Cultural field-oriented corpus data generation method

Abstract

The invention discloses a cultural field-oriented corpus data generation method. The method comprises the following steps: S1, constructing a triple set based on a data text of a cultural field; s2, generating a multi-style corpus set by the triple through a basic template and a large language model; s3, performing style transformation on the multi-style corpus based on a manually defined rule to generate a basic corpus set; and S4, constructing culture corpus data by using the basic corpus which simultaneously meets double verification of structure alignment and semantic consistency. The culture corpus can be generated accurately, controllably and diversely.

Description

Technical Field

[0001] This invention relates to the field of computer natural language artificial intelligence, and in particular to a method for generating corpus data for the cultural domain. Background Technology

[0002] Language models are natural language processing models trained on large-scale corpora using deep learning techniques. By learning from a vast amount of data, the quality and effectiveness of large language models can be improved. However, if the training of large language models lacks structural knowledge support, it is easy to produce illusions or miss cultural elements. Currently, the application of large language models in the cultural field is becoming increasingly widespread. However, the sources of corpus data are mixed, and accuracy is difficult to guarantee; many knowledge points are expressed in a single way, and the language style is lacking; although there are existing text generation and pre-training methods, there is a lack of structure-driven enhancement mechanisms for cultural knowledge graphs, as well as a lack of semantic accuracy verification systems. There is an urgent need for a technical method that can combine knowledge graphs to generate accurate, controllable, and diverse corpora. Summary of the Invention

[0003] The purpose of this application is to provide a method for generating corpus data in the cultural field, aiming to solve the technical problems of existing corpora having a single language style and low semantic accuracy.

[0004] To achieve the above objectives, this application proposes a method for generating corpus data in the cultural domain, characterized in that the method includes: S1. Construct a set of triples based on textual data from the cultural domain; S2. The triples generate multi-style corpus sets through the basic template and the large language model, respectively; S3. Based on manually defined rules, perform style transformation on multi-style corpora to generate a basic corpus set; S4. Construct cultural corpus data from the basic corpus that simultaneously satisfies both structural alignment and semantic consistency checks.

[0005] Preferably, step S1 includes: S11. Obtain and clean the text data in the cultural field; S12. Identify cultural entities in the cleaned text; S13. Construct a set of triples by extracting the relationships between cultural entities in the text.

[0006] Preferably, step S13 includes: S131. Extract entity pair relation sets using either supervised relation extraction or open information extraction methods based on the text type; S132. Iterate through all entity pairs and their relations, and combine them into a set of triples; S133. The confidence of triples is predicted by the pre-trained model Bert, and triples with confidence exceeding the threshold are retained.

[0007] Preferably, step S12 includes: Cultural entities in the cleaned text are identified using the NER model; semantic similarity is used to normalize these cultural entities.

[0008] Preferably, the semantic similarity method includes: generating word vectors for cultural entities using an embedding method, and calculating the semantic similarity between any two entities based on cosine similarity. : ; in: : Semantic vectors generated based on embedding for cultural entities; The closer the value is to 1, the more similar the semantics. Set a phase velocity threshold δ, if ,Will If the corresponding entities are the same entity, then they are merged; If the merged entity has a standard entry in the cultural dictionary, it is named according to the standard entry; otherwise, the entity is named using the expression that appeared most frequently or was shortest before the merge.

[0009] Preferably, step S2 includes: The triples were converted into different corpora using templates and large models respectively. The basic templates were generated into different template sets through manual definition and automatic induction methods. The large models included ChatGPT, DeepSeek, and Qwen3.

[0010] Preferably, the structural alignment verification method in step S4 includes: Extract triples from the basic corpus again, and determine whether the triples exist in the set of triples generated in step S1. If they do, the basic corpus satisfies the structural alignment check.

[0011] Preferably, the semantic consistency verification method in step S4 includes: The triples generated in step S1 are converted into natural language summaries. The cosine similarity between the summation and the basic corpus is calculated using sentence vector encoding. If the similarity is greater than a set threshold, the basic corpus satisfies the semantic consistency check.

[0012] One or more technical solutions proposed in this application have at least the following technical effects: This invention classifies texts based on differences in structure and language style, and selects appropriate relation extraction methods. By first classifying the texts and then selecting different extraction methods, the accuracy and diversity of the extraction results can be improved. Since the same entity may have different spellings, problems such as redundant entity recognition, duplicated knowledge structures, and ambiguous relation extraction occur, which affect the quality of subsequent knowledge graph construction and the consistency of corpus generation. After named entity recognition is completed, this invention introduces an entity normalization processing mechanism. Based on semantic similarity vector merging and other methods, entities with different expressions but the same semantics are uniformly represented as standardized terms, forming a unique set of cultural entities. By constructing semantic-driven templates based on basic templates and large language models, and integrating artificial language design with real-world corpus summarization, we provide a basic structure for cultural content control, style control, and language security for subsequent language model generation. Based on a combination of scenario-oriented, user-segmented, corpus-based, and manual verification methods, a batch of style tags with clear structure, stable expression, effective control, and strong scalability were selected to improve the style diversity and controllability of the corpus. By using dual verification of structural alignment and semantic consistency to filter the basic corpus, the accuracy and semantic consistency