parameters in the multi-granularity cross-attention disambiguation module 3. For the evolution tracking module, the feedback signal prompts it to optimize the learning of word evolution patterns, making the semantic evolution representation more accurately reflect historical changes. For the disambiguation module, the feedback signal drives it to adjust the allocation of attention weights at each granularity, improving its ability to distinguish difficult samples. This bidirectional parameter adjustment forms a closed-loop optimization mechanism driven by annotation quality, enabling the system to continuously improve itself when handling difficult examples.

[0086] The four core modules form a deeply coupled collaborative relationship through a carefully designed data flow and parameter sharing mechanism, achieving a synergistic effect where 1+1>2.

[0087] First layer of coupling: Evolutionary semantic representation output by the cross-era semantic evolution tracking module 1 The query vector is dimensionally aligned using a feature mapping layer and a multi-granularity cross-attention disambiguation module 3. Specifically, the evolutionary semantic representation first undergoes a linear transformation:

[0088] ,

[0089] in, For the mapping weight matrix, This is the bias vector. The mapped representation. This serves as prior knowledge input to the query vector of the disambiguation module, enhancing its ability to perceive the historical semantic changes of words. This coupling enables the disambiguation module to dynamically adjust its semantic understanding strategy based on the historical evolution of words when processing cross-era documents.

[0090] Second-layer coupling: Enhanced semantic representation output by the TCM knowledge fusion enhancement module 2 The key-value vector of the multi-granularity cross-attention disambiguation module 3 is enhanced using an attention weighting method. This enhancement is achieved by calculating the semantic relevance between knowledge graph entities and text words.

[0091] ,

[0092] ,

[0093] ,

[0094] in, For word-level query vectors, and It is a word-level key-value vector. To enhance the weight of knowledge graphs, and This is the enhanced key-value vector. This coupling directly injects knowledge from the field of traditional Chinese medicine into the attention calculation process of the disambiguation module, enabling it to accurately determine the semantics of words by leveraging relationships such as compatibility and treatment in the knowledge graph.

[0095] The third layer of coupling: The disambiguation results of the multi-granularity cross-attention disambiguation module 3 are fed back to the cross-era semantic evolution tracking module 1 through a contrastive learning mechanism to optimize its evolution pattern learning. For multiple semantics of the same word in different historical periods, the correct annotation results of the disambiguation module are used as supervision signals to guide the evolution tracking module to adjust the parameters of the temporal graph convolutional network, so that the generated evolutionary semantic representation can better distinguish the semantic differences in different periods.

[0096] The fourth layer of coupling: The feedback signal generated by the adaptive closed-loop feedback optimization module 4 simultaneously acts on both the cross-generational semantic evolution tracking module 1 and the multi-granularity cross-attention disambiguation module 3. For the evolution tracking module, the feedback signal updates the parameters of the temporal graph convolutional layer through gradient backpropagation, especially the weights of the forget and update gates, making the model's capture of semantic evolution patterns more accurate. For the disambiguation module, the feedback signal adjusts the parameters of each granularity attention layer, optimizes the fusion weights of character-level, word-level, and sentence-level features, and improves the ability to distinguish between complex and ambiguous meanings.

[0097] The complete workflow of the four modules working together is as follows: The input ancient Chinese text first enters the cross-era semantic evolution tracking module 1, which generates a semantic representation containing historical evolution information; at the same time, the text is input into the traditional Chinese medicine knowledge fusion enhancement module 2, which generates an enhanced representation of fused domain knowledge through entity links and graph attention networks; the outputs of the two modules are jointly input into the multi-granularity cross-attention disambiguation module 3, which performs collaborative semantic understanding and word meaning disambiguation at the three granularities of characters, words, and sentences, generating an integrated annotation result; the annotation result enters the adaptive closed-loop feedback optimization module 4 for quality evaluation. For low-confidence samples, this module generates feedback signals to adjust the parameters of the evolution tracking and disambiguation annotation modules in reverse, forming a closed-loop optimization; after several rounds of iteration, the system performance is continuously improved, and finally high-quality annotation results are output.

[0098] This deeply coupled closed-loop collaborative mechanism achieves synergistic effects at multiple levels: evolutionary semantic representation provides historical priors for disambiguation, knowledge graph embedding injects domain wisdom into disambiguation, disambiguation results are fed back to optimize evolutionary modeling, and quality assessment drives global parameter adjustment. The four modules promote each other and synergize, making the overall system performance far exceed the simple sum of the individual modules working independently, achieving a synergistic effect of 1+1>2.

