Chinese language text semantic association analysis method and system

Abstract

The invention discloses a Chinese language text semantic association analysis method and system. The method comprises the following steps: constructing a historical corpus benchmark reference system of time slices, and establishing a semantic feature library and a syntactic structure feature library containing distribution weight for each time slice; parallel computing is carried out on a dominant semantic age value and a recessive structure age value of the to-be-analyzed text through two channels, and a time sequence deviation between the dominant semantic age value and the recessive structure age value is obtained; and generating a dynamic structure and semantic weight coefficient by using a nonlinear mapping function according to the time sequence deviation, and finally outputting an associated text list based on the mixed similarity. According to the method, high-order structure Hash and an inverse time frequency weighting strategy are combined, adaptive switching of retrieval weights is driven by time sequence deviation, the problem that time sequence dislocation texts such as antique texts and counterfeit texts are difficult to recognize in traditional semantic matching is effectively solved, and the accuracy of cross-time text association analysis is remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of natural language processing and information retrieval technology, specifically to a method and system for semantic association analysis of Chinese text. Background Technology

[0002] With the development of digital humanities and intelligent information processing technologies, the correlation analysis of massive historical documents and modern texts has become an important requirement for academic research and copyright protection. Existing text correlation analysis techniques mainly rely on vector space models or semantic representation models based on deep learning (such as Word2Vec and BERT), which calculate similarity by mapping text to high-dimensional semantic vectors. This semantic-centric matching mechanism performs well in handling topic classification, sentiment analysis, or conventional information retrieval tasks, accurately capturing the relevance of texts at the content level.

[0003] However, in specific scenarios involving cross-era text retrieval, ancient book collation, and the identification of forgeries, analysis relying solely on semantic dimensions has significant limitations. The evolution of Chinese texts is not only reflected in the changes in lexical semantics but also profoundly in the transformation of syntactic structures and sentence-building habits. Existing semantic analysis models tend to focus on the co-occurrence and semantic distance of content words, often neglecting the syntactic topology hidden beneath the lexical surface. When faced with texts exhibiting a specific style of imitation, such as forgeries written by modern people imitating classical Chinese or collection-sentence imitations, these texts may highly overlap with authentic historical documents in terms of vocabulary usage, making them extremely difficult to distinguish in the semantic vector space.

[0004] Existing technical solutions typically fail to effectively separate and independently quantify the "content era" and "structural era" of text. When there is a temporal misalignment between the semantic content and syntactic structure of text, traditional methods, lacking independent modeling of syntactic evolution and adaptive weight adjustment mechanisms based on temporal deviations, are often misled by superficial semantic similarity. They cannot accurately identify related texts that are similar in content but have vastly different writing styles, or similar in style but different in content. This results in the system's precision and robustness failing to meet the high standards of practical applications in complex retrieval tasks that require in-depth consideration of writing style and era characteristics. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for semantic association analysis of Chinese texts. This solves the problem that existing technologies, when processing cross-era text association analysis, rely solely on semantic keyword matching while ignoring the evolution of the text's inherent syntactic structure, thus failing to effectively identify texts with temporal misalignment characteristics (such as antiquarianism, forgeries, or imitations of specific writing styles).

[0006] The method mainly includes the following processing steps: constructing a historical corpus benchmark reference system based on time slices, and establishing a semantic feature library and a syntactic structure feature library containing distributed weights for each time slice; receiving the text to be analyzed, and calculating the explicit semantic age value and implicit structural age value of the text through semantic analysis path and structural analysis path respectively; then calculating the absolute value of the difference between the explicit semantic age value and the implicit structural age value to obtain the time series deviation value; based on the time series deviation value, generating dynamic structural weight coefficients and semantic weight coefficients using a preset nonlinear mapping function; finally, calculating the mixed similarity between the text to be analyzed and the candidate texts based on these weight coefficients, and outputting a list of related texts according to the similarity ranking.

[0007] In the above technical solution, a high-order structure hashing strategy is adopted to construct the syntactic structure feature library. Specifically, dependency parsing and delexicalization are performed on the text in the corpus, retaining only the part-of-speech tags of nodes and the dependency relation tags of edges to construct an abstract syntactic graph. Subsequently, an iterative aggregation mechanism is used to process the abstract syntactic graph. Through multiple iterations, the neighborhood structure information of each node is aggregated and encoded into a unique hash identifier. This processing method can transform a variable-length graph structure into a fixed-length set of syntactic fingerprints, enabling the system to capture deep syntactic topological features across the lexical surface in the text, achieving a digital representation of the text's structural style.

[0008] Furthermore, to accurately quantify the temporal characteristics of syntactic structures, this invention employs an inverse temporal frequency weighting strategy to determine distribution weights. The system dynamically assigns weights based on the prevalence of each syntactic subtree's hash identifier across different time slices. For common syntactic structures appearing in the vast majority of time slices, their weights are reduced through a logarithmic penalty mechanism to suppress noise; for specific syntactic structures appearing only in specific time slice sets, they are assigned high weights. This mechanism enhances the structural fingerprint's ability to identify specific historical periods, making the analysis results more reflective of the text's temporal uniqueness.

[0009] In the two-dimensional date feature extraction stage, semantic centroid projection and probabilistic sharpening mechanisms were used to calculate the explicit semantic date value. This step maps the text to be analyzed into document vectors in the semantic vector space, calculates the similarity between the document vectors and the reference centroids of each historical time slice, and introduces a normalized exponential function with a temperature coefficient to perform a nonlinear transformation on the similarity distribution. Adjusting the temperature coefficient can concentrate the probability distribution towards the time slice with the highest similarity, thereby enhancing the significance of the text's date attribution determination and reducing the ambiguity of time judgment while preserving distribution information.

[0010] Meanwhile, a Bayesian smoothing correction mechanism is introduced to address the data sparsity problem in short texts when calculating the latent structure age value. This mechanism incorporates a global historical period prior distribution as a regularization term and dynamically adjusts the confidence weights of the observation probability and the prior distribution based on the total number of valid structural fingerprints in the text to be analyzed. When the text is short or there are few valid fingerprints, the weight of the prior distribution is increased, so that the final posterior probability of the structure converges to the global distribution, thereby avoiding zero-probability misjudgment due to insufficient samples and ensuring the robustness of the age determination.

[0011] To determine temporal discrepancies in text, this method establishes a dynamic judgment threshold through a statistical process control mechanism. By statistically analyzing the natural deviation distribution of explicit and implicit date values ​​in native texts from a benchmark corpus, a confidence interval is constructed using the statistical mean and standard deviation. The upper limit of this confidence interval is set as the judgment threshold. Only when the temporal discrepancy value of the text to be analyzed exceeds this statistical confidence interval is the text considered to have significant temporal dissonance.

[0012] Based on the above criteria, the dynamic weight generation process employs a continuous nonlinear response mechanism based on an S-shaped curve. This mechanism constructs a soft-switching control logic based on temporal deviation values: when the temporal deviation value is within the judgment threshold range, it indicates that the text's temporal characteristics are consistent, and the system maintains the dominant position of the semantic weight coefficient, focusing on content matching; when the temporal deviation value exceeds the judgment threshold, it indicates that the text has temporal misalignment, and as the deviation increases, the structural weight coefficient is controlled to rise rapidly along an S-shaped curve and tend to saturate. This mechanism realizes the automatic switching of the retrieval focus from semantic content to syntactic structure, enabling more accurate capture of related texts with similar writing styles.

