A traditional chinese medicine ancient book symptom standardization processing method based on a large language model

Abstract

The present application relates to the field of artificial intelligence and traditional Chinese medicine information processing technology, in particular to a traditional Chinese medicine ancient book symptom standardization processing method based on a large language model, comprising: obtaining a standard symptom term library, a conversion rule knowledge base and original text data; performing term standardization and semantic coding on the term library to construct a standard symptom ideal semantic field; performing rule injection on the ideal semantic field and the rule base to determine a simulated symptom semantic field and a theoretical residual space; slicing the original text and using a preset large language model to identify candidate symptom segments, encode to obtain an initial symptom vector and determine a real residual space; determining a direct standard mapping result, a composite split mapping result and an unmatched record according to the real residual space and the theoretical residual space, clustering and inducing the unmatched record, determining a new symptom candidate result and updating the conversion rule knowledge base; the present application realizes unified processing of direct mapping, composite splitting and unknown symptom discovery of ancient book symptoms.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and traditional Chinese medicine information processing technology, specifically a method for standardizing the processing of symptoms in ancient Chinese medicine texts based on a large language model. Background Technology

[0002] The standardization of symptoms in ancient Chinese medicine texts aims to map the symptom descriptions in ancient texts to modern standard Chinese medicine symptom terms. Its technical essence is to establish a semantic association between unstructured ancient texts and modern standard medical ontology, so as to facilitate subsequent clinical retrieval and academic collation.

[0003] Existing text standardization mapping methods typically employ direct text similarity matching. However, due to the prevalence of variant characters, lack of punctuation, compression of classical Chinese lexical structures, and multiple symptoms being written together in classical Chinese texts, traditional direct similarity matching methods struggle to cope with the disturbances in classical Chinese contexts. This coarse-grained matching method is prone to mistaking fever and chills as a single symptom and to incorrectly breaking down whole-item terms such as epigastric fullness and hardness, resulting in low mapping accuracy.

[0004] Currently, for business scenarios involving the digitization of massive amounts of ancient Chinese medicine texts, it is necessary to capture more granular semantic features and word formation rules to assist in symptom identification. Therefore, how to automatically and accurately achieve standardized conversion and mapping of symptom texts when there are variant expressions and complex perturbations in ancient texts has become a technical problem to be solved. Summary of the Invention

[0005] The purpose of this invention is to provide a method for standardizing symptoms in ancient Chinese medicine texts based on a large language model, thereby solving the following technical problems:

[0006] It avoids misjudgment or incorrect splitting of symptoms caused by using only direct text similarity matching, and makes it easier to achieve unified and stable processing of direct mapping, compound splitting, and discovery of unknown symptoms.

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] A method for standardizing symptoms in ancient Chinese medicine texts based on a large language model, the method comprising:

[0009] Acquire a standard symptom terminology database, a conversion rule knowledge base, and raw text data;

[0010] The standard symptom terminology database is subjected to terminology normalization and semantic encoding to determine a standard symptom vector set, and a standard symptom ideal semantic field is constructed based on the standard symptom vector set, wherein the standard symptom ideal semantic field is a vector space composed of the standard symptom vector set;

[0011] The standard symptom ideal semantic field and the transformation rule knowledge base are subjected to rule injection processing to determine the simulated symptom semantic field, wherein the simulated symptom semantic field is composed of multiple simulation vectors; and the theoretical residual space is determined based on the set of difference vectors between the simulation vectors in the simulated symptom semantic field and the standard symptom vectors in the standard symptom ideal semantic field.

[0012] The original text data is processed into text segments to determine target text segments, and a preset large language model is used to perform symptom recognition processing on the target text segments to determine candidate symptom fragments.

[0013] The candidate symptom fragments are semantically encoded to determine the initial symptom vector, and the actual residual space is determined based on the set of difference vectors between the initial symptom vector and the standard symptom vector in the ideal semantic field of the standard symptom.

[0014] Based on the actual residual space and the theoretical residual space, a coupling decision result is determined. The coupling decision result includes a direct standard mapping result, a composite split mapping result, and unmatched records. The unmatched records include unmatched candidate symptom fragments and their corresponding difference vectors.

[0015] The unmatched records are clustered and summarized to determine new symptom candidate results, and the conversion rule knowledge base is updated based on the new symptom candidate results.

[0016] Preferably, the standard symptom terminology database is subjected to terminology normalization and semantic encoding processing to determine a standard symptom vector set, and an ideal semantic field of standard symptoms is constructed based on the standard symptom vector set, including:

[0017] The standard symptom terminology database is subjected to terminology standardization processing to determine the standard symptom set;

[0018] The standard symptom set is semantically encoded to determine the standard symptom vector set;

[0019] Construct the ideal semantic field of standard symptoms based on the set of standard symptom vectors.

[0020] Preferably, rule injection processing is performed on the standard symptom ideal semantic field and the transformation rule knowledge base to determine the simulated symptom semantic field, wherein the simulated symptom semantic field is composed of multiple simulated vectors; and the theoretical residual space is determined based on the set of difference vectors between the simulated vectors in the simulated symptom semantic field and the standard symptom vectors in the standard symptom ideal semantic field, including:

[0021] Based on the ancient and modern terminology conversion rules in the conversion rule knowledge base, the standard symptom vector in the standard symptom ideal semantic field is converted to determine the conversion simulation vector;

[0022] Based on the compound word formation rules in the transformation rule knowledge base, the standard symptom vectors in the standard symptom ideal semantic field are combined to determine the compound simulation vector;

[0023] The simulated symptom semantic field is determined based on the transformed simulation vector and the composite simulation vector;

[0024] The theoretical residual space is determined based on the difference vector between the simulation vector in the simulated symptom semantic field and the corresponding standard symptom vector in the standard symptom ideal semantic field.

[0025] Preferably, the method includes: performing slicing processing on the original text data to determine the target text slice;

[0026] Using the preset large language model, symptom recognition processing is performed on the target text slice to determine candidate symptom segments, wherein the preset large language model is a large language model that has been finely adjusted by the corpus of ancient Chinese medicine texts.

[0027] The candidate symptom fragments are subjected to location word recognition, property word recognition, and connector word recognition to determine location features, property features, and parallel connection features;

[0028] The candidate symptom fragments are semantically encoded, wherein the semantic encoding process uses the same or mapped-aligned semantic encoding model as the standard symptom ideal semantic field to determine the initial symptom vector;

[0029] The real residual space is determined based on the set of difference vectors between the initial symptom vector and the standard symptom vector in the standard symptom ideal semantic field, combined with the location feature, the property feature and the parallel connection feature.

[0030] Preferably, the coupling decision result is determined based on the actual residual space and the theoretical residual space, including:

[0031] Structural similarity, direct semantic similarity, and transformed semantic similarity are calculated for the real residual space and the theoretical residual space to determine structural similarity, direct semantic similarity, and transformed semantic similarity. The structural similarity characterizes the consistency between the feature extraction structure of the candidate symptom fragment and the corresponding structural template of the preset rule path in the transformed rule knowledge base. The direct semantic similarity characterizes the vector similarity between the initial symptom vector and the corresponding single vector in the standard symptom vector set. The transformed semantic similarity characterizes the vector similarity between the initial symptom vector and the corresponding simulation vector in the simulated symptom semantic field.

[0032] The comprehensive coupling value is determined based on a weighted combination of the structural similarity, the direct semantic similarity, and the transformed semantic similarity.

[0033] If the overall coupling value is greater than or equal to a preset high threshold, the coupling decision result is determined to be a direct standard mapping result; if the overall coupling value is less than or equal to a preset low threshold, the coupling decision result is determined to be an unmatched record; if the overall coupling value is greater than the preset low threshold and less than the preset high threshold, the coupling decision result is determined to be a composite split mapping result.

[0034] The preset high threshold and the preset low threshold are discrimination thresholds predetermined based on the comprehensive coupling value distribution in the labeled sample set, and the preset high threshold is greater than the preset low threshold.

[0035] Preferably, if the comprehensive coupling value is greater than or equal to a preset high threshold, then the coupling decision result is determined to be a direct standard mapping result, including:

[0036] The target standard symptom is determined based on the optimal matching standard symptom vector in the ideal semantic field of the standard symptoms from the initial symptom vector.

[0037] The direct standard mapping result is determined based on the target standard symptoms and the standard symptom terminology database.

[0038] Preferably, if the comprehensive coupling value is greater than the preset low threshold and less than the preset high threshold, then the coupling decision result is determined to be a composite split mapping result, including:

[0039] Extract the location features, property features, and parallel connection features corresponding to the candidate symptom fragments from the real residual space;

[0040] The composite boundary is determined based on the location characteristics, the property characteristics, and the parallel connection characteristics;

[0041] Based on the composite boundary, the candidate symptom fragments are segmented to determine the split symptom sequence;

[0042] The composite splitting mapping result is determined based on the splitting symptom sequence and the standard symptom terminology database.

[0043] Preferably, the unmatched records are clustered and summarized to determine candidate results for new symptoms, including:

[0044] Clustering is performed on the residual vectors of the unmatched records to determine candidate clusters;

[0045] Using the preset large language model, the candidate clusters are semantically summarized to determine the cluster labels and definition descriptions that represent the common semantics of the candidate clusters;

[0046] Based on the cluster labels and their definitions, the candidate results for the new symptoms are determined.

[0047] Preferably, updating the conversion rule knowledge base based on the newly added symptom candidate results includes:

[0048] Based on the newly added symptom candidate results, new alternative name rules, new compound word formation rules, and new semantic mapping relationships are determined;

[0049] The newly added alternative name rules, the newly added compound word formation rules, and the newly added semantic mapping relationships are written into the conversion rule knowledge base to determine the updated conversion rule knowledge base.