[0099] The system training employs a multi-stage strategy. In the first stage, a large-scale classical Chinese corpus is used to pre-train the cross-era semantic evolution tracking module 1, learning the historical evolution patterns of words. The training corpus includes classic texts from the pre-Qin period to the Ming and Qing dynasties, totaling approximately 500 million characters. In the second stage, a knowledge graph is constructed using ancient Chinese medicine texts, and this graph is used to train the traditional Chinese medicine knowledge fusion enhancement module 2. The knowledge graph contains 21,647 entities and 53,829 relationships, and the training corpus includes classic texts such as the *Huangdi Neijing*, *Shanghan Lun*, *Jinkui Yaolue*, and *Bencao Gangmu*, totaling approximately 20 million characters. In the third stage, manually annotated classical Chinese corpus is used to supervise the training of the multi-granularity cross-attention disambiguation module 3. The annotated corpus includes sentence segmentation, word segmentation, part-of-speech tagging, and entity tagging, totaling approximately 10 million characters. In the fourth stage, the four modules are integrated for end-to-end joint training, and the adaptive closed-loop feedback optimization module 4's dynamic adjustment mechanism is enabled to continuously optimize system parameters during training.

[0100] The model was implemented using the PyTorch framework, trained on eight NVIDIA A100 GPUs with a batch size of 32 and 50 training epochs. The Adam optimization algorithm was used with an initial learning rate of 0.0001, adjusted using warmup and cosine annealing. To prevent overfitting, Dropout (scale 0.1) and weight decay (coefficient 0.01) regularization techniques were employed.

[0101] When the system is applied, the user inputs the ancient Chinese text to be annotated. The system automatically completes sentence segmentation, word segmentation, part-of-speech tagging, and named entity recognition, and outputs an integrated annotation result. For ancient Chinese medical books, the system can also identify proper nouns such as prescription names, medicinal material names, and disease names, and associate with relevant entities in the knowledge graph to provide users with rich semantic information and domain knowledge. The system's annotation speed is about 1000 characters per second, with an efficiency improvement of 85% compared to manual annotation, and an annotation accuracy rate of 92.3%.

[0102] Example 1: Annotate the sentence "Taiyang disease, headache and stiff neck with aversion to cold, treated by Cinnamon Twig Decoction" in Treatise on Febrile Diseases.

[0103] First, the system analyzes the historical semantics of each word through the cross-era semantic evolution tracking module 1. For "Taiyang disease", the module identifies that it specifically refers to a type of exogenous disease in Han Dynasty medical literature, which is significantly different from the meaning of "sunlight" in pre-Qin literature. For "Cinnamon Twig Decoction", the module identifies it as a prescription name that began to be widely used in the Eastern Han Dynasty.

[0104] Next, the traditional Chinese medicine knowledge fusion enhancement module 2 retrieves relevant entities in the knowledge graph. It identifies "Taiyang disease" as a disease entity and links it to the knowledge graph node; identifies "Cinnamon Twig Decoction" as a prescription entity and retrieves its constituent medicinal materials "cinnamon twig, white peony root, licorice root, fresh ginger, and Chinese date"; identifies "headache and stiff neck" and "aversion to cold" as symptom descriptions; and establishes the relationship that "Cinnamon Twig Decoction" treats "Taiyang disease". Through the graph attention network, the module integrates the embedding vectors of these entities and their associated relationships into the text representation.

[0105] The multi-granularity cross-attention disambiguation module 3 conducts semantic analysis at three granularities. At the character level, the attention identifies "Taiyang" as a whole word, "disease" as a disease marker, "cinnamon twig" as a medicinal material name, and "decoction" as a prescription marker. At the word level, the attention models the parallel structure of "headache and stiff neck", the verb-object structure of "aversion to cold", and the subject-verb-object relationship of "treated by Cinnamon Twig Decoction". At the sentence level, the attention captures the causal logic expressed in the whole sentence: because the symptoms of "Taiyang disease" "headache and stiff neck with aversion to cold" appear, so "Cinnamon Twig Decoction" is used for treatment. Based on the multi-granularity fusion features, the system outputs the annotation result:

[0106] Taiyang disease / disease name, / punctuation headache / noun stiff / verb pain / verb and / conjunction aversion / verb cold / noun, / punctuation Cinnamon Twig Decoction / prescription name treat / verb it / pronoun. / punctuation.

[0107] The adaptive closed-loop feedback optimization module 4 calculates the confidence of each labeled position. For proper nouns such as "Taiyang disease" and "Guizhi Decoction", the confidence exceeds 0.95 and no optimization is required. For the词性标注 of "headache and stiffness", the initial confidence in predicting "stiffness" as an adjective is 0.68, which is lower than the threshold of 0.75. The module generates a feedback signal, adjusts the parameters of the disambiguation module, combines syntactic analysis and domain knowledge, and finally confirms that "stiffness" here is a verb meaning "rigid", and the confidence is increased to 0.89.

[0108] Example 2: Annotate the sentence "Licorice, with a neutral nature, a sweet taste, and belongs to the heart, lung, spleen, and stomach meridians" in the Compendium of Materia Medica in the Ming Dynasty.