[0013] In the final association retrieval stage, hybrid similarity is measured based on feature-weighted overlap. This measurement method comprehensively considers the cosine similarity between the text to be analyzed and the candidate texts in the semantic vector space and the set overlap in the syntactic structure space. When calculating structural overlap, the aforementioned distributed weights are used to weight each common syntactic hash identifier, ensuring that the retrieval results preferentially match those sparse syntactic structures with high discriminative power. In addition, the retrieval process is configured with an adaptive dual-path indexing strategy, pre-constructing a first index based on lexical inversion and a second index based on syntactic hash fingerprint inversion. The system dynamically adjusts the retrieval path according to the real-time generated structural weight coefficients: when the structural weight coefficients are low, the first index is mainly used for recall; when the structural weight coefficients increase significantly, the traversal depth of the second index is activated and deepened, thereby efficiently capturing related texts with similar syntactic styles but different vocabulary.

[0014] A second aspect of this invention provides a semantic association analysis system for Chinese language text.

[0015] The system comprises a benchmark construction unit, a feature extraction unit, and an association analysis unit. The benchmark construction unit is configured to build a benchmark reference system including a time-segmented semantic mapping module and a high-order structure hashing module. The feature extraction unit is configured to calculate explicit semantic age values ​​through a semantic centroid projection module and implicit structural age values ​​through a structural fingerprint matching module and a Bayesian smoothing correction module. The association analysis unit is configured to obtain temporal deviation through a temporal deviation calculation module, determine weight coefficients through a dynamic weight generation module, and output a list of associated texts based on the weight coefficients through a hybrid similarity measurement module.

[0016] This invention, through the aforementioned technical solution, effectively solves the problem that traditional methods cannot distinguish between the surface semantic content and deep structural style of text by comparing the dual temporal characteristics of semantics and structure. In particular, by introducing a dynamic weighting mechanism driven by temporal deviation, the system can adaptively process normal text and text with temporal discrepancies, significantly improving the accuracy of complex text association analysis such as cross-era imitations and forgeries while maintaining the accuracy of conventional retrieval.

[0017] This invention provides a method and system for semantic association analysis of Chinese text. It has the following beneficial effects: 1. This invention effectively solves the problem of traditional methods' difficulty in identifying texts with temporal misalignment by constructing a dual-channel dating feature analysis mechanism based on semantics and structure. Existing technologies often rely on lexical co-occurrence, which is easily misled by surface semantics. This invention calculates the deviation between explicit semantic dating values ​​and implicit structural dating values, and dynamically adjusts the retrieval weights based on this using a nonlinear function. When a significant temporal deviation is detected, the system automatically reduces the weight of semantic matching and strengthens the matching of syntactic structure and style. This mechanism enables the system to penetrate surface lexical differences and accurately locate related texts with deep structural homology when processing archaic documents, forgeries, or texts imitating specific writing styles.

[0018] 2. This invention employs a high-order structure hashing and inverse-era frequency weighting strategy to achieve accurate quantification and noise suppression of text syntactic styles. Through delexical dependency parsing and iterative hash encoding, the variable-length syntactic tree is transformed into a fixed-length fingerprint, thereby capturing the topological features of the text without relying on specific keywords. Combined with inverse-era frequency weighting, the system can automatically reduce interference from common syntactic structures such as subject-verb-object and significantly increase the weight of sentence structures specific to particular historical periods. This makes the retrieval results more focused on core syntactic patterns that reflect the characteristics of an era, improving the signal-to-noise ratio for identifying specific writing styles.

[0019] 3. The Bayesian smoothing correction and adaptive indexing mechanism introduced in this invention significantly enhance the robustness and efficiency of the system in short text analysis and large-scale retrieval scenarios. To address the zero-probability misjudgment problem that may arise from the sparsity of structural features in short texts, Bayesian smoothing introduces a prior distribution for regularization correction, ensuring the stability of date determination. Simultaneously, the adaptive dual-path indexing strategy based on dynamic weights prioritizes the use of low-cost lexical indexes during regular retrievals, activating high-cost structural indexes only when deep structural analysis is required. This on-demand approach effectively reduces system computational overhead and response latency while maintaining high recall. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall architecture of the Chinese language text association analysis system of the present invention; Figure 2 This is a schematic diagram of the internal structure of the reference building unit of the present invention; Figure 3 This is a logical architecture diagram of the feature extraction unit of the present invention; Figure 4 This is a logical architecture diagram of the correlation analysis unit of the present invention; Figure 5 This is a schematic diagram comparing the metachronization error (MAE) of the present invention and the benchmark model under different text lengths; Figure 6 This is a schematic diagram of the accuracy confusion matrix of the present invention in a cross-generational retrieval task; Figure 7 This is a comparative ablation experiment showing the impact of the inverse era frequency weight (IEF) of this invention on feature discrimination. Figure 8 This is a schematic diagram of the S-shaped response curve of the dynamic weight generation function in this invention.

[0021] Among them, 100 is the Chinese language text association analysis system; 110 is the benchmark construction unit; 111 is the time-segmented semantic vector module; 112 is the high-order structure hash module; 113 is the time-series distribution weighting module; 120 is the feature extraction unit; 121 is the semantic centroid projection module; 122 is the semantic centroid projection module; 123 is the Bayesian smoothing correction module; 130 is the association analysis unit; 131 is the time-series deviation calculation module; 132 is the dynamic weight generation module; and 133 is the semantic centroid projection module. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] See attached document Figure 1 , Figure 1 This is a schematic diagram of the overall architecture of a Chinese language text association analysis system 100 according to an embodiment of the present invention. The present invention provides a Chinese language text association analysis system 100, which runs on a computing device including processing and storage hardware, and mainly consists of three core logical units: a baseline construction unit 110, a feature extraction unit 120, and an association analysis unit 130. The units are connected through data interfaces to collaboratively complete text analysis tasks.

[0024] The baseline construction unit 110 is used to process historical corpus data and establish an analysis baseline in the offline stage. This unit includes a time-segmented semantic vector module 111 and a high-order structure hashing module 112. The time-segmented semantic vector module 111 is used to divide the historical timeline into discrete time slices, train word vectors for different time slices for content words, and establish the mapping relationship between content words, time, and vectors. The high-order structure hashing module 112 is used to perform dependency parsing on the corpus, extract delexicalized syntactic subtree fingerprints, statistically analyze the distribution probability of each fingerprint in different time slices, and calculate the inverse epoch frequency weight of each fingerprint to suppress common sentence pattern interference.

[0025] The feature extraction unit 120 is used to process the input text and quantify its temporal features during the online phase. This unit includes a semantic centroid projection module 121 and a structural fingerprint matching module 122. The semantic centroid projection module 121 calculates the explicit semantic age value based on the distribution of content words in the input text and the data from the time-segmented semantic vector module 111. The structural fingerprint matching module 122 calculates the implicit structural age value based on the syntactic fingerprint of the input text and the data from the higher-order structure hash module 112, combined with a Bayesian smoothing algorithm to correct for short text bias.

[0026] The association analysis unit 130 is used to adjust strategies based on temporal feature differences and output results. This unit includes a temporal deviation calculation module 131, a dynamic weight generation module 132, and a hybrid similarity measurement module 133. The temporal deviation calculation module 131 calculates the deviation between the explicit semantic age value and the implicit structural age value to obtain the temporal detuning rate. The dynamic weight generation module 132 generates dynamic weight coefficients based on the temporal detuning rate, increasing the weight of structural features when the temporal detuning rate exceeds a threshold. The hybrid similarity measurement module 133 combines semantic similarity with structural similarity that applies inverse epochal frequency weights, uses the dynamic weight coefficients to calculate the final association score, and outputs a list of associated texts.