[0050] Preferably, the original text data is ancient Chinese medicine text data, wherein the ancient Chinese medicine text data includes text without punctuation, text with variant characters, and text describing symptoms in classical Chinese.

[0051] The standard symptom terminology database contains pre-built ontology data of modern TCM standard symptoms;

[0052] The conversion rule knowledge base includes ancient and modern Chinese morphological rules, compound word formation rules, part-of-speech substitution rules, and measurement expression rules.

[0053] The beneficial effects of this invention are:

[0054] 1. This invention constructs a standard symptom ideal semantic field and injects transformation rules to generate a simulated symptom semantic field; by comparing the real residual space and theoretical residual space of the original text for coupling judgment, simple text matching is transformed into stable residual coupling judgment; effectively solving the problem that traditional keyword matching is difficult to adapt to the misidentification caused by different names and symptoms in ancient books and parallel writing;

[0055] 2. This invention combines structural similarity, direct semantic similarity, and transformed semantic similarity to calculate a comprehensive coupling value, and intelligently divides the identification results into direct mapping, composite splitting, and unmatched records accordingly. This multi-dimensional joint evaluation method overcomes the defects of traditional mechanical rule-based segmentation, can accurately identify and retain indivisible overall symptoms, and effectively avoids erroneous segmentation.

[0056] 3. This invention proactively utilizes the rules for converting ancient and modern usage and the rules for compound word formation to combine and transform the standard symptom vector, generating a theoretical residual space in advance. This mechanism pre-parameterizes the predictable contextual changes and alternative name substitution phenomena in ancient books, making up for the shortcomings of traditional dictionary merging methods in covering unregistered expressions and complex ancient text variants.

[0057] 4. This invention performs clustering processing on the residual vectors of unmatched candidate symptoms and uses a large language model to perform semantic induction to determine the candidate results of new symptoms, and then writes back to update the knowledge base of transformation rules; this breaks through the bottleneck of traditional methods being unable to discover new symptoms, transforms unknown abnormal expressions into knowledge increments, and realizes the closed-loop self-iterative improvement of the knowledge base. Attached Figure Description

[0058] The invention will now be further described with reference to the accompanying drawings.

[0059] Figure 1 This is a flowchart illustrating a method for standardizing symptoms in ancient Chinese medicine texts based on a large language model, as provided in an embodiment of this application. Detailed Implementation

[0060] 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.

[0061] Please see Figure 1 A method for standardizing symptoms in ancient Chinese medicine texts based on a large language model is proposed. The method includes: acquiring a standard symptom terminology database, a conversion rule knowledge base, and original text data; performing terminology standardization and semantic encoding on the standard symptom terminology database to determine a set of standard symptom vectors; and constructing a standard symptom ideal semantic field based on the set of standard symptom vectors, wherein the standard symptom ideal semantic field is a vector space composed of the set of standard symptom vectors.

[0062] The standard symptom ideal semantic field and the transformation rule knowledge base are subjected to rule injection processing to determine the simulated symptom semantic field, wherein the simulated symptom semantic field is composed of multiple simulation vectors; and the theoretical residual space is determined based on the set of difference vectors between the simulation vectors in the simulated symptom semantic field and the standard symptom vectors in the standard symptom ideal semantic field.

[0063] The original text data is processed into text segments to determine target text segments. A preset large language model is used to perform symptom recognition processing on the target text segments to determine candidate symptom fragments. The candidate symptom fragments are then processed into semantic encoding to determine initial symptom vectors. Based on the set of difference vectors between the initial symptom vectors and the standard symptom vectors in the standard symptom ideal semantic field, the real residual space is determined.

[0064] Based on the actual residual space and the theoretical residual space, i.e., the set of residual vectors, a coupling decision result is determined. The coupling decision result includes the direct standard mapping result, the composite split mapping result, and the unmatched record. The unmatched record includes the unmatched candidate symptom fragment and its corresponding difference vector. The unmatched record is then subjected to clustering and induction processing to determine the new symptom candidate result. The transformation rule knowledge base is then updated based on the new symptom candidate result.

[0065] This embodiment provides a standardized processing mechanism for symptoms in ancient Chinese medicine texts based on a large language model. Specifically, this embodiment uses the construction of an ancient text symptom knowledge base system to build a knowledge foundation for symptoms in ancient texts, which will be used for subsequent clinical auxiliary retrieval and academic collation as a unified main scenario. The system continuously receives scanned and transcribed texts of ancient texts on typhoid fever, febrile diseases, and miscellaneous diseases, aiming to map the symptom descriptions in ancient texts to modern standard Chinese medicine symptom terminology, so that doctors can retrieve historical evidence of different names for the same symptoms in ancient and modern times.

[0066] Because ancient books commonly contain variant characters, lack punctuation, compression of classical Chinese lexical structures, and multiple symptoms written together, if only direct text similarity matching is used, it is easy to mistake fever and chills for a single symptom, and it is also easy to break up the epigastric fullness and hardness. Therefore, a dual-track coupling processing flow of ideal semantic field, simulated semantic field, real residual space and theoretical residual space is adopted.

[0067] To avoid ambiguity caused by multiple names for the same technical object throughout the text, in this embodiment and subsequent embodiments, the term "ideal semantic field" is only used as an abbreviation for "standard symptom ideal semantic field", the term "simulation semantic field" is only used as an abbreviation for "simulation symptom semantic field", the term "term library" is only used as an abbreviation for "standard symptom term library", and the term "rule library" is only used as an abbreviation for "conversion rule knowledge base". None of the above abbreviations introduce new technical objects.

[0068] Accordingly, the names of standard symptom vector, simulation vector, initial symptom vector, theoretical residual space, real residual space, coupling decision result, and unmatched record, etc., retain a unique meaning throughout the text and do not change their referent due to changes in the scenario;

[0069] The specific processing steps are as follows: Three types of basic data are acquired. The first type is a standard symptom terminology database, such as the standard symptom ontology of modern Chinese medicine, which includes standard entries like fever, chills, headache, abdominal distension, and epigastric fullness and hardness, and records their standardized names, synonyms, and hierarchical relationships. The second type is a conversion rule knowledge base, used to record rules for converting ancient and modern usage, rules for compound word formation, rules for using place names, and rules for measurement and expression. The third type is the original text data, i.e., the transcribed text stream of ancient books. For example, a slice of the original text on one page might describe Taiyang disease with fever, chills, headache, and neck stiffness, while another page might describe epigastric fullness and hardness without pain.

[0070] The standard symptom terminology database is subjected to terminology normalization and semantic encoding. Terminology normalization is used to eliminate synonymous differences. For example, aversion to cold is uniformly merged into the set of alternative names for aversion to cold, and headache and neck stiffness are separated to extract the standardized association information between headache and neck stiffness. Semantic encoding can use a pre-trained medical text encoder or a vector encoder adapted to TCM corpus, such as a medical language model based on the BERT architecture, the RoBERTa model, or the Word2Vec algorithm, to encode each standard term into a standard symptom vector of fixed dimensions.

[0071] For ease of explanation, assuming the encoding dimension is 3-dimensional, then fever can correspond to the vector (0.90, 0.10, 0.20), chills to the vector (0.12, 0.91, 0.18), and epigastric fullness to the vector (0.22, 0.34, 0.88). The set of vectors consisting of all standard symptom vectors constitutes the ideal semantic field of standard symptoms. This ideal semantic field can be understood as a standard reference space without the noise of classical Chinese expression.

[0072] Based on this, rule injection processing is performed on the ideal semantic field and the knowledge base of transformation rules to obtain the simulated symptom semantic field. The so-called rule injection is to actively apply the common expression transformation methods in ancient texts to the standard vector to form the ancient text expression vectors that may appear in theory. For example, according to the ancient and modern usage conversion rules, the alternative name of chills, aversion to cold, can be mapped to a simulated vector that is adjacent to but not exactly the same as chills. According to the compound word formation rules, fever and chills can be combined in a parallel connection pattern to form a compound simulated vector of fever and chills.

[0073] Continuing with the 3D example, if fever is (0.90, 0.10, 0.20) and chills are (0.12, 0.91, 0.18), then the composite simulation vector obtained through the combination rule can be (0.51, 0.53, 0.19). The difference between this simulation vector and the corresponding standard vector, for example, the difference with fever is (-0.39, 0.43, -0.01), and the difference with chills is (0.39, -0.38, 0.01). The set of such difference vectors constitutes the theoretical residual space. This space does not express the specific symptoms themselves, but rather the interpretable offset patterns brought about by known ancient text transformations, alternative name changes, and compound writing.

[0074] The original text data is then processed by text slicing. Slicing can be done by sentence candidate points, function word boundaries, medical record entry boundaries, or fixed word count windows. For example, a slice S1 can be formed by taking fever, chills, headache, and neck pain in Taiyang disease as an example. For example, a slice S2 can be formed by taking epigastric fullness and hardness without pain as an example. For each slice, a preset large language model is called to perform symptom recognition and output candidate symptom fragments.

[0075] For example, candidate segments P1 (fever and chills) and P2 (head and neck stiffness) are identified in S1; candidate segment P3 (epigastric fullness and hardness) is identified in S2; then the candidate segments are semantically encoded to obtain the initial symptom vector; for example, P1 is encoded as (0.49, 0.55, 0.17), and P3 is encoded as (0.23, 0.33, 0.86);

[0076] The initial symptom vector is then subtracted from each standard vector in the ideal semantic field to obtain the real residual space. Taking P1 as an example, its difference from fever is (-0.41, 0.45, -0.03), and its difference from chills is (0.37, -0.36, -0.01). This residual contains both real semantic information and the bias caused by the classical Chinese expression.