[0109] The cross-era semantic evolution tracking module 1 analyzes the semantic evolution of "Licorice". In pre-Qin literature, "Licorice" mainly referred to herbaceous plants with a sweet taste; in Han Dynasty medical literature, it gradually became a specific medicinal material name; in Ming and Qing medical books, a complete theory of medicinal properties had been formed. The module captures this evolution trajectory to provide an accurate semantic representation for the current context.

[0110] The traditional Chinese medicine knowledge fusion enhancement module 2 retrieves the "Licorice" entity in the knowledge graph, identifies its medicinal properties as "neutral nature, sweet taste", and its meridian tropism as "heart meridian, lung meridian, spleen meridian, stomach meridian". The module also retrieves relevant compatibility information: Licorice is often compatible with qi-tonifying herbs such as Astragalus membranaceus and Codonopsis pilosula, and can harmonize the medicinal properties when compatible with yang-warming herbs such as Aconite root and dried ginger. This knowledge is integrated into the text representation through a graph attention network.

[0111] The multi-granularity cross-attention disambiguation module 3 performs annotation by integrating features at three granularities. The character-level attention identifies "Licorice" as a two-character word, and in "belongs to the heart", "heart" refers to the heart meridian rather than emotion. The word-level attention models the parallel structure of "neutral nature, sweet taste" and the enumeration structure of "belongs to the heart, lung, spleen, stomach meridians". The sentence-level attention understands that the whole sentence is a systematic description of the medicinal properties of Licorice. The system outputs the annotation:

[0112] Licorice / herb name, / punctuation nature / noun neutral / adjective, / punctuation taste / noun sweet / adjective, / punctuation belong / verb heart / meridian name, / punctuation lung / meridian name, / punctuation spleen / meridian name, / punctuation stomach / meridian name meridian / noun. / punctuation.

[0113] The adaptive closed-loop feedback optimization module 4 evaluates the annotation quality. For the entity recognition of "Licorice", the confidence is 0.97; for the annotation of the meridian names of "heart, lung, spleen, stomach", the confidence exceeds 0.90. For the词性标注 of the character "belong", the initial confidence in predicting it as a preposition is 0.72, slightly lower than the threshold. The module generates a feedback signal, combines the semantics of the traditional Chinese medicine term "meridian tropism", and after adjusting the parameters, confirms that "belong" is a verb meaning "belong to a certain meridian", and the confidence is increased to 0.88.

[0114] As can be seen from the above embodiments, the system of the present invention can accurately capture the historical semantic evolution of ancient Chinese words, deeply integrate knowledge in the field of traditional Chinese medicine, achieve accurate word meaning disambiguation and integrated annotation based on multi-granularity collaborative semantic understanding, and continuously optimize the annotation quality through an adaptive closed-loop feedback mechanism, which significantly improves the level of automated processing of ancient Chinese, especially ancient Chinese medicine literature.

[0115] It should be understood that the above embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Various modifications and variations can be made to the present invention by those skilled in the art without departing from the spirit and scope of the invention, and all such modifications and variations should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A deep learning-based automatic annotation and semantic analysis system for classical Chinese documents, characterized in that, include: The cross-era semantic evolution tracking module is used to construct a temporal evolution map of ancient Chinese corpus containing different historical periods. It uses a temporal graph neural network to extract semantic evolution feature vectors of words in different dynasties and generate evolutionary semantic representations containing the time dimension. The TCM knowledge fusion and enhancement module is used to construct a TCM knowledge graph, including prescription entities, medicinal material entities, disease entities and their relationships. Through graph attention network, the entity embedding vectors in the knowledge graph are fused with text semantic features across modalities to generate an enhanced semantic representation that integrates TCM knowledge. The multi-granularity cross-attention disambiguation module is used to receive the evolved semantic representation and the enhanced semantic representation, calculate self-attention weights and cross-attention weights at three granularities: character level, word level and sentence level, capture semantic dependencies between different granularities through a multi-head attention mechanism, perform context-aware word semantic disambiguation for the polysemy of a word in Classical Chinese, and generate disambiguated annotation results. An adaptive closed-loop feedback optimization module is used to calculate the confidence score of the annotation results, judge the annotation quality based on the confidence threshold, and generate a feedback adjustment signal if the confidence score is lower than the preset threshold. The feedback adjustment signal dynamically adjusts the evolution weight parameters of the cross-generational semantic evolution tracking module and the attention weight parameters of the multi-granularity cross-attention disambiguation module through the gradient backpropagation mechanism, forming a closed-loop optimization mechanism driven by annotation quality.