[0027] See attached document Figure 2 , Figure 2 This is a flowchart of a Chinese language text association analysis method according to an embodiment of the present invention. It includes the following steps: S10, construct a two-dimensional historical benchmark library; the benchmark construction unit 110 processes the historical corpus with time attributes, establishes a semantic vector space containing the time dimension through the time-segmented semantic vector module 111, and establishes a syntactic topology graph containing inverse era frequency weights through the high-order structure hash module 112. S20, extract the dual-channel temporal features of the input text; the feature extraction unit 120 receives the text to be analyzed, calculates the explicit semantic age value of the text through the semantic centroid projection module 121, and calculates the implicit structural age value of the text through the structural fingerprint matching module 122 combined with the Bayesian smoothing mechanism. S30, calculate the semantic-structural temporal misalignment rate; the temporal deviation calculation module 131 calculates the absolute value of the difference between the explicit semantic age value and the implicit structural age value, and obtains the temporal misalignment rate that represents the degree of temporal misalignment between content and form. S40, Generate adaptive dynamic weights; The dynamic weight generation module 132 generates a weight adjustment coefficient based on the temporal detuning rate through a nonlinear function. When the temporal detuning rate is greater than a preset threshold, the coefficient causes the association calculation strategy to tilt towards structural similarity. S50, calculate the hybrid association score and output the result; the hybrid similarity measurement module 133 calculates the semantic similarity and weighted structural similarity between the target text and the candidate text, uses the weight adjustment coefficient to fuse the two to obtain the final score, and outputs the list of associated texts based on the score.

[0028] In step S10, which involves constructing a two-dimensional diachronic benchmark library, the system accesses a historical corpus containing temporal metadata. This corpus includes, but is not limited to, digitized historical texts, collected works from various dynasties, or newspaper and periodical documents with explicit dates. This step aims to establish a semantic and structural reference system for subsequent comparative analysis, and is implemented through the following sub-steps: S101, construct the diachronic semantic vector space.

[0029] To address the semantic shift caused by language evolution over time, this step employs a strategy of independent training across different time periods. The system divides the timeline of the historical corpus into a set of discrete time slices. ,in The total number of slices is defined, and the granularity of the segmentation can be set according to the corpus density, such as by dynasty, century, or fixed year interval. This applies to content words in the corpus that appear more frequently than a preset threshold (e.g., more than 50 times or in the top 5000 words). The system utilizes each time slice. The text data within the slice is used to train the corresponding center word vectors. Specifically, for each time slice... Initialize an independent word embedding model (such as the Skip-gram architecture of Word2Vec). The model's input is the segmented text sequence within the given time slice. The training objective is to maximize the log-likelihood probability of surrounding words within a given center word's predicted context window (typically 5 to 10 words in size). After training, extract the model's hidden layer weight matrix as the semantic vector library for that time slice. Through this process, a mapping function is established. ,in content words In time slices Below 3D vector representation.

[0030] S102, extract higher-order syntactic topological features.

[0031] The system performs dependency parsing on text in the corpus, generating a dependency parsing tree containing part-of-speech (POS) tags and dependency relation tags. For the specific implementation of dependency parsing, those skilled in the art can use transfer-based or graph-based dependency parsing algorithms, which are well-known technologies in the field and will not be elaborated here. To shield the structural analysis from interference from specific text content, the system performs delexicalization on the syntactic tree, replacing specific words on tree nodes with corresponding POS tags, retaining only the tree's topology, node POS tags, and dependency relation types on edges. Subsequently, the system uses the Weisfeiler-Lehman (WL) graph kernel algorithm to extract syntactic subtree features. This algorithm aggregates node neighborhood information iteratively: in the... In the next iteration (1) Iteration depth Typically, the value is 2 to 3. For each node, its current label is sorted and concatenated with the labels of all its neighbors to form a new multiset signature string; then, a hash function is used to map this signature string to a new discrete label. After each iteration, the final label of each node in the tree represents the set of elements centered at that node. A higher-order syntactic subtree fingerprint of skip-range dependency structures. All collected fingerprints are denoted as a set. .

[0032] S103, calculate the inverse age frequency weight (IEF).

[0033] To quantify the discriminative power of each syntactic subtree fingerprint in the time dimension, the system calculates the fingerprint for each fingerprint. The IEF weights are designed based on the IDF concept in information retrieval, which states that the more frequently a syntactic structure is used across dynasties, the lower its information entropy for dating analysis, and therefore the smaller its weight should be; conversely, if it appears only in a few dynasties, it has high discriminative power. The system statistically analyzes each fingerprint. Slices at different times Normalized frequency of occurrence, i.e., conditional probability. Based on this probability distribution, the reverse-era frequency weight is calculated. The calculation formula is as follows: ; In the formula, This indicates the total number of time slices. Representation of syntactic subtree fingerprints Appearing in time slices The conditional probability; This is an indicator function; it takes the value 1 when the condition within the parentheses is met, and 0 otherwise. To determine the probability threshold of the syntactic structure's activity within a certain time slice, this threshold is used to filter out occasional noise. Its value is typically set between [0.01, 0.05], or set to... The dynamic value; the 1 in the denominator is a smoothing term to prevent the denominator from being zero. Using this formula, for common syntactic structures widely used in all dynasties, the denominator value is large, leading to a calculated... The value approaches 0; while for syntactic structures used only in a few specific dynasties, its The value is relatively large, thus it dominates in subsequent similarity calculations.

[0034] S104, Construct a syntax-temporal probability mapping table.

[0035] The system will calculate the conditional probability distribution. and the corresponding reverse-era frequency weight The data is stored in a key-value pair index structure. This mapping table records the distribution pattern of each syntactic topology throughout history and its confidence level as a basis for dating, forming a structural reference system for determining the age of implicit structures. For specific storage implementation, efficient key-value databases such as Redis or RocksDB can be used for persistent storage to support millisecond-level queries during subsequent online analysis.

[0036] In step S20, which extracts the dual-channel temporal features of the input text, the system receives the target text to be analyzed. The system initiates parallel processing of explicit semantic channels and implicit structural channels. This step is based on the asynchronous nature of text semantic representation and syntactic topology in the diachronic evolution dimension; that is, the semantic content (what is said) and syntactic structure (how it is said) of a text often have different evolution rates over time. The system quantifies the chronological characteristics of these two dimensions separately, providing a quantitative basis for subsequent identification of temporal discrepancies. This step utilizes the baseline data constructed in step S10 and introduces a Bayesian smoothing mechanism to correct the bias caused by the sparsity of short text features. Specifically, it is implemented through the following sub-steps: S201, calculate the explicit semantic age value.

[0037] The system first processes the target text Preprocessing is performed, including Chinese word segmentation and stop word filtering, to obtain a set of effective content words. To aggregate discrete lexical features into document-level semantic representations, the system computes text... Document center vector This vector is composed of the set All content words in the text are sliced ​​at a specific time. The arithmetic mean of the word vectors is obtained. Then, the document center vector is calculated. With this time slice Global average semantic vector Cosine similarity between To transform the similarity scores into a probability distribution, the system uses the Softmax function for normalization and calculates the text similarity scores. Belonging to various time slices semantic posterior probability : ; In the formula, represents the cosine similarity value, ranging from [-1, 1]; The scaling factor is used to adjust the sharpness of the probability distribution, typically ranging from [5, 20]. A larger value results in a more concentrated probability distribution around the time slices with the highest similarity. Based on this probability distribution, the explicit semantic date value is calculated. The calculation formula is as follows: ; In the formula, This represents the total number of time slices. Time slice The corresponding numerical center year (e.g., if) Representing the Tang Dynasty, The possible value is 750 AD. As a scalar, it intuitively reflects the central position of the words used in the text and their meanings on the historical timeline.