[0077] The actual residual space and the theoretical residual space are coupled for judgment. If the actual residual of a candidate segment is very close to a certain theoretical residual, it indicates that the abnormal offset of the segment mainly comes from the ancient Chinese expression mode that can be explained by known rules. At this time, the direct standard mapping result or the composite split mapping result can be output. If the actual residual cannot be fully explained by the theoretical residual, it is retained as an unmatched record.

[0078] Taking P1 fever and chills as an example, its actual residual is highly consistent with the theoretical residual generated by the fever + chills combination rule. Therefore, it is judged as an interpretable compound symptom and enters the compound splitting mapping path to output the two standard symptoms of fever and chills. Taking P3 epigastric fullness and hardness as another example, if its actual residual is closer to the single standard vector of epigastric fullness and hardness in the ideal semantic field, and not close to any splitting and combination rule, it is directly mapped to the standard symptom epigastric fullness and hardness.

[0079] If the deviation value of the calculated residual of another candidate segment P4 with indescribable pain in joints is greater than the preset deviation threshold from any existing theory residual, it is recorded as an unmatched segment, and its residual vector is saved at the same time;

[0080] Further perform clustering and induction processing on the unmatched records; specifically, multiple unmatched residual vectors can be input into a clustering algorithm, such as density clustering or hierarchical clustering, to form several candidate clusters; then call the preset large language model to generalize the commonalities of the original text segments in each cluster, and output possible new symptom candidate labels and definition descriptions;

[0081] For example, if multiple unmatched segments all show symptoms such as pain in joints, soreness in limbs, and pain in all joints, etc., the candidate label "symptoms like general joint pain" can be generalized; transfer the candidate result to the artificial knowledge maintenance end for review, and write it into the conversion rule knowledge base or expand the term library after passing the review;

[0082] In the exception handling mechanism, if there are duplicate entries in the standard symptom term library or inconsistent encodings from different sources, first perform term de-duplication and vector re-encoding to avoid self-contradiction within the ideal semantic field;

[0083] If the length of a certain text slice is too short, such as only the single character "pain", it is marked as a low-information slice and does not directly enter the standardization process, but is retried after being spliced with adjacent slices; if the large language model outputs multiple conflicting candidate segments for the same slice, retain the top several results with the highest confidence for parallel evaluation, and the number of candidate segments entering parallel evaluation is limited by the preset reserved candidate number N, and the N is preferably 2 to 5;

[0084] If the coupling values of the retained results are all lower than the lowest threshold, they are uniformly entered into the unmatched records; if the sample size of the unmatched clustering is insufficient, such as a certain residual cluster has only 1 record, it is first retained as an observation cluster and the knowledge base is not updated immediately to avoid accidental noise contaminating the rule system;

[0085] Exemplarily, in an actual data processing task, the system extracts the text "Taiyang disease, fever and aversion to cold, stiffness and pain in the head and nape" from the relevant ancient books of Treatise on Cold Damage. After slicing, recognition, encoding and dual-track coupling, the standard symptoms "fever and aversion to cold, headache and stiffness in the nape" are output; for "epigastric fullness and hardness", the overall mapping is maintained without incorrect splitting; for several uncommon symptom descriptions that frequently appear in the collected medical records of epidemic febrile diseases but have not been included in the term library, they are aggregated as new candidates for subsequent knowledge review;

[0086] The purpose of this step is to establish a processing framework that first constructs a pure standard benchmark, then actively simulates ancient text perturbations, and compares the real text with the theoretical perturbations. This transforms the standardization of ancient text symptoms from a simple text matching problem into a more stable residual coupling judgment problem, thereby achieving unified processing of direct mapping, composite decomposition, and the discovery of unknown symptoms.

[0087] In a preferred embodiment of the present invention, the standard symptom terminology database is subjected to terminology normalization and semantic encoding processing to determine a standard symptom vector set, and an ideal semantic field of standard symptoms is constructed based on the standard symptom vector set. This includes: performing terminology normalization processing on the standard symptom terminology database to determine a standard symptom set; performing semantic encoding processing on the standard symptom set to determine a standard symptom vector set; and constructing the ideal semantic field of standard symptoms based on the standard symptom vector set.

[0088] This embodiment provides a refinement mechanism for constructing an ideal semantic field for standard symptoms. Specifically, in the aforementioned document digitization center scenario, if only the original names in the terminology database are directly fed into the encoder, the multiple centers of synonyms and their proper names will interfere with each other, resulting in the ideal semantic field itself not being pure.

[0089] This would cause the reference frame for subsequent residual calculations to drift. Therefore, this embodiment first standardizes the standard terminology and then performs semantic encoding.

[0090] The specific processing steps are as follows: In the terminology standardization stage, candidate term sets are first extracted from the standard symptom terminology database, and a structured record is established for each term: standard name - alternative name - non-splitting tag - splittable tag - hierarchical parent node; for example, chills is the standard name, and aversion to cold and fear of cold are alternative names.

[0091] "Hardness and fullness in the chest" is the standard name and is marked as not to be split. "Strong head and neck" can be recorded as a combined expression and associated with the two standard nodes of "headache" and "stiff neck". Then, terminology merging is performed to unify spelling differences, variant character differences, and synonym differences. For example, if "irritability" and "restlessness" are confirmed by artificial knowledge rules to be different standard symptoms, they are retained separately. If "aversion to cold" and "chills" are synonymous, the former is folded into an alternative name entry for the latter and is no longer retained as a separate standard center.

[0092] After standardization, the standard symptom set is semantically coded. Instead of using a single phrase, a combined input method of standard name + definition description + hierarchical relationship prompts is adopted to make the coding results more stable. For example, the input of chills is not just the two words "chills", but can also include chills, which means that the patient's subjective feeling of cold is obvious and is not based solely on changes in external temperature.

[0093] The input for "epigastric fullness and hardness" can include fullness and hardness in the epigastric region, representing a specific overall symptom, and is not subject to the restriction of splitting according to general parallel rules; the resulting vector better reflects the standard semantic boundary; for ease of explanation, assume that only three standard symptoms are retained after normalization: A is fever, B is chills, and C is epigastric fullness and hardness, with corresponding encoding vectors VA=(0.90,0.10,0.20), VB=(0.12,0.91,0.18), and VC=(0.22,0.34,0.88), respectively; the set of all standard vectors constitutes the ideal semantic field;

[0094] The construction of an ideal semantic field can be represented as either a vector table or an indexed vector space structure. For example, the system establishes a main index table, where the first row corresponds to fever and vector VA, the second row corresponds to chills and vector VB, and the third row corresponds to epigastric fullness and vector VC.

[0095] At the same time, an adjacency index is established to record which standard symptoms are allowed to participate in compound word formation in the subsequent rule injection stage, and which standard symptoms are prohibited from being split; in this way, the ideal semantic field is not only a set of vectors, but also a benchmark space that facilitates subsequent retrieval and comparison.

[0096] In the anomaly handling mechanism, if two standard term vectors are found to be too close after normalization, for example, the cosine similarity exceeds 0.98, they are not directly merged, but are put into the manual review list to prevent near-synonymous but different symptoms from being folded in error; if a standard term lacks a definition description, its parent node and similar examples are used to automatically complete a brief description before encoding.

[0097] If completion fails, it is encoded only with the standard name and marked as a low-confidence standard point in the index for recalculation in subsequent iterations; if a term has both splittable and non-splittable conflicting labels, the non-splittable label takes priority to avoid destroying the overall symptom boundary when performing subsequent compound splitting.

[0098] For example, when the center was organizing ancient books of the Jin Gui category, the staff sent the terms "heart-lower epigastric fullness" and "heart-lower epigastric hardness" into the standardization module; the system determined that "heart-lower epigastric hardness" was a standard item for the whole symptoms according to the definition in the terminology library, and did not establish independent centers for the three parts of "heart-lower epigastric hardness"; while fever and chills were retained as independent standard items, and allowed to generate a combined simulation state of fever and chills in subsequent composite rules.

[0099] The purpose of this step is to establish a stable, clean, and clearly defined reference benchmark for the entire standardization process, thereby enabling interpretability and consistency of subsequent residual calculations.

[0100] In a preferred embodiment of the present invention, rule injection processing is performed on the standard symptom ideal semantic field and the transformation rule knowledge base to determine the simulated symptom semantic field, wherein the simulated symptom semantic field is composed of multiple simulated vectors; and the theoretical residual space is determined according to the set of difference vectors between the simulated vectors in the simulated symptom semantic field and the standard symptom vectors in the standard symptom ideal semantic field, including: transforming the standard symptom vectors in the standard symptom ideal semantic field according to the ancient and modern terminology transformation rules in the transformation rule knowledge base to determine the transformed simulation vectors;

[0101] Based on the compound word formation rules in the transformation rule knowledge base, the standard symptom vectors in the standard symptom ideal semantic field are combined to determine the compound simulation vector; based on the transformed simulation vector and the compound simulation vector, the simulated symptom semantic field is determined; based on the difference vector between the simulation vector in the simulated symptom semantic field and the corresponding standard symptom vector in the standard symptom ideal semantic field, the theoretical residual space is determined.

[0102] This embodiment provides a rule injection mechanism for ancient Chinese context perturbation. Specifically, in the case of only an ideal semantic field, although the system knows the data features of modern standard symptoms, it does not know how these symptoms will appear in ancient books. If this step is missing, then the ancient and modern variant names and parallel abbreviations appearing in real texts will be mistakenly regarded as abnormal. Therefore, this embodiment further constructs a simulated symptom semantic field to generate interpretable offsets in advance.