2. The system according to claim 1, characterized in that, The cross-generational semantic evolution tracking module also includes: The temporal corpus construction unit is used to divide the ancient Chinese corpus into six time periods according to the dynasty: pre-Qin, Qin and Han, Wei, Jin and Northern and Southern Dynasties, Sui, Tang and Five Dynasties, Song, Liao, Jin and Yuan, and Ming and Qing, and to construct a temporal semantic node sequence for each word. The evolution graph generation unit is used to construct a cross-era word evolution graph, where nodes represent the semantic state of words in a specific period, and edges represent the semantic evolution relationship between adjacent time periods; The temporal graph convolutional unit is used to apply a temporal graph convolutional network to the evolution graph, aggregate the features of temporal neighborhood nodes and extract the evolution patterns, and generate word representation vectors containing historical evolution information.

3. The system according to claim 1, characterized in that, The module for integrating and enhancing traditional Chinese medicine knowledge also includes: The knowledge graph construction unit is used to extract prescriptions, medicinal materials, diseases, acupoints, and meridian entities from ancient Chinese medicine literature and their treatment, compatibility, and meridian tropism relationships to construct a Chinese medicine knowledge graph containing no less than 20,000 entities and 50,000 relationships. The entity linking unit is used to semantically match the candidate entities identified in the text to be labeled with the standard entities in the knowledge graph, calculate the cosine similarity between the text representation and the entity embedding in the knowledge graph, and determine the entity correspondence. The cross-modal fusion unit is used to calculate the attention weights between entity nodes in the knowledge graph using a graph attention network, and integrates the entity embedding vectors into text features through attention weighting to generate a contextual representation of fused domain knowledge.

4. The system according to claim 1, characterized in that, The multi-granularity cross-attention disambiguation module also includes: Character-level attention units are used to calculate self-attention distributions at the character level, capturing the positional features and word formation patterns of individual characters within words; Word-level attention units are used to calculate cross-word attention weights at the word granularity, and to model the grammatical dependencies and semantic collocation patterns between words. Sentence-level attention units are used to compute global contextual attention at the sentence level, extracting long-distance semantic dependencies and discourse coherence features; The multi-granularity fusion unit is used to adaptively and weightfully fuse the attention features at the character, word, and sentence levels through a gating mechanism, and dynamically adjust the contribution weight of each granularity feature for different semantic disambiguation tasks.

5. The system according to claim 1, characterized in that, The adaptive closed-loop feedback optimization module also includes: The confidence calculation unit is used to calculate the labeled confidence based on the probability distribution output by the multi-granularity cross-attention disambiguation module, and to quantify the prediction uncertainty using the entropy method. A quality assessment unit is used to compare the confidence score with a dynamically adjusted confidence threshold to identify low-confidence labeled samples; The feedback signal generation unit is used to generate a feedback signal containing the gradient adjustment direction and adjustment magnitude for low-confidence samples. The parameter adaptive adjustment unit is used to update the network parameters of the cross-generational semantic evolution tracking module and the multi-granularity cross-attention disambiguation module through backpropagation based on the feedback signal, so as to achieve continuous optimization based on annotation quality.

6. The system according to claim 2, characterized in that, The temporal graph convolutional unit adopts a gated temporal graph convolutional network, which introduces a gating mechanism when aggregating features of adjacent nodes in time. The forget gate controls the degree of retention of historical semantic information, and the update gate determines the absorption ratio of semantic features in the current period.

7. The system according to claim 3, characterized in that, When calculating graph attention weights, the cross-modal fusion unit comprehensively considers the structural importance and semantic relevance of entities in the knowledge graph. It assigns higher attention weights to entities related to traditional Chinese medicine prescriptions and dynamically adjusts the weights of entities related to common medicinal materials based on their compatibility relationships.

8. The system according to claim 4, characterized in that, The multi-granularity fusion unit adopts a hierarchical gating mechanism. First, a first-level fusion is performed between character-level and word-level features. Then, the fusion result is fused with sentence-level features in a second-level fusion. The gating weight of each fusion layer is adaptively determined by the semantic disambiguation difficulty of the current task.

9. The system according to claim 1, characterized in that, The system also includes an annotation output module, which generates an integrated annotation result containing sentence segmentation markers, word boundary markers, part-of-speech tags, and named entity tags, and generates fine-grained entity type annotations for TCM proper nouns, including prescription names, medicinal material names, and disease names.

10. The system according to claim 1, characterized in that, The evolutionary semantic representation of the cross-generational semantic evolution tracking module is dimensionally aligned with the query vector of the multi-granularity cross-attention disambiguation module through a feature mapping layer. The enhanced semantic representation of the TCM knowledge fusion enhancement module enhances the key-value vector of the multi-granularity cross-attention disambiguation module through attention weighting. The feedback signal of the adaptive closed-loop feedback optimization module simultaneously acts on the graph convolutional layer parameters of the cross-generational semantic evolution tracking module and the attention layer parameters of the multi-granularity cross-attention disambiguation module.

Images