[0038] S202, Calculate the implicit structure period value. This sub-step aims to mine the syntactic topological features of the text, i.e., the period to which the syntactic style belongs. The system processes the target text... Perform the same dependency parsing and WL graph kernel hash extraction operations as in step S102 to obtain the set of syntactic subtree fingerprints of the target text. ,in The system queries the syntax-temporal probability mapping table generated in step 5104 to obtain each fingerprint. conditional probability and its reverse-era frequency weight .

[0039] S203, perform Bayesian smoothing correction.

[0040] Since the input text may be short (such as a five-character quatrain), direct statistical analysis may lead to data sparsity or overfitting (e.g., misjudging the era based on a single, accidental sentence structure). Bayesian smoothing works by introducing prior knowledge to weigh the observed data. When the amount of observed data (i.e., the text length) is insufficient, the system relies more on the prior distribution; as the amount of data increases, the result gradually approximates the actual observed distribution. In practice, the unsmoothed observation probability distribution is first calculated. ; Subsequently, a global prior distribution is introduced. and smoothing hyperparameters Calculate the corrected structural posterior probability. : ; In the formula, The effective structural node length of the text, i.e., the size of the fingerprint set. This variable serves as a confidence factor; the longer the text, the greater the weight of the observation. The prior probability of time slices is usually set as the natural or uniform distribution of the amount of text in each dynasty in the corpus (i.e., The smoothing hyperparameter, also known as the pseudo-count, is used to control the influence of the prior distribution. Its typical value range is [10, 50]. The larger the value, the more conservative the system's judgment of short texts tends to be (tending towards the prior distribution), thereby avoiding extreme misjudgments.

[0041] S204, Generating the final structural age value based on a smoothed and corrected probability distribution. Calculate the age value of hidden structures : ; In the formula, The definition is consistent with that in S201. The result obtained through the above calculations... Eliminating interference from common sentence structures (by) It guarantees and overcomes the data noise of short texts (guaranteed by Bayesian smoothing), and can accurately characterize the temporal attributes of text in syntactic topology.

[0042] In step S30, which calculates the semantic-structural temporal misalignment rate, the system performs a temporal consistency check based on the explicit semantic age value and implicit structural age value output in step S20. The underlying technical principle of this step lies in the asynchronous nature of language evolution: in the history of the Chinese language, the replacement of content words and the evolution of their meanings (semantic dimension) are usually sensitive to social changes and have a high temporal rate of change; while syntactic topology (such as inverted structures and specific collocation patterns of function words) is constrained by the inertia of language habits and has a strong evolutionary lag and stability. Therefore, by quantifying the deviation between these two dimensions, it is possible to reveal whether there is a temporal misalignment between the "content" and "form" of the text. This is specifically implemented through the following sub-steps: S301, Perform timing deviation calculation.

[0043] The system receives the target text. explicit semantic date value and latent structure age values Both values ​​are continuous time scalars calculated based on probability distributions, rather than discrete dynasty labels. This ensures that the calculation results retain fine-grained temporal characteristics. The system calculates the absolute value of the difference between the two values, which is the time-series detuning rate. The calculation formula is as follows: ; In the formula, This is the expected value of semantic age inferred based on word distribution and vector space mapping; The expected value of the structural age is inferred based on the syntactic subtree fingerprint and the inverse age frequency weight; This represents absolute value operations. The physical meaning of this metric lies in measuring the projection distance of text features in the semantic-structural two-dimensional time space. This metric is a non-negative real number, and its magnitude directly maps to the stylistic properties of the text: a lower value indicates a higher stylistic property. The value represents the original temporal text, that is, the author's expression using the natural syntax of their era; higher values ​​indicate a more natural or chronological text. The value represents a text with temporal disharmony, which usually corresponds to antiquarian creation, collection of sentences, or cross-era literary style inheritance. That is, the text reflects the physical creation era in terms of content, while retaining the early era characteristics of the imitated object in terms of syntactic structure.

[0044] S302, Determine the detuning threshold.

[0045] In order to make continuous To convert this into a control signal for weight adjustment in subsequent step S40, the system needs to determine a reference detuning threshold. To meet the requirement of sufficient disclosure, this threshold... The data is not arbitrarily assigned, but rather calculated through statistical analysis of a benchmark corpus. In practice, the system selects validation set samples known as "native texts" from the benchmark corpus (i.e., regular historical texts confirmed to be neither archaic nor forged), and calculates the semantic-structural age difference set of all samples in this validation set. Subsequently, the system calculates the average value of this difference set. and standard deviation Detuning threshold The calculation formula is set as follows: ; In the formula, is the confidence coefficient, used to control the system's tolerance to normal language evolution fluctuations. Typically, it is taken as... The range [2.0, 3.0] corresponds to a 95% to 99% confidence interval in statistics. Taking ancient Chinese literary corpus as an example, statistical calculations show that this threshold... The typical value range is between [50, 100] years. When the input text... When this happens, the system determines that the text has experienced significant temporal disharmony, thereby triggering a subsequent dynamic weight adjustment mechanism.

[0046] In step S40, which generates adaptive dynamic weights, the system calculates the timing detuning rate based on the results of step S30. The system dynamically generates a weight allocation scheme for subsequent association calculations through a nonlinear mapping mechanism. This step is based on the "signal-to-noise ratio modulation" principle in signal processing: for temporally detuned texts (such as later imitations of ancient works), although their explicit semantic features reflect the physical creation time, for the task of finding their stylistic origins, these modern semantic words constitute interference noise. Conversely, their syntactic topology retains the characteristics of the era being imitated, belonging to high-value signals. To optimize the retrieval signal-to-noise ratio, the system uses the temporal detuning rate as a control variable to dynamically adjust the contribution ratio of semantic and structural features in similarity calculation. This is specifically implemented through the following sub-steps: S401, Construct a nonlinear weight adjustment model.

[0047] The system pre-sets a weight adjustment function. Used to convert continuous timing detuning rates Weight coefficients mapped to structural similarity To achieve a smooth transition in weight allocation from "semantic-driven" to "structure-driven" and to prevent drastic oscillations in the retrieval strategy caused by small fluctuations in the detuning rate near the threshold, the function is configured as a smooth switching model based on an S-curve. This model design ensures that the system maintains weight stability in the range far from the threshold, while providing continuous and differentiable gradient changes in the transition range near the threshold.

[0048] S402, Calculate the structure weight coefficients. The system uses a preset generalized logistic function to calculate the structure weight coefficients corresponding to the target text. The calculation formula is as follows: ; In the formula, The target text timing detuning rate output in step S301; The detuning determination threshold determined in step S303 defines the center point of weight switching; The preset lower limit value for the structural weight is set to [0.3, 0.4]. The basis for setting this parameter is that even for raw text, the syntactic structure still provides about 3096 to 40% of auxiliary discrimination information, which cannot be completely ignored. The upper limit of the structural weight is set to [0.8, 0.9]. The reason for setting this parameter to less than 1.0 is that even for heavily archaic texts, retaining 10% to 20% of the semantic weight helps to distinguish different themes under the same text style and prevents the matching results from being completely unrelated to the content. The sensitivity coefficient is used to adjust the width of the weight transition interval, and its value range is usually set to [0.1, 0.5]. The larger the value, the steeper the S-curve, and the more violently the system reacts to detuning phenomena exceeding the threshold, which is similar to on / off control; the smaller the value, the flatter the curve, and the system tends to maintain the gradual change characteristics of the mixed weights over a larger range.

[0049] S403, generate complementary semantic weights.

[0050] Following the principle of weight normalization, the system synchronously generates weight coefficients for semantic similarity. The calculation formula is as follows: ; Based on the above calculations, the system generates a set of texts targeting the current target. Adaptive weight pair This mechanism enables the system to adaptively perceive and adjust policies based on text attributes: when the input... much smaller hour, converges to The system executes a conventional retrieval strategy based primarily on semantic matching; when Significantly greater than hour, Rapidly climbed and approached The system automatically switches to a source retrieval strategy based on syntactic topology matching, thereby achieving adaptive processing of native and archaic texts without manual intervention.