[0103] To avoid multiple interpretations of the corresponding standard symptom vector under different rule paths, this embodiment makes a unified agreement on its determination method: when the simulation vector is generated by the ancient and modern terminology conversion rule, its corresponding standard symptom vector is the source standard symptom vector that generated the simulation vector, which is a one-to-one correspondence;

[0104] When the simulation vector is generated by the compound word formation rule, its corresponding standard symptom vector is the source standard symptom vectors that participate in the compound word formation, which is a one-to-many correspondence; accordingly, each residual record in the theoretical residual space, in addition to the residual vector itself, also stores the rule type, source standard symptom identifier and the composition order identifier in the compound case, so as to ensure that when the actual residual and the theoretical residual are coupled and compared later, it can be clearly identified whether the offset comes from the alternative name conversion path or from the compound word formation path;

[0105] The specific processing procedure is as follows: Based on the ancient and modern terminology conversion rules in the conversion rule knowledge base, the standard symptom vector is converted. The conversion process involves applying a regularized offset to the standard vector, bringing it closer to common expressions in ancient texts. Specifically, this is done by introducing a regularized offset vector; the formula for calculating the converted simulation vector is as follows: ,in For the corresponding standard symptom vector, This is a preset offset extracted based on a knowledge base of transformation rules for specific ancient and modern terms;

[0106] For example, the knowledge base records that aversion to cold often corresponds to aversion to cold and fear of cold, abdominal distension often corresponds to abdominal fullness, and headache can be seen in some ancient books as a compound expression prefix of head and neck stiffness; for ease of explanation, the transformation function can be understood as adding a small offset to the original vector; if the standard vector of aversion to cold is VB=(0.12,0.91,0.18), and the transformation offset for the aversion to cold rule is set as Δ1=(0.02,-0.03,0.01), then the transformation simulation vector VB1=(0.14,0.88,0.19) is obtained;

[0107] If the offset for the rule of fear of cold is set as Δ2=(0.01,-0.04,0.00), then VB2=(0.13,0.87,0.18) is obtained; these transformation simulation vectors represent reasonable variations of the same standard symptom in the ancient text environment;

[0108] Based on the rules of compound word formation, standard vectors are combined to obtain compound simulation vectors. The combination is not a simple addition, but takes into account participles, morphological words, conjunctions, and classical Chinese omission habits. The compound simulation vector is obtained by weighted summation of multiple standard symptom vectors involved in the combination; the calculation formula is as follows: ,in For the first to participate in the compound A standard symptom vector, The combined weight coefficients are assigned in the rule base according to the aforementioned conjunctions and the omission habits of classical Chinese; for example, fever and chills are typical parallel compound structures, and the fever vector VA and the chills vector VB can be weighted and combined; headache and neck stiffness can be regarded as a connected expression of headache + neck stiffness.

[0109] For example, if VA=(0.90,0.10,0.20) and VB=(0.12,0.91,0.18), and the weights of the parallel combination are set to 0.5 and 0.5 respectively, then the composite simulation vector VAB=(0.51,0.505,0.19) is obtained; if the more core expression habits of the former in ancient books are considered, the weights can also be set to 0.6 and 0.4, resulting in VAB'=(0.59,0.424,0.192); the system can retain multiple simulation states for the same rule path to cover different expression habits;

[0110] After obtaining the transformed simulation vector and the composite simulation vector, a simulation symptom semantic field is formed; each simulation vector in this semantic field has a source label, such as being generated by the rule of ancient and modern alternative names of the chills meridian, or by the rule of parallel composite of fever and chills meridian.

[0111] Furthermore, the system calculates the difference between each simulation vector and its corresponding standard vector to form a theoretical residual; for example, the difference between VAB and VA is (-0.39, 0.405, -0.01), and the difference between VAB and VB is (0.39, -0.405, 0.01); these differences are not directly used for mapping, but are used to describe how much and in which direction the real text should deviate from the standard if it deviates from the standard only because of the rule-interpretable ancient text transformation.

[0112] In the context of compound word formation, the same compound simulation vector can generate multiple theoretical residual records. Each record is stored in relation to a source standard symptom vector to avoid logical confusion between the compound simulation vector relative to the entire combination and the compound simulation vector relative to a single source standard symptom during subsequent comparisons.

[0113] In the anomaly handling mechanism, if the simulation vector generated by a certain rule is too far from the original standard vector, for example, the cosine similarity is less than 0.5, it means that the rule may cause too much semantic deviation. The system will not write it into the simulation semantics, but will record it as an abnormal rule candidate and wait for manual revision.

[0114] If multiple rules generate the same or similar simulation vectors, deduplication is performed, and the number of rule sources is retained as a weight to improve the stability of subsequent coupling decisions. If a composite combination violates the prohibition on splitting, for example, attempting to split the heart-lower-pit-hard into heart-lower+pit-hard or heart-lower+pit+hard, the system directly blocks the combination path and does not generate the corresponding simulation vector.

[0115] For example, when the system processes ancient books on typhoid fever, the knowledge base has pre-set high-frequency parallel rules for fever + chills, as well as pre-set rules for the ancient and modern usage of chills ←→ aversion to cold.

[0116] Therefore, the system actively generates several composite simulation states and alternative simulation states for fever and chills in the ideal semantic field; when fever and chills or aversion to cold actually appear in the subsequent text, the system does not treat these expressions as completely unfamiliar inputs, but uses the real residuals to find these pre-generated theoretical offset trajectories.

[0117] The purpose of this mechanism is to pre-parameterize the predictable contextual changes, alternative name substitutions, and compound word formation phenomena in ancient books, thereby achieving interpretable modeling of real-world text offsets and reducing the reliance of subsequent judgments on a single similarity index.

[0118] In a preferred embodiment of the present invention, the method includes: performing slicing processing on the original text data to determine target text slices;

[0119] Using the pre-defined large language model, symptom recognition processing is performed on the target text slices to determine candidate symptom fragments. The pre-defined large language model is a large language model fine-tuned from ancient Chinese medicine texts. Specifically, it can be a generative or interpretative pre-trained language model based on the Transformer architecture, such as ChatGLM or LLaMA, trained using fine-tuning techniques based on ancient Chinese medicine text instructions. The candidate symptom fragments are then subjected to location word recognition, property word recognition, and conjunction word recognition to determine location features, property features, and parallel connection features. Finally, the candidate symptom fragments undergo semantic encoding processing to determine the initial symptom vector.

[0120] The real residual space is determined based on the set of difference vectors between the initial symptom vector and the standard symptom vector in the standard symptom ideal semantic field, combined with the location feature, the property feature and the parallel connection feature.

[0121] This embodiment provides a mechanism for extracting the actual residuals of the original text of ancient books. Specifically, based on the above, theoretical residuals alone are not enough to make a decision, because the system also needs to know the actual offset pattern presented by the ancient text. If the entire segment is directly input into the encoder, the cause, treatment, dosage and symptoms will often be mixed into the same vector. Therefore, this embodiment first performs text slicing, and then performs symptom recognition and feature extraction.

[0122] The specific processing steps are as follows: In the original text slicing stage, candidate punctuation marks, function words, pause words, and fixed windows can be used. For ancient texts lacking punctuation or with non-standard punctuation, a dynamic sentence segmentation algorithm based on character mutual information and boundary entropy is introduced before slicing. Specifically, the algorithm determines the segmentation probability of adjacent characters by calculating the probability of segmentation. The algorithm formula is as follows: ,in For pointwise mutual information between two adjacent characters, The right boundary entropy of the left character. The left boundary entropy of the right-hand character. , , The model adjustment coefficient is the splitting probability. Segmentation is performed when the value exceeds a preset dynamic threshold;

[0123] For example, for the text "Taiyang disease with fever, chills, headache, neck stiffness, and floating pulse", several candidate windows can be formed at the locations of the symptoms, and then the target text slices can be selected based on the symptom density score. The specific calculation method of the symptom density score can be: the ratio of the number of TCM symptom feature words in the candidate window to the total number of characters in the window. The higher the ratio, the higher the score of the window. In a simplified example, the system slices the original text into S1 fever and chills, S2 headache and neck stiffness, and S3 floating pulse.

[0124] Since this system mainly deals with symptoms, S1 and S2 are retained as target text slices, while S3 can be sent to the vital signs submodule or processed temporarily.

[0125] Then, a large language model fine-tuned from ancient Chinese medicine texts is used for symptom identification. The model does not output the final standard terms, but rather candidate symptom fragments. For example, for S1, the model outputs P1 fever and chills; for S2, the model outputs P2 headache and neck pain; for epigastric fullness and pain without pain, the model can output P3 epigastric fullness and pain and P4 fullness without pain. At this stage, the model is responsible for pointing out the boundaries of suspected symptoms from complex ancient texts, rather than directly judging the standard mapping results.

[0126] The candidate symptom fragments are identified by location words, nature words, and connector words. Taking P2 (head and neck pain) as an example, "head and neck" is a location word, and "pain" is a nature word. The connection relationship is parallel and adherent. Taking P1 (fever and chills) as an example, there is no explicit connector word, but it can be identified as an implicit parallel connection feature according to the parallel habit of ancient Chinese.

[0127] Taking P3 epigastric fullness and hardness as an example, epigastric fullness is a location feature, and fullness and hardness is a combination of properties. However, since this expression is marked as a whole symptom in the terminology database, its overall priority is higher than that of the conventional split features. Through this step, the system obtains additional evidence at the structural level, which supports the subsequent residual interpretation. Semantic encoding is performed on the candidate symptom fragments to obtain the initial symptom vector.