[0051] In this embodiment, in step S50, which calculates the hybrid association score and outputs the result, the system performs multi-dimensional similarity calculation and fusion between the target text and the texts in the candidate corpus based on the adaptive weighting scheme generated in step S40. This process is mainly based on the technical principle of multi-view fusion ranking, which treats the text as a composite object composed of a "semantic content view" and a "syntactic structure view". Since these two views have different dimensional definitions and measurement scales in the vector space, directly concatenating the vectors will lead to the curse of dimensionality or feature submersion. Therefore, this embodiment adopts a decision-level fusion strategy, first calculating the single-view similarity separately, and then using dynamic weights to synthesize them, thereby outputting a comprehensive evaluation index that combines content relevance and stylistic inheritance. Specifically, the following processing steps are included: S501, calculate global semantic similarity.

[0052] The system calls the historical semantic vector space constructed in step S101 to calculate the target text. With candidate text set Each candidate text The semantic similarity is considered. To eliminate the influence of text length differences and highlight the role of core vocabulary, this embodiment preferably adopts a document vector synthesis method based on TF-DF weighting. Specifically, for each content word in the text... Obtain its word vectors and the corresponding TF-IDF weights Document semantic vectors are generated by weighted summation and normalization. Subsequently, the cosine similarity algorithm was used to calculate the target text vector. With candidate text vectors semantic similarity between The calculation formula is as follows: ; In the formula, · represents the vector dot product operation. This represents the Euclidean norm of a vector. The value ranges from [0,1] (negative values ​​are truncated to 0). This value is used to quantify the degree of overlap between two texts in terms of the topic, object, and specific content being discussed, without regard to their expression.

[0053] S502, calculate the weighted structural similarity.

[0054] This sub-step is a crucial step in implementing "Inverse Era Frequency Weighting (IEF)," aiming to evaluate the homology of texts at the syntactic topological level. Unlike the dense nature of semantic vectors, syntactic fingerprint features belong to high-dimensional sparse data; therefore, this embodiment employs a generalized Jaccard similarity algorithm suitable for sparse set comparisons. For the target text... and candidate text The system constructs sparse feature vectors respectively. and Each dimension of the vector corresponds to a syntactic subtree fingerprint existing in the entire database. Its dimension value The similarity is determined by the product of the term frequency (TF) of the fingerprint in the text and the inverse epochal frequency weight (IEF) calculated in step S103. The calculation formula is as follows: ; In the formula, The union of all non-zero syntactic fingerprints in the two texts; fingerprints In the text Weighted values ​​in The denominator of this formula represents the union weights of the two text structural features, and the numerator represents the intersection weights. By introducing... The formula naturally suppresses the contribution of high-frequency general sentence structures (such as simple subject-predicate structures) to similarity, while significantly amplifying the weight of low-frequency syntactic structures with specific era characteristics (such as special inversions or consecutive use of function words), ensuring that high scores can accurately reflect the inheritance relationship of literary style.

[0055] S503, Synthetic Mixed Association Score.

[0056] The system receives the adaptive weight pair output in step S403. The semantic similarity and structural similarity calculated above are linearly weighted and fused to generate the final hybrid association score. The calculation formula is as follows: ; During this calculation process, the system dynamically reconstructs the retrieval strategy based on the temporal attributes of the text itself: if the target text is determined to be severely out of tune (i.e., (larger), the final score will be mainly based on structural similarity. The decision prioritizes displaying texts with similar syntactic styles (such as original classical Chinese texts) in the results list; if the target text is a native time-series text, the score will be primarily based on semantic similarity. The decision was made to prioritize displaying contemporary texts related to the content.

[0057] S504 generates a list of associations and attribute labels. The system scores based on mixed associations. Sort the candidate text set in descending order and extract the top texts with scores higher than a preset output threshold (e.g., 0.5). Each text is used as a list of associated results. Simultaneously, the system combines the timing detuning rate calculated in step S30. Generate attribute labels for the target text. To explain the source of the high similarity to the user, the system executes a key structure backtracking algorithm: for structural similarity... For higher-quality matching pairs, the system calculates a common syntactic fingerprint for each pair. Contribution to similarity score That is, in the numerator term The system will select the top contributors. Each syntax subtree fingerprint is labeled as a "key feature". Ultimately, if... The system adds an "archaic / traditional" label to the output and highlights the text fragments corresponding to the key syntactic subtrees in the visualization interface using different colors; if The system adds "native / contemporary" tags. This structured data is output through API interfaces or data buses for users to conduct historical research or textual source analysis.

[0058] See attached document Figure 2 , Figure 2This is a schematic diagram of the structure of a benchmark construction unit 110 according to an embodiment of the present invention. As a core basic component of the system, the benchmark construction unit 110 is configured to access a historical corpus with precise chronological metadata (such as the *Complete Tang Poems*, *Complete Song Lyrics*, or a collection of historical materials with chronological information), and accordingly establish a vector reference system for explicit semantic comparison and a topological reference system for implicit structural comparison. At the hardware level, this unit is supported by a high-performance computing cluster or dedicated server equipped with GPU accelerator cards, and logically includes a time-division semantic mapping module 111, a high-order structure hashing module 112, and a time-series weighted distribution module 113.

[0059] The time-segmented semantic mapping module 111 is configured to perform the task of constructing a diachronic vector space. This embodiment is based on the distributional semantics hypothesis, namely, "words with similar contexts also have similar semantics," mapping discrete symbols to a continuous vector space. The module receives historical text data via a data bus and uses an internally integrated time-series slicer to divide the corpus into discrete time series. For each time slice The module instantiates an independent neural network embedding engine, which preferably adopts a Skip-gram architecture. The model's internal structure consists of three layers: the input layer receives one-hot encoded center word vectors, the dimension of which is the vocabulary size for that time slice. The projection layer does not contain an activation function and is used to compress high-dimensional sparse vectors into low-dimensional dense vectors. The range is typically set to [100, 300]. The output layer uses Hierarchical Softmax or negative sampling techniques to predict the probability of co-occurring words within the context window of the center word. During training, the engine optimizes the objective function using stochastic gradient descent to maximize the log-likelihood probability of actual co-occurring word pairs, and finally extracts the weight matrix from the input layer to the projection layer as the semantic feature library for that time slice. To support subsequent millisecond-level retrieval, the semantic mapping module 111 constructs a multi-dimensional inverted index structure. Its storage format is This ensures that, given any words and time points, the system can... Obtain the corresponding semantic vector within the time complexity.

[0060] The high-order structure hashing module 112 connects to the data preprocessing interface and is configured to extract non-semantic skeleton features from the text. This module aims to decouple formal and content features, ensuring that the extracted fingerprints reflect only syntactic style rather than specific meaning. Internally, the module integrates a dependency parser and a graph kernel calculator. The dependency parser employs a transformation-based system, outputting a directed graph structure containing part-of-speech tags and dependency relation tags. To remove content interference, this module performs delexicalization, traversing all nodes of the syntax tree, removing specific lexical text, and retaining only part-of-speech tags (such as "Noun" and "Verb") as node attributes. Subsequently, the graph kernel calculator executes the WL algorithm, which is solidified as an iterative aggregation logic: in the initialization phase, the features of each node are set to the POS tag of that node; in the second phase... In this iteration, for each node Collect all its child nodes Feature tags and dependency labels on the connecting edges Sorted lexicographically and then concatenated into a string sequence Next, a non-cryptographic hash function (such as MurmurHash) is used to hash the string. Mapped to a unique integer ID as the new node label .go through After the nth iteration ( (The preset subtree depth is typically 2 to 3). The module outputs a set of syntactic subtree fingerprints at all different levels contained in the text. .