[0128] For example, after encoding P1, we get VP1=(0.49,0.55,0.17), after encoding P2, we get VP2=(0.61,0.28,0.47), and after encoding P3, we get VP3=(0.23,0.33,0.86). The system calculates the difference between these initial vectors and the standard vectors in the ideal semantic field one by one to form a set of difference vectors. If we subtract the heat vector and the chills vector from P1 respectively, we get two difference vectors.

[0129] Subtracting the subcortex vector from P3 yields a single-unit difference value that is less than the preset difference threshold. Combining the aforementioned location features, property features, and parallel connection features, the actual residual space can be formed. In other words, the actual residual is not simply a numerical deviation, but a composite representation of numerical deviation and structural labeling.

[0130] In the anomaly handling mechanism, if the model cannot identify any candidate symptom segments in a target text slice, an adjacent slice re-splitting strategy can be adopted, such as merging S1 and S2 and identifying them again; if there are still no results, it is marked as unrecognized text and sent to the manual sampling set; if multiple overlapping candidates are identified for the same segment, such as head pain being identified as both the whole and headache, all of them are retained first, and then the coupled decision module decides whether to retain the whole or split it based on the matching of the actual residual and the theoretical residual; if the identified location words and property words conflict with the prohibited split entries in the terminology database, the prohibited split entries are given priority to avoid erroneously amplifying structural features;

[0131] For example, when the system processes ancient books on typhoid fever, if the symptoms are slight chills and fullness in the chest and hypochondrium, the system will extract three target segments: afternoon fever, slight chills, fullness in the chest and hypochondrium; then, candidate segments will be obtained through identification; the fullness in the chest and hypochondrium will be extracted as a location feature and fullness in the hypochondrium as a nature feature; finally, the actual residual will be formed, which will be used to compare with the theoretical residual that has been prepared in advance.

[0132] The purpose of this step is to project the true expression form in the ancient texts as completely as possible into the computable space, thereby achieving a joint description of semantic deviation and word formation deviation, laying the foundation for subsequent detailed judgment.

[0133] In a preferred embodiment of the present invention, determining the coupling decision result based on the real residual space and the theoretical residual space includes: performing structural similarity calculation, direct semantic similarity calculation, and transformed semantic similarity calculation on the real residual space and the theoretical residual space to determine structural similarity, direct semantic similarity, and transformed semantic similarity. The structural similarity is used to characterize the consistency between the feature extraction structure of the candidate symptom fragment and the corresponding structure template of the preset rule path in the transformed rule knowledge base. The direct semantic similarity is used to characterize the vector similarity between the initial symptom vector and the corresponding single vector in the standard symptom vector set. The transformed semantic similarity is used to characterize the vector similarity between the initial symptom vector and the corresponding simulation vector in the simulated symptom semantic field.

[0134] A comprehensive coupling value is determined based on a weighted combination of the structural similarity, the direct semantic similarity, and the transformed semantic similarity. If the comprehensive coupling value is greater than or equal to a preset high threshold, the coupling decision result is determined to be a direct standard mapping result. If the comprehensive coupling value is less than or equal to a preset low threshold, the coupling decision result is determined to be an unmatched record.

[0135] If the comprehensive coupling value is greater than the preset low threshold and less than the preset high threshold, then the coupling decision result is determined to be a composite splitting mapping result; wherein, the preset high threshold and the preset low threshold are discrimination thresholds predetermined based on the distribution of comprehensive coupling values ​​in the labeled sample set, and the preset high threshold is greater than the preset low threshold.

[0136] This embodiment provides a coupled decision mechanism; specifically, if only the direct semantic similarity between candidate segments and standard terms is compared, complex expressions such as fever and chills are easily misjudged as low-quality matches of a single symptom; if only the distance to the simulation vector is compared, some truly unknown new symptoms are easily mistaken for known variants; therefore, this embodiment introduces the joint calculation of structural similarity, direct semantic similarity and transformed semantic similarity.

[0137] To ensure that the technical meaning of the three types of similarity remains unique throughout the text, this embodiment further stipulates that: structural similarity only applies to structural information such as location words, property words, conjunction words, and disjointness markers of candidate symptom fragments, and is unrelated to vector distance;

[0138] Direct semantic similarity is calculated only between the initial symptom vector and the standard symptom vector in the ideal semantic field of standard symptoms, without introducing any simulation vectors; transformed semantic similarity is calculated only between the initial symptom vector and the simulation vector in the simulated symptom semantic field, without directly comparing it with the standard symptom vector.

[0139] If a candidate segment has multiple comparable standard symptom vectors or multiple comparable simulation vectors, then the one with the highest score in each set is selected as the input for the comprehensive coupling value of the direct semantic similarity and transformed semantic similarity of the segment. The highest score always refers to the optimal matching score of the same segment in the corresponding candidate set, without changing the name and meaning of the three types of similarity.

[0140] The specific processing procedure is as follows: Structural similarity is used to measure whether the word formation structure of the candidate segment is consistent with the rule path template. Specifically, the corresponding similarity score can be obtained by comparing the preset structural matching mapping table or rule graph. Taking fever and chills as an example, its structural feature is property word + property word, which is implicitly parallel.

[0141] If a symptom A + symptom B parallel combination template exists in the rule base, the structural consistency is considered high, and a value of 0.90 can be assigned according to the mapping table. For example, although epigastric fullness can also be mechanically separated into location + nature, this term is prohibited from being split into a whole item in the terminology database. Therefore, its structural consistency with the whole symptom template is high, triggering the whole matching rule, and a value of 0.92 can be assigned, while its consistency with ordinary parallel templates is low.

[0142] Direct semantic similarity is used to measure the closeness between the initial symptom vector and the standard symptom vector. Taking P1 fever and chills as an example, its direct similarity with fever may be 0.61, with chills 0.60, and with epigastric fullness and hardness 0.21. Transformed semantic similarity measures the closeness between the initial symptom vector and the simulation vector. If the similarity between P1 and the simulated vector of fever + chills combination reaches 0.88, it indicates that the segment is more like a rule-interpretable composite expression than a completely new standard unit.

[0143] The overall coupling value can be obtained by weighted combination of the three factors; its calculation formula can be expressed as: Overall Coupling Value = Structural similarity + Direct semantic similarity + Transform semantic similarity, where For the preset weighting coefficients, and For example, weights of 0.3, 0.3, and 0.4 can be used. For P1, the structural similarity is 0.90, the direct semantic similarity is the optimal single-unit similarity of 0.61, and the transformed semantic similarity is 0.88. Therefore, the comprehensive coupling value is 0.3×0.90+0.3×0.61+0.4×0.88=0.805.

[0144] If the preset high threshold is 0.85 and the low threshold is 0.45, then P1 enters the intermediate value zone. Therefore, according to the judgment logic defined in the embodiment, its coupling judgment result is determined to be a composite split mapping result. Taking P3 as an example, if the structural similarity is 0.92, the direct semantic similarity is 0.95, and the transformed semantic similarity is 0.78, then the comprehensive coupling value is approximately 0.877, entering the high value zone, and is therefore determined to be a direct standard mapping result. As another example, for P4, if the structural similarity is 0.38, the direct semantic similarity is 0.41, and the transformed semantic similarity is 0.36, then the comprehensive coupling value is approximately 0.381, falling into the low value zone, and is recorded as unmatched.

[0145] In practical implementation, the preset high and low thresholds can be determined in advance based on the comprehensive coupling value distribution in the labeled sample set. For example, if statistics are performed on known directly mappable samples, splittable samples, and unknown samples respectively, and if high-value samples mainly fall above 0.84, intermediate-value samples are mainly distributed between 0.46 and 0.84, and low-value samples mainly fall below 0.46, then the high threshold can be set to 0.85 and the low threshold to 0.45. This can reduce boundary jitter caused by rule assumptions.

[0146] Furthermore, to avoid ambiguity, the comprehensive coupling value in this embodiment can be understood as the basis for the diversion after the summation of the three sources of evidence: the high value area corresponds to the candidate fragments that can be stably accepted by the single standard point, the middle value area corresponds to the composite expression that is more suitable for entering the splitting process, and the low value area corresponds to the unknown expression that is difficult to explain by the existing rules.

[0147] In other words, whether to directly map in this embodiment is determined by the interval where the comprehensive coupling value is located, while which standard term to map to or which standard symptoms to break it down into will be further refined in subsequent implementations;

[0148] In the anomaly handling mechanism, if the comprehensive coupling value of a certain segment falls near the threshold, such as 0.849 or 0.451, the system can add a buffer and send it to the review queue instead of immediately giving an irreversible judgment; if the structural similarity is very high but the semantic similarity is very low, it means that it looks like a known rule in form but is not close in content. The system will first judge it as a non-match rather than an incorrect match of a rule.

[0149] If both direct semantic similarity and transformed semantic similarity are high, but the overall coupling value is still in the middle range, then the composite splitting candidate attribute is maintained first, and it will be expanded according to the composite boundary in Implementation 7, instead of falling directly into the single standard item in this step; if the transformed semantic similarity of multiple rule paths is the same or approximately the same, then the rule path with higher structural similarity is selected first; if the structural similarity is also the same, then the rule path with higher priority or more sample support times in the transformation rule knowledge base is selected first, so as to ensure that the same candidate segment outputs the same value when processed repeatedly;

[0150] For example, when the center processes ancient books on typhoid fever and febrile diseases in batches, the system calculates the comprehensive coupling value of fever and chills to be between the low and high thresholds. Therefore, the fragment is identified as an interpretable complex expression and enters the splitting process. For epigastric fullness and hardness, the comprehensive coupling value is higher than the high threshold and is directly mapped to a single standard symptom. For multiple obscure descriptions that are difficult to interpret, unmatched records are generated to provide an entry point for the subsequent discovery of new terms.

[0151] The purpose of this step is to suppress misjudgments caused by a single indicator through multi-dimensional joint evaluation, thereby achieving stable diversion of the three results: direct mapping, composite splitting, and unknown records.

[0152] In a preferred embodiment of the present invention, if the comprehensive coupling value is greater than or equal to a preset high threshold, the coupling decision result is determined to be a direct standard mapping result, including: determining the target standard symptom based on the optimal matching standard symptom vector of the initial symptom vector in the ideal semantic field of the standard symptom; and determining the direct standard mapping result based on the target standard symptom and the standard symptom terminology database.

[0153] This embodiment provides a direct standard mapping mechanism. Specifically, the previous embodiment has given the determination framework for the comprehensive coupling value and diverted the samples in the high-value area to the direct standard mapping results. On this basis, this embodiment further explains how the samples in the high-value area are mapped to specific standard terms to avoid unclear or untraceable output content in high-confidence scenarios.

[0154] The specific processing process is as follows: When the comprehensive coupling value of a certain candidate symptom segment is not lower than the preset high threshold, the system first searches for the optimal matching standard symptom vector in the ideal semantic field of standard symptoms. Here, the "optimal" can comprehensively consider the cosine similarity, Euclidean distance, and term priority label. For example, for "epigastric fullness and hardness", the initial vector VP3=(0.23, 0.33, 0.86), and the similarity with the standard vector VC=(0.22, 0.34, 0.88) is 0.95. Moreover, this standard item has the overall priority and non-splittable marks, so its optimal standard symptom is determined to be "epigastric fullness and hardness".

[0155] After determining the target standard symptom, the direct standard mapping result is output according to the term library. The output content can not only include the standard name, but also carry the standard code, the symptom category it belongs to, the definition description, and the source ancient text segment. For example, the returned result includes: the original segment "epigastric fullness and hardness"; the standard term "epigastric fullness and hardness"; the standard code SYM-02318; the mapping method direct; the source slice number S2. This is convenient for subsequent knowledge retrieval and audit tracking.

[0156] To illustrate this process, assume that a certain candidate segment Q1 is "aversion to cold with shivering". The similarity between its direct vector and the standard "aversion to cold" is 0.91, and the similarity with other standard terms is lower than 0.60. At the same time, its comprehensive coupling value reaches 0.88. Although "aversion to cold with shivering" is literally different from "aversion to cold" on the text surface, since its best single-body match in the ideal semantic field is still "aversion to cold", the standard result "aversion to cold" is output instead of retaining the original word or splitting it.

[0157] Furthermore, this embodiment is applicable to the candidate segments that have been determined to enter the high-value area by the previous embodiment. If a certain candidate segment is relatively close to a certain composite simulation path, but its comprehensive coupling value still falls between the low threshold and the high threshold, it does not enter this embodiment but is instead processed by the composite splitting mapping process. Through this front-to-back connection, it is possible to avoid the logical intersection between the direct mapping in the high-value area and the splitting mapping in the intermediate-value area.

[0158] In the exception handling mechanism, if the difference between the similarities of two optimal standard symptoms in the high-value area is less than the preset difference threshold, for example, the difference is less than 0.01, then further reference is made to the non-splittable mark, the upper symptom category, and the context part constraint for disambiguation.

[0159] If they still cannot be distinguished, they are temporarily saved as dual-candidate direct mapping records and handed over to manual review; if a candidate segment is highly similar to a certain standard point, but the standard point is in a discontinued or historical version state, the system will prioritize mapping to the current valid version terminology and retain the old name as a historical alias in the output.

[0160] If a valid code for the target standard symptom cannot be found in the terminology database, only the standard name and temporary internal number are returned to avoid processing link interruption;

[0161] For example, when the system processes ancient books on typhoid fever, the fragment "heart lower back pain" enters the high-value area. The system retrieves the optimal standard point in the ideal semantic field, which is still "heart lower back pain", and then directly generates a standard mapping record. Similarly, although the fragment "aversion to cold" comes from an ancient alternative name, it is also uniformly mapped to the standard term "aversion to cold".

[0162] The purpose of this step is to quickly complete the standardized table in high-confidence scenarios, reduce unnecessary splitting and repetitive reasoning, and thus achieve stable and traceable standard symptom output.

[0163] In a preferred embodiment of the present invention, if the comprehensive coupling value is greater than the preset low threshold and less than the preset high threshold, then the coupling decision result is determined to be a composite splitting mapping result, including: extracting the location features, property features, and parallel connection features corresponding to the candidate symptom fragment from the real residual space; determining the composite boundary based on the location features, property features, and parallel connection features; segmenting the candidate symptom fragment according to the composite boundary to determine the split symptom sequence; and determining the composite splitting mapping result based on the split symptom sequence and the standard symptom terminology database.

[0164] This embodiment provides a composite splitting and mapping mechanism. Specifically, within the intermediate value range, the system already knows that the candidate fragment is similar to certain regularized simulation states, but it cannot be directly placed on the table as a single standard item. If the splitting step is missing, composite expressions such as fever, chills, chest and rib fullness will be stored as a whole, affecting the consistency of subsequent retrieval. Therefore, this embodiment performs boundary recognition and splitting mapping on such fragments.

[0165] The specific processing procedure is as follows:

[0166] Read the corresponding location features, property features, and parallel connection features from the real residual space; for example, fever and chills can be extracted as two property features, fever and chills, and there is an implicit parallel connection.

[0167] Head and neck pain can be extracted by identifying two site features of the head and neck and two nature features of pain; chest and rib fullness can be extracted by identifying chest and rib fullness as a complex site feature and fullness as a combination of nature features; the system does not always cut according to the number of characters, but first observes the combination relationship of these structural features; the complex boundary is determined based on the above features;

[0168] In a simplified example of boundary determination, if a property word + property word is detected and the two correspond to two independent standard symptoms in the rule base, then a boundary is set between them; if a location word 1 + location word 2 + property word is detected and the rule base supports apportioned properties, then a boundary is set according to the pattern of headache item strength.

[0169] If a forbidden phrase in the overall dictionary is hit, no boundary is set; taking fever and chills as an example, the system sets a boundary between fever and chills to get two segments; taking epigastric fullness and hardness as an example, although the location and nature can be identified, the boundary number is 0 because the forbidden phrase is hit, so no splitting is performed.

[0170] Candidate segments are segmented based on the composite boundary to determine the split symptom sequence; for example, fever and chills are split into R1 fever and R2 chills; headache and neck stiffness can be split into R3 headache and R4 neck stiffness according to the rules; chest and rib fullness can be split into chest fullness or retained as chest and rib fullness in some knowledge base versions, depending on whether the chest and ribs are considered fixed composite sites in the rule base; after splitting, each segment is matched with the standard symptom terminology database to form the composite splitting mapping result;

[0171] For ease of explanation, the comprehensive coupling value of the sample in the intermediate range of fever and chills can be set at 0.78, indicating that it is close to the composite rule but not sufficient to be directly mapped as a single standard item. After the system extracts two property features and confirms the existence of an implicit parallel relationship, it defines the boundary and outputs the standard symptom sequence [fever, chills]. For example, the comprehensive coupling value of headache and neck stiffness is 0.73. The system identifies the property combination of two parts of the head and neck and the stiffness, and outputs [headache, neck stiffness] according to the location allocation + property mapping rule in the rule base.

[0172] In the exception handling mechanism, if there is more than one boundary candidate, for example, a segment can be split into two or three segments, the scheme that maximizes the average matching degree of each segment in the terminology database is selected first. If the average matching degree of all splitting schemes is lower than the preset value, it is reverted to an unmatched record instead of being forcibly split. If an empty segment, a single field, or a fragment without a valid term appears after splitting, the fragment is merged into the adjacent segment for recalculation. If recalculation still fails, the entire record enters the unmatched path. If a segment hits both the overall disjoint dictionary and the composite rule template, the overall disjoint dictionary takes precedence.

[0173] For example, when the system processes ancient books on typhoid fever, fever and chills are identified by the system as a compound expression in the intermediate value area and are eventually separated into fever and chills; while epigastric fullness and hardness can be separated into location and nature on the surface, but since it has been clearly registered in the overall dictionary, it is not separated to avoid forming an incorrect epigastric fullness and hardness record.

[0174] The purpose of this step is to provide a stable, structured approach for interpretable but not directly mapped classical Chinese expressions, thereby enabling the standardized conversion of complex symptoms into standard symptom sequences.

[0175] In a preferred embodiment of the present invention, clustering and inductive processing is performed on the unmatched records to determine candidate results for new symptoms, including: clustering the residual vectors in the unmatched records to determine candidate clusters; using the preset large language model, performing semantic inductive processing on the candidate clusters to determine cluster labels and definition descriptions representing the common semantics of the candidate clusters; and determining the candidate results for new symptoms based on the cluster labels and definition descriptions.

[0176] This embodiment provides a mechanism for discovering new symptom candidates. Specifically, the aforementioned process has already precipitated fragments that cannot be fully explained by existing rules as unmatched records. However, if these records are scattered and stored for a long time without being summarized, they cannot form knowledge increments. Therefore, this embodiment forms verifiable new symptom candidates by clustering residual vectors and combining them with semantic induction.