[0061] The temporal distribution weighting module 113 communicates with the higher-order structure hashing module 112, configuring it for quantifying the temporal discrimination power of syntactic structures. This module maintains a global counter matrix and a probability calculation engine. The global counter is used to accumulate and record each syntactic fingerprint. Slices at various times Frequency of occurrence in Based on accumulated frequency data, the probability calculation engine periodically updates the conditional probability distribution. and reverse-era frequency weight The computational engine performs the following: The logical operations are as follows: ; In the formula, The total number of time slices; This is an indicator function; it takes the value 1 when the condition inside the parentheses is true, and 0 otherwise. This is the activity threshold, used to filter statistical noise and ensure that only cases where the structure was the mainstream usage in a given era are included in the statistics. The value of is usually determined based on the long-tail cutoff point of Zipf's law, and the typical range is

[10] . -4 50 -3 The physical meaning of this formula is that if a syntactic structure appears frequently in all eras (such as a simple subject-predicate structure), its... The value will approach 0; if it is only active in a few eras (such as a specific classical Chinese inversion), its weight will increase significantly. This module ultimately serializes and stores the calculation results in a key-value database, where the key is the hash value of the syntactic fingerprint, and the value contains the weight. And binary objects of probability distribution vectors. This pre-computation and storage structure enables subsequent units to obtain all temporal attributes of the syntactic structure with a single memory query during online analysis, thereby supporting real-time response.

[0062] See attached document Figure 3 , Figure 3 This is a logical architecture diagram of the feature extraction unit 120 according to an embodiment of the present invention. The feature extraction unit 120, as the system's online inference engine, is configured to receive the target text to be analyzed and, based on the reference frame generated by the baseline construction unit 110, decouples the explicit semantic age value and the implicit structural age value in parallel. This unit adopts a dual-channel parallel processing architecture, independently projecting the "lexical semantics" and "syntactic topology" features of the text into the time dimension. Its internal logical processing flow specifically includes the following steps: S121, Perform semantic centroid projection.

[0063] This step, performed by the semantic centroid projection module 121, aims to map the unstructured target text into a diachronic semantic vector space to measure the semantic fit between its content and various historical periods. The module is responsible for analyzing the text. Perform cleaning and word segmentation to filter out stop words that do not contribute semantically, and obtain a sequence of content words. To represent semantic cores at the document level, the module employs a weighted average word vector model. Specifically, for each historical time slice... The semantic index built by the baseline building unit 110 Extracting sequences All content words in the current time slice The document vector is obtained by representing it as a vector and calculating its arithmetic mean. Simultaneously, the module retrieves the pre-calculated reference centroids for each time slice. (That is, the mean of all baseline text vectors during that period). The module then calculates the document vector. With reference centroid cosine similarity To transform cosine similarity (typically between |-1, 1|) into a statistically significant probability distribution and enhance salient features, the module introduces a Softmax function with a temperature coefficient to calculate the semantic posterior probability. : ; In the formula, This is the temperature scaling factor, used to control the entropy value of the probability distribution. The value range is usually set to [5, 20]. When When a larger value is taken, the model acts as a "sharpening" element, forcing the probability quality to concentrate on the time slice with the highest similarity, thereby reducing the ambiguity of time determination; when When smaller values ​​are taken, the distribution tends to be smoother. This mechanism ensures that the model can significantly distinguish the most likely semantic era of the text, rather than outputting an invalid distribution that is uniform.

[0064] S122, calculate the probability of the original structure.

[0065] This step is performed by the structural fingerprint matching module 122, which aims to calculate the original likelihood of the text belonging to different eras at the syntactic structure level. The module performs this step on the target text. Perform dependency parsing and WL subtree hash extraction to generate a set of syntactic fingerprints. Based on the Naive Bayes independence assumption, the module assumes that the choices of various syntactic structures in the text are independent of each other. The module queries the key-value database to obtain each fingerprint. Time slices conditional probability and its reverse-era frequency weight The module calculates the target text in time slices using a weighted summation method. The unnormalized observation scores are then converted into the original weighted observation probabilities. : ; This formula is not directly used as the final result because for short texts, the number of observed syntactic fingerprints is relatively small. Direct statistics are easily affected by random noise caused by data sparsity, leading to zero-probability estimates for some unobserved time slices. Therefore, statistical correction is required. S123, Perform Bayesian smoothing correction. This step is performed by the Bayesian smoothing correction module, configured to improve the model's robustness when processing short texts. The module is based on the Dirichlet smoothing principle and introduces a global prior distribution as a regularization term. The module combines prior probabilities... The above original observation probabilities After making corrections, the final structural posterior probability is obtained. The calculation formula is as follows: ; In the formula, The effective structural length of the text is defined as follows: This variable quantifies the total amount of effective discriminative information contained in the text, rather than simply the length of the words; The prior probability of a time slice is usually assumed to be a uniform distribution in the absence of prior knowledge. ; The smoothing hyperparameter (Pseudo-count) controls the strength of prior beliefs, i.e., the extent to which the system trusts the prior distribution when observational data is insufficient. The value range is usually set to [10, 50]. When the text contains a large number of high-weight fingerprints ( When the text is very short or the fingerprint weight is low, the results depend primarily on the observed data; when the text is extremely short or the fingerprint weight is low ( When the sample size is too small, the result smoothly degenerates into the prior distribution, thus avoiding extreme misjudgments caused by insufficient sample size.

[0066] S124, outputs two-dimensional age values. The feature extraction unit 120 calculates the two probability distributions obtained from the above steps. and Calculate the explicit semantic age value respectively. Age values ​​of hidden structures To obtain continuous time values, a probability-weighted expectation method is used here: ; In the formula, Time slice The central year value. Through the above processing, the feature extraction unit 120 outputs two scalar values. and It accurately quantifies the temporal center of gravity of the text in two independent dimensions: content and form, providing a decoupled quantization input for subsequent temporal detuning calculations.

[0067] See attached document Figure 4 , Figure 4 This is a logical architecture diagram of the association analysis unit 130 according to an embodiment of the present invention. The association analysis unit 130 is connected to the feature extraction unit 120 and the historical corpus database, serving as the system's decision-making center. It is configured to adaptively adjust the retrieval strategy and generate the final association results based on the temporal differences between explicit and implicit features. Logically, this unit is divided into a temporal deviation calculation module 131, a dynamic weight generation module 132, and a hybrid similarity measurement module 133. The specific processing flow is as follows: S131, Execute timing deviation calculation.

[0068] The timing deviation calculation module 131 receives the explicit semantic age value output by the feature extraction unit 120. and latent structure age values This step, based on Statistical Process Control (SPC) principles, aims to detect abnormal deviations in text over time. The module integrates a numerical comparator and a statistical threshold register. The comparator performs interpolation to obtain the temporal detuning rate. The statistical threshold register stores the decision threshold. This threshold is not a fixed constant, but a statistic determined through offline analysis of a benchmark corpus. The system selects native texts with definite dates from the corpus that are confirmed not to be archaic as the sample set, and calculates the average semantic-structural age difference of all texts in this sample set. and standard deviation Determination threshold Set as ,in The confidence coefficient is preferably set within the range of [2.0, 3.0]. Based on Chebyshev's inequality or the normal distribution assumption, this range covers 95% to 99% of the original text variation interval; therefore, values ​​exceeding this threshold are considered invalid. The statistical significance indicates that the text exhibits an unnatural stylistic lag. The module generates a status signal based on the comparison result, which is used for logical branch control in subsequent modules.

[0069] S132, generate dynamic weight parameters.