[0177] The specific processing steps are as follows: Clustering is performed on the residual vectors in the unmatched records; the clustered objects are not the original text, but the uninterpreted semantic offsets retained after dual-track judgment; for example, the system accumulates 5 unmatched segments: U1 joint pain, U2 pain in all joints, U3 soreness in limbs, U4 pain between bones, U5 heart pain; among them, the residual vectors of U1 to U4 are close to each other, while U5 is significantly far away; after density clustering, candidate cluster C1={U1,U2,U3,U4} can be obtained, and U5 forms a discrete point or a separate small cluster C2;

[0178] The pre-defined large language model is used to semantically summarize each candidate cluster. The input content can include original fragments within the cluster, high-frequency location words, high-frequency property words, and several reference items that are closest to the existing terminology database but do not reach the threshold. For C1, the large language model can summarize the common semantics as symptoms of pain and soreness in multiple joints and limbs, thereby generating the cluster label "joint pain-like symptoms" and a brief definition describing pain and soreness in bones, limbs, or all joints, which is difficult to classify into a single local pain. For C2 "heart-wrenching pain", if the sample size is insufficient, the model only outputs a descriptive summary and does not immediately assign a stable label.

[0179] Based on the cluster labels and definition descriptions, the candidate results for new symptoms can be determined; here, "new" does not mean that it will immediately become a formal standard term, but rather that it will enter the candidate pool as a knowledge review object;

[0180] Candidate results may include: candidate name, definition description, representative original text example sentences, relevant residual center vector, list of nearest standard symptoms, and frequency of occurrence; for example, for C1, candidate item joint pain-like symptoms can be generated, along with 4 classical Chinese example sentences and residual center vector;

[0181] In the anomaly handling mechanism, if there is a large semantic discrepancy within a candidate cluster, such as both joint pain and chest and rib pain in the same cluster, it indicates that the clustering granularity is too coarse, and the system should increase the clustering threshold and re-cluster; if the number of samples in a cluster is less than the preset lower limit, such as less than 3, it will not be directly formed as a new candidate, but will be retained as an observation cluster.

[0182] If the cluster labels generated by the large language model highly overlap with existing standard terms, such as generating chills-like symptoms, but the system verification finds that a chills standard item already exists, then the cluster is marked as a rule missing problem and is handled by the rule base update process instead of adding new symptoms.

[0183] For example, after processing hundreds of medical records of febrile diseases at the center, the system found that many unmatched records had co-occurrence patterns such as joints, joints, limbs and pain, soreness and pain. Through clustering and semantic induction, new candidate joint pain-like symptoms were formed. The candidate was entered into the expert review stage to determine whether the terminology database needed to be expanded or only alternative names and mapping relationships needed to be added.

[0184] The purpose of this step is to transform scattered, unmatched anomaly expressions into reviewable, definable, and sustainably accumulating knowledge candidates, thereby enabling the system to iteratively expand its knowledge of symptoms from ancient texts.

[0185] In a preferred embodiment of the present invention, updating the conversion rule knowledge base based on the newly added symptom candidate results includes: determining new alternative name rules, new compound word formation rules, and new semantic mapping relationships based on the newly added symptom candidate results; writing the new alternative name rules, the new compound word formation rules, and the new semantic mapping relationships into the conversion rule knowledge base; and determining the updated conversion rule knowledge base.

[0186] This embodiment provides a mechanism for updating a conversion rule knowledge base. Specifically, if a new candidate is only output for manual review and is not updated to the underlying rule system, the same mismatch problem will still occur repeatedly when the system processes similar texts in the next round. Therefore, this embodiment converts the approved new candidates into rule increments to achieve closed-loop updates.

[0187] The specific processing procedure is as follows: Based on the results of the newly added symptom candidates, the system determines which type of rule should be formed; if the candidate symptom is actually an ancient alternative name for a known standard symptom, then a new alternative name rule is generated; for example, if clustering and induction find that aversion to cold repeatedly and stably corresponds to chills, then aversion to cold → chills is written as an alternative name mapping rule; if the candidate reflects a new compound word formation method, then a new compound word formation rule is generated.

[0188] For example, if a large number of ancient books find that chest and rib fullness should usually be broken down into chest fullness + rib fullness, then corresponding boundary rules and location allocation rules can be added; if the candidate expression is a new semantic unit that has not been fully defined before but has a stable adjacency relationship with the existing standard symptoms, then a new semantic mapping relationship can be established to record its close neighbor relationship or superior affiliation with the existing standard system.

[0189] When writing to the knowledge base, it's not simply about appending a line of text, but rather creating a structured rule record. A rule record includes at least the rule type, triggering condition, applicable priority, source cluster number, effective version, and rollback flag. For example, the rule "Aversion to Cold" can be recorded as an alias, with a medium applicable priority, sourced from cluster C17, and version V3.2; the rule "Fullness in the Chest and Ribs" can be recorded as a composite splitting rule, along with a dictionary of prohibited splitting conditions that were not met. This facilitates subsequent rule conflict management and version control.

[0190] To illustrate the update process, we can assume that the newly added candidate H1, joint pain-like symptoms, is deemed by experts to be temporarily unsuitable as a formal standard symptom, but can be closely related to joint pain and limb soreness, so a new semantic mapping relationship is written; and we can assume that the newly added candidate H2, aversion to cold, is verified to be an ancient alternative name for chills, so a new alternative name rule is written.

[0191] If the newly added candidate H3 head item strong pain stably corresponds to headache + item strong in the specified classical system, then it is written into the new compound word formation rule; in this way, when processing similar texts in the next round, the system can directly use these new rules to generate more accurate theoretical residuals;

[0192] In the exception handling mechanism, if a new rule conflicts with an existing rule, for example, if an existing rule treats a certain expression as a whole while the new rule attempts to split it, the system will not automatically override it. Instead, it will enter the conflict arbitration table, where experts will select the priority or limit the scope of applicable literature.

[0193] If the supporting samples for a new rule are insufficient, such as only one candidate cluster and two samples, it will be written in a trial state and will not be immediately included in the main version of the official rule library. If backtracking verification after writing reveals that the new rule has caused an increase in the misjudgment rate of historical samples, the rule version will be revoked by marking it as rolled back.

[0194] For example, after the center completed the collation of a batch of ancient books on febrile diseases, experts confirmed that aversion to cold should be regarded as an ancient alternative name for aversion to cold, and that headache and stiff neck in most cases correspond to headache + stiff neck in this literature system.

[0195] Therefore, the system writes these two types of knowledge into the alternative name rules and compound word formation rules respectively; when processing similar documents again in the future, the same expression will be more easily interpreted as a known offset, rather than repeatedly falling into the unmatched set;

[0196] The purpose of this step is to continuously feed the incremental knowledge obtained from manual review and machine summarization back into the underlying rule system, thereby achieving self-iterative improvement of the system's recognition capabilities.

[0197] In a preferred embodiment of the present invention, the original text data is ancient Chinese medicine text data, wherein the ancient Chinese medicine text data includes text without punctuation, text with variant characters, and ancient Chinese symptom description text; the standard symptom terminology database contains pre-constructed modern Chinese medicine standard symptom ontology data; the conversion rule knowledge base includes ancient and modern Chinese lexical rules, compound word formation rules, location substitution rules, and measurement expression rules.

[0198] This embodiment provides a data organization mechanism suitable for the scenario of ancient Chinese medicine books; specifically, in the aforementioned scheme, the source and type of data directly affect the slice quality, rule injection effect and final mapping accuracy; if the boundaries of ancient book data, standard terminology ontology and rule knowledge base are not specifically set, the system is difficult to operate stably, so this embodiment further explains the applicable data types;

[0199] The specific processing procedure is as follows: The original text data is preferably selected from ancient Chinese medicine texts, including texts without punctuation, texts with variant characters, and texts describing symptoms in classical Chinese. Texts without punctuation refer to the original transcription that retains the continuous writing form of ancient texts, such as fever, chills, headache, and neck stiffness in Taiyang disease. Texts with variant characters refer to texts that retain or annotate variant characters and phonetic loan characters, such as the coexistence of ribs and slits.

[0200] Classical Chinese symptom description text refers to phrases with the characteristics of compressed expression in classical Chinese, such as "heartache and restlessness, insomnia, and chest pain." This type of data is different from modern medical record descriptions and cannot be directly applied to modern word segmentation rules. Therefore, it is more suitable to process it using the above-mentioned slicing, rule injection, and residual coupling methods.

[0201] The standard symptom terminology database preferably includes pre-built modern TCM standard symptom ontology data; this ontology data may include standard term names, definitions, superordinate categories, alternative names, disambiguation prohibition markers, and term codes; for example, aversion to cold belongs to the cold and heat category of symptoms, and epigastric fullness and hardness belong to the chest and abdomen category and have disambiguation prohibition markers; through ontology organization, the system can not only obtain standard names, but also use hierarchical relationships to perform disambiguation and anomaly tolerance processing;

[0202] The conversion rule knowledge base preferably includes ancient and modern Chinese word - formation rules, compound word - formation rules, part - denotation rules, and measurement expression rules. Among them, the ancient and modern Chinese word - formation rules are used to handle alias mappings such as "憎寒" → "恶寒"; the compound word - formation rules are used to handle juxtaposed or glued expressions such as "发热恶寒头项强痛".

[0203] The part - denotation rules are used to handle the correspondence between ancient Chinese part expressions such as "心下", "胁下", "少腹" and modern standard part words; the measurement expression rules are used to explain the offset effect of degree words or quantifiers such as "微恶寒", "大热", "少腹满" on the semantic center. Through these rules, the system can generate a more accurate simulation semantic field and theoretical residual space.

[0204] In a simplified example, assume that the original text input is "少腹急结发热恶寒", and there are already three standard terms "少腹拘急发热恶寒" in the term base. The rule base contains that "少腹" is a part - denotation, "急结" is a property combination, and "发热恶寒" is a juxtaposed compound rule. The system can first split the original text into two candidate segments and then enter the real - world residual extraction and coupling judgment processes respectively.

[0205] In the exception - handling mechanism, if the original text contains both scanning recognition errors and mixed variant characters, for example, "惡寒" is recognized as "悪寒" or "愕寒", it will first go through basic character normalization and variant - character mapping and then enter the slicing module. If some necessary fields are missing in the standard term ontology, for example, the split - prohibition marker is missing, the term is defaulted to be processed as a splittable candidate, but a boundary - unconfirmed marker will be attached to the output record. If the rule knowledge base does not yet have part - denotation rules specific to a certain type of literature, the system will appropriately increase the unmatched retention rate in the corresponding literature batch to avoid misapplying rules of other literature systems.

[0206] Exemplarily, when the system processes ancient medical books of the "Shanghan" category, the system first unifies the variant characters, then uses the modern standard symptom ontology of traditional Chinese medicine as an ideal benchmark, and completes the whole set of standardization processes in combination with rules such as ancient and modern word - formation, compound word - formation, part - denotation, and measurement expression.

[0207] For ancient Chinese symptom descriptions such as "微恶寒大热少腹急结", the system can obtain a consistent calculation entry under the above - mentioned data organization framework.

[0208] The purpose of this step is to clarify the data boundaries and knowledge sources adapted by the present invention, so as to achieve the stable implementation of the entire standardization process in the real application environment of ancient Chinese medical books.

[0209] The above has described an embodiment of the present invention in detail, but the described content is only a preferred embodiment of the present invention and cannot be considered as limiting the scope of implementation of the present invention. All equivalent changes and improvements made according to the scope of the application of the present invention should still fall within the patent coverage scope of the present invention.

Claims

1. A method for standardizing symptoms in ancient Chinese medical texts based on a large language model, characterized in that, The method includes: Acquire a standard symptom terminology database, a conversion rule knowledge base, and raw text data; The standard symptom terminology database is subjected to terminology normalization and semantic encoding to determine the standard symptom vector set, and an ideal semantic field of standard symptoms is constructed based on the standard symptom vector set. The standard symptom ideal semantic field and the transformation rule knowledge base are subjected to rule injection processing to determine the simulated symptom semantic field; and the theoretical residual space is determined based on the set of difference vectors between the simulated vectors in the simulated symptom semantic field and the standard symptom vectors in the standard symptom ideal semantic field. The original text data is processed into text segments to determine target text segments, and a preset large language model is used to perform symptom recognition processing on the target text segments to determine candidate symptom fragments. The candidate symptom fragments are semantically encoded to determine the initial symptom vector, and the real residual space is determined based on the set of difference vectors between the initial symptom vector and the standard symptom vector in the ideal semantic field of the standard symptom. Based on the actual residual space and the theoretical residual space, the coupling decision result is determined, which includes the direct standard mapping result, the composite split mapping result, and the unmatched record; The unmatched records are clustered and summarized to determine new symptom candidate results, and the conversion rule knowledge base is updated based on the new symptom candidate results.

2. The method for standardizing symptoms in ancient Chinese medical texts based on a large language model according to claim 1, characterized in that, The process of performing terminology normalization and semantic encoding on the standard symptom terminology database to determine a set of standard symptom vectors, and constructing an ideal semantic field for standard symptoms based on the set of standard symptom vectors, includes: The standard symptom terminology database is subjected to terminology standardization processing to determine the standard symptom set; The standard symptom set is semantically encoded to determine the standard symptom vector set; Construct the ideal semantic field of standard symptoms based on the set of standard symptom vectors.

3. The method for standardizing symptoms in ancient Chinese medical texts based on a large language model according to claim 1, characterized in that, The step involves injecting rules into the standard symptom ideal semantic field and the transformation rule knowledge base to determine the simulated symptom semantic field, wherein the simulated symptom semantic field consists of multiple simulated vectors; and the theoretical residual space is determined based on the set of difference vectors between the simulated vectors in the simulated symptom semantic field and the standard symptom vectors in the standard symptom ideal semantic field, including: Based on the ancient and modern terminology conversion rules in the conversion rule knowledge base, the standard symptom vector in the standard symptom ideal semantic field is converted to determine the conversion simulation vector; Based on the compound word formation rules in the transformation rule knowledge base, the standard symptom vectors in the standard symptom ideal semantic field are combined to determine the compound simulation vector; The simulated symptom semantic field is determined based on the transformed simulation vector and the composite simulation vector; The theoretical residual space is determined based on the difference vector between the simulation vector in the simulated symptom semantic field and the corresponding standard symptom vector in the standard symptom ideal semantic field.

4. The method for standardizing symptoms in ancient Chinese medical texts based on a large language model according to claim 1, characterized in that, include: The original text data is sliced ​​to determine the target text slice; Using the preset large language model, symptom recognition processing is performed on the target text slice to determine candidate symptom segments, wherein the preset large language model is a large language model that has been finely adjusted by the corpus of ancient Chinese medicine texts. The candidate symptom fragments are subjected to location word recognition, property word recognition, and connector word recognition to determine location features, property features, and parallel connection features; The candidate symptom fragments are semantically encoded to determine the initial symptom vector; The real residual space is determined based on the set of difference vectors between the initial symptom vector and the standard symptom vector in the standard symptom ideal semantic field, combined with the location feature, the property feature and the parallel connection feature.

5. The method for standardizing symptoms in ancient Chinese medical texts based on a large language model according to claim 1, characterized in that, The step of determining the coupling decision result based on the actual residual space and the theoretical residual space includes: Structural similarity, direct semantic similarity, and transformed semantic similarity are calculated for the real residual space and the theoretical residual space to determine structural similarity, direct semantic similarity, and transformed semantic similarity. The structural similarity characterizes the consistency between the feature extraction structure of the candidate symptom fragment and the corresponding structural template of the preset rule path in the transformed rule knowledge base. The direct semantic similarity characterizes the vector similarity between the initial symptom vector and the corresponding single vector in the standard symptom vector set. The transformed semantic similarity characterizes the vector similarity between the initial symptom vector and the corresponding simulation vector in the simulated symptom semantic field. The comprehensive coupling value is determined based on a weighted combination of the structural similarity, the direct semantic similarity, and the transformed semantic similarity. If the overall coupling value is greater than or equal to a preset high threshold, the coupling decision result is determined to be a direct standard mapping result; if the overall coupling value is less than or equal to a preset low threshold, the coupling decision result is determined to be an unmatched record; if the overall coupling value is greater than the preset low threshold and less than the preset high threshold, the coupling decision result is determined to be a composite split mapping result. The preset high threshold and the preset low threshold are discrimination thresholds predetermined based on the comprehensive coupling value distribution in the labeled sample set, and the preset high threshold is greater than the preset low threshold.

6. The method for standardizing symptoms in ancient Chinese medical texts based on a large language model according to claim 5, characterized in that, If the overall coupling value is greater than or equal to a preset high threshold, then determining the coupling decision result as a direct standard mapping result includes: The target standard symptom is determined based on the optimal matching standard symptom vector in the ideal semantic field of the standard symptoms from the initial symptom vector. The direct standard mapping result is determined based on the target standard symptoms and the standard symptom terminology database.

7. The method for standardizing symptoms in ancient Chinese medical texts based on a large language model according to claim 5, characterized in that, If the overall coupling value is greater than the preset low threshold and less than the preset high threshold, then the coupling decision result is determined to be a composite splitting mapping result, including: Extract the location features, property features, and parallel connection features corresponding to the candidate symptom fragments from the real residual space; The composite boundary is determined based on the location characteristics, the property characteristics, and the parallel connection characteristics; Based on the composite boundary, the candidate symptom fragments are segmented to determine the split symptom sequence; The composite splitting mapping result is determined based on the splitting symptom sequence and the standard symptom terminology database.

8. The method for standardizing symptoms in ancient Chinese medical texts based on a large language model according to claim 1, characterized in that, The process of clustering and summarizing the unmatched records to determine candidate results for new symptoms includes: Clustering is performed on the residual vectors of the unmatched records to determine candidate clusters; Using the preset large language model, the candidate clusters are semantically summarized to determine the cluster labels and definition descriptions that represent the common semantics of the candidate clusters; Based on the cluster labels and their definitions, the candidate results for the new symptoms are determined.

9. A method for standardizing symptoms in ancient Chinese medical texts based on a large language model, as described in claim 1, is characterized in that... The step of updating the conversion rule knowledge base based on the newly added symptom candidate results includes: Based on the newly added symptom candidate results, new alternative name rules, new compound word formation rules, and new semantic mapping relationships are determined; The newly added alternative name rules, the newly added compound word formation rules, and the newly added semantic mapping relationships are written into the conversion rule knowledge base to determine the updated conversion rule knowledge base.

10. A method for standardizing symptoms in ancient Chinese medical texts based on a large language model, as described in claim 1, is characterized in that... The original text data is ancient Chinese medicine text data, which includes text without punctuation, text with variant characters, and text describing symptoms in classical Chinese. The standard symptom terminology database contains pre-built ontology data of modern TCM standard symptoms; The conversion rule knowledge base includes ancient and modern Chinese morphological rules, compound word formation rules, part-of-speech substitution rules, and measurement expression rules.

Images