[0070] The dynamic weight generation module 132 is connected to the timing deviation calculation module 131 and configured to calculate the timing detuning rate. The mapping is a fusion of semantic and structural weights. To avoid drastic shifts in the retrieval strategy between "content-first" and "form-first" due to minor fluctuations in values ​​near the threshold, this module employs a soft-switching control strategy based on a generalized logistic function. The module reads preset control parameters, including the lower limit of structural weights. Upper limit and sensitivity coefficient Structural weight coefficients The formula for generating it is as follows: ; In the formula, The value is set to [0.3, 0.4] to ensure that even for raw text, the system retains approximately 30% to 40% of the structural weight to assist in the judgment. Set to [0.8, 0.9] to ensure that at least 10% of the semantic weight is retained to maintain basic topic relevance even in cases of severe deharmonicity; The value is set to [0.1, 0.5] to control the steepness of the transition curve. The module calculates... Then, based on the normalization constraints Synchronously output semantic weight coefficients.

[0071] S133, calculate the mixed similarity and sort it.

[0072] The hybrid similarity measurement module 133 is the core component for performing large-scale comparisons. To achieve real-time response in large-scale corpora (millions of documents), this module employs a dual-path inverted index strategy for initial screening: the first path is a "vocabulary-document" index, and the second path is a "syntactic fingerprint-document" index. The system then... The size determines the recall strategy: when When the value is high, relax the retrieval depth of the syntactic fingerprint index; when At lower levels, the system primarily relies on the vocabulary index. A set of candidate texts is obtained through a union operation. For sets Each candidate text in The module computes similarity in parallel across two dimensions: in the semantic dimension, the module computes the target text vector. With candidate text vectors cosine similarity In the structural dimension, module-based weighted Jaccard similarity is calculated. To accurately quantify the overlap of sparse structural features, the computational logic incorporates inverse epochal frequency weights stored in the benchmark database. The calculation formula is as follows: ; In the formula, The union of the non-zero syntactic fingerprints of the two texts; Fingerprint is defined as... In the text The original frequency in Multiply by its reverse-era frequency weight The physical meaning of this formula lies in: extracting the common syntactic feature strengths of two texts through the minimum operator in the numerator, and normalizing them through the maximum operator in the denominator. This is because it introduces... Rare syntactic structures with strong contemporary characteristics (such as special inverted sentence structures) will make a decisive contribution to the similarity score once a match is successful.

[0073] Finally, the module uses the dynamic weights generated in step S132 to perform a weighted fusion of the two similarity scores to obtain the final score. The module's built-in sorter ranks the candidate set in descending order based on the score and truncates the top candidates. The module outputs the results. Additionally, for the Top-N records in the results list, the module performs a "contribution backtracking" operation to identify the contributions... The fingerprints of the syntactic subtrees with the largest contribution values ​​(i.e., the weighted values ​​in the numerator) are output along with the results as explanatory evidence to help users understand the source of the high similarity.

[0074] To more intuitively understand the operating mechanism of this invention when processing complex temporal text, the following explanation is based on specific examples.

[0075] Scenario: A user uploaded a poem titled "Imitating the Ancients" by the Qing Dynasty poet Nalan Xingde (1655-1685). While using elegant language common in the Qing Dynasty, the poem deliberately imitates the archaic style of Han and Wei Dynasty folk songs in its syntax. The goal is to correctly identify its "imitation of the ancients" attribute and retrieve its source (i.e., original texts from the Han and Wei periods), rather than simply retrieving poems by other Qing Dynasty poets.

[0076] Processing flow: 1. Feature decoupling and chronological projection: Semantic Dimension (S121): Semantic centroid projection module 121 analyzes content words (such as "west wind" and "round fan") in the text. Although the poet intended to revive the past, his word usage habits (co-occurrence probabilities of words) inevitably bear the statistical characteristics of the 17th century. The model calculates explicit semantic date values. It was 1670 (corresponding to the reign of Emperor Kangxi of the Qing Dynasty).

[0077] Structural dimension: The structural fingerprint matching module 122 extracts the short sentence patterns and specific dependency subtrees of the poem (such as the "verb preposition + double object" structure). These structures were extremely active during the Han and Wei periods (approximately 200-300 AD), but were rare in spoken language or general classical Chinese during the Qing Dynasty. After combining Bayesian smoothing correction, the model calculates the latent structure age value. It is 220 years.

[0078] 2. Timing mistuning detection and strategy adjustment: Deviation calculation: System calculates timing mistuning rate Year.

[0079] Threshold determination: Preset determination threshold (Based on raw text statistics) 150 years. Because... The system determined that the text had significant temporal discord, i.e., it exhibited "archaism".

[0080] Weight Generation: The dynamic weight generation module 132 maps large deviation values ​​to high-structure weights using the Sigmoid function. Assuming... Calculations yielded , This means that subsequent searches will focus heavily on the similarity of the "syntactic skeleton" rather than the "lexical content".

[0081] Related Search and Result Output: The system searches the candidate database. Because... Extremely high. Those Han and Wei Yuefu poems with different content but highly similar syntax (such as "Han Yuefu: Shang Xie") received extremely high mixed scores.

[0082] Output: The first item in the list returned by the system is not a work by any other poet from the Qing Dynasty, but rather "Bai Tou Yin" from the Han Dynasty.

[0083] Explanatory label: The system outputs the label "Ancient / Inheritance (Time Deviation: 14.5 centuries)" and highlights the key syntax subtree fingerprints (such as specific opening phrases) shared by the two on the interface, accurately revealing Nalan Xingde's inheritance relationship with Han Yuefu.

[0084] Experimental verification and effect comparison To verify the effectiveness of the association analysis method based on temporal decoupling proposed in this invention, this embodiment constructs a Chinese historical corpus (covering five dynasties: Tang, Song, Yuan, Ming, and Qing) containing 2 million articles with exact date metadata for testing.

[0085] Experimental setup and comparison benchmarks: The experiment included the following three comparison schemes: Baseline Group A (pure semantic model): Only uses the BERT pre-trained model to extract semantic vectors for cosine similarity retrieval, without including structural features and temporal weighting.

[0086] Baseline B (Static Fusion Model): Uses both semantic vectors and syntactic tree kernels, but with a fixed weight of 0.5:0.5, and does not include... Reverse-era frequency weighting.

[0087] The present invention employs a complete dual-channel timing decoupling and adaptive weight adjustment mechanism.

[0088] Analysis of experimental results: Analysis of dating accuracy and robustness: See attached document Figure 5 , Figure 5 This is a comparison curve of Mean Absolute Error (MAE) for different models under different text lengths.

[0089] Experimental results: The horizontal axis represents the text length (number of words), and the vertical axis represents the mean absolute error (in years) of the predicted year.

[0090] like Figure 5 As shown, in the long text (>100 characters) range, the three models show little difference in performance. However, in the short text (<20 characters, such as a five-character quatrain) range, the errors of the baseline groups A and B spike significantly (MAE>200 years), exhibiting severe overfitting or feature sparsity problems. The proposed solution (solid line), thanks to the introduction of the Bayesian smoothing correction module 123, can still control the MAE within 80 years for extremely short texts. This demonstrates that introducing a prior distribution has a significant robustness improvement effect on short text analysis.

[0091] Analysis of the accuracy of archaic text retrieval: See attached document Figure 6 , Figure 6 It is a confusion matrix heatmap for the retrieval accuracy of special "archaic" samples.

[0092] Experimental design: Select 500 known "Ming Dynasty imitations of Tang poems" as the query objects. The correct search results should include the "original Tang Dynasty texts" that they actually imitated.

[0093] Experimental results: The search results for baseline group A were mainly concentrated on "other poems from the Ming Dynasty" (contemporary interference), with a Top-20 recall rate of only 12%.

[0094] The Top-20 recall rate of the present invention reaches 78%.

[0095] Analysis: This is directly thanks to the dynamic weight generation module 132. When the system detects temporal misalignment, it automatically switches to the "structure-driven" mode, thereby penetrating the temporal interference at the lexical level and successfully locking in syntactic cognates spanning hundreds of years.

[0096] Ablation experiments using inverse epoch frequency weighting (IEF); See attached document Figure 7 , Figure 7 It is a bar chart showing the discriminative power of certain specific syntactic structures.

[0097] Experiment content: Compare the effects of specific syntactic structures (such as "object fronting") on the introduction of... The contribution of the preceding and following values ​​to the similarity calculation.

[0098] Analysis: For example Figure 7 As shown, in the unweighted case, high-frequency common syntax (such as "subject-verb-object") dominates the similarity contribution, resulting in low distinguishability. (Introduction) Subsequently, the weight of general syntax is suppressed (approaching 0), while the weight of rare syntax with strong temporal characteristics is amplified. Experiments show that introducing... Subsequently, the system's F1 score for stylistic classification improved by approximately 15 percentage points.

[0099] Adaptive weighted response curve: See attached document Figure 8 , Figure 8 It is structural weight With the deviation of time Response curve of the change.

[0100] This figure visually illustrates the soft-switching mechanism of the Sigmoid function in the formula.

[0101] when Less than the threshold (e.g., 150 years) Maintaining a low value (approximately 0.35) ensures semantic priority retrieval of the original text; once... Exceeding the threshold It rapidly and smoothly climbs to a high level (approximately 0.85). This non-linear response mechanism effectively solves the shortcomings of traditional linear weighting, which cannot simultaneously address both "regular retrieval" and "cross-generational tracing".

[0102] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for semantic association analysis of Chinese text, characterized in that, The method includes the following steps: A historical corpus benchmark reference system is constructed by dividing it into time slices, and a semantic feature library and a syntactic structure feature library containing distributed weights are established for each time slice. Receive the text to be analyzed and calculate its explicit semantic age value and implicit structural age value through semantic and structural dual channels respectively; The absolute value of the difference between the explicit semantic age value and the implicit structural age value is calculated to obtain the time series deviation value; Based on the time-series deviation value, dynamic structural weight coefficients and semantic weight coefficients are generated using a nonlinear mapping function; The mixed similarity between the text to be analyzed and the candidate texts is calculated based on the weighting coefficients, and a list of related texts is output.

2. The method for semantic association analysis of Chinese text according to claim 1, characterized in that, The construction of the syntactic structure feature library adopts a high-order structure hashing strategy, which specifically includes: Dependency parsing and delexicalization are performed on the text in the corpus, retaining only the part-of-speech tags of nodes and the dependency relation tags of edges, and an abstract syntactic graph is constructed. An iterative aggregation mechanism is used to process the abstract syntactic graph. Through multiple iterations, the neighborhood structure information of each node is aggregated and encoded into a unique hash identifier, thereby transforming the variable-length graph structure into a fixed-length syntactic fingerprint set. The hash identifier is able to capture deep syntactic topological features across the lexical surface of the text.

3. The method for semantic association analysis of Chinese text according to claim 1, characterized in that, The distribution weights are determined using an inverse epochal frequency-weighted strategy: The system dynamically assigns weights based on the extent to which the hash identifiers of each syntax subtree are distributed across different time slices; For common syntactic structures that appear in the vast majority of time slices, the system reduces their weight through a logarithmic penalty mechanism; For specific syntactic structures that appear only in a specific time slice set, the system assigns them high weights to enhance the ability of the structure fingerprint to identify specific historical periods.

4. The method for semantic association analysis of Chinese text according to claim 1, characterized in that, The calculation of explicit semantic age values ​​employs a semantic centroid projection and probabilistic sharpening mechanism: The text to be analyzed is mapped to a document vector in the semantic vector space, and its similarity to the reference centroid of each historical time slice is calculated. A normalized exponential function with a temperature coefficient is introduced to perform a nonlinear transformation on the similarity distribution; By adjusting the temperature coefficient, the probability distribution is concentrated towards the time slice with the highest similarity, thereby sharpening the determination of the text's era while preserving the distribution information and reducing temporal ambiguity.

5. The method for semantic association analysis of Chinese text according to claim 1, characterized in that, The calculation of the latent structure age value introduces a Bayesian smoothing correction mechanism to handle the data sparsity problem: The system incorporates the global historical prior distribution as a regularization term; The confidence weights of the observation probability and the prior distribution are dynamically adjusted based on the total number of valid structural fingerprints in the text to be analyzed. When the text is short or there are few valid fingerprints, the weight of the prior distribution is increased, so that the final structural posterior probability degenerates and approaches the global distribution, thereby avoiding zero-probability misjudgment due to insufficient samples.

6. The method for semantic association analysis of Chinese text according to claim 1, characterized in that, The method also includes an anomaly detection step based on corpus statistical features: Based on the chronological characteristics of native texts in the benchmark corpus, we statistically analyzed the natural deviation distribution of explicit and implicit chronological values. Confidence intervals are constructed using the statistical mean and standard deviation, and the upper limit of the confidence interval is set as the judgment threshold. A text is considered to have significant temporal misalignment only when the temporal deviation value of the text being analyzed exceeds the statistical safety boundary.

7. The method for semantic association analysis of Chinese text according to claim 6, characterized in that, The nonlinear mapping function is configured as a continuous nonlinear response logic based on an S-shaped curve: This logic constructs a soft switch based on timing deviation values; When the timing deviation value is within the determination threshold range, the semantic weight coefficient maintains its dominant position. When the time-series deviation value exceeds the determination threshold, as the deviation increases, the structural weight coefficient is controlled to rise rapidly along an S-shaped curve and tend to saturate, so that when a significant deharmonicity is detected, the retrieval focus is automatically switched from semantic content to syntactic structure.

8. The method for semantic association analysis of Chinese text according to claim 1, characterized in that, The hybrid similarity is measured based on feature-weighted overlap: The analysis takes into account both the cosine similarity between the text to be analyzed and the candidate texts in the semantic vector space and the set overlap in the syntactic structure space. In calculating the set overlap of the syntactic structure space, the distributed weights in step S1 are used to weight each common syntactic hash identifier to ensure that the retrieval results preferentially match those sparse syntactic structures with high discriminative power.

9. The method for semantic association analysis of Chinese text according to claim 1, characterized in that, The candidate text is obtained through an adaptive dual-path indexing strategy, which specifically includes: Pre-build a first index based on lexical inversion and a second index based on syntactic hash fingerprint inversion; The system dynamically adjusts the retrieval path based on the real-time generated structural weight coefficients: when the structural weight coefficients are low, the first index is mainly used for recall to obtain candidate texts; when the structural weight coefficients increase significantly, the traversal depth of the second index is activated and deepened to capture related texts with similar syntactic styles but different vocabulary as candidate texts.

10. A Chinese language text semantic association analysis system, comprising a Chinese language text semantic association analysis method according to any one of claims 1-9, characterized in that, The system includes: The benchmark building unit is configured to build a benchmark reference system that includes a time-segmented semantic mapping module and a high-order structure hash module; The feature extraction unit is configured to calculate the explicit semantic age value through the semantic centroid projection module, and to calculate the implicit structural age value through the structural fingerprint matching module and the Bayesian smoothing correction module. The association analysis unit is configured to obtain the time series deviation through the time series deviation calculation module, determine the weight coefficients through the dynamic weight generation module, and output a list of associated texts based on the weight coefficients through the hybrid similarity measurement module.

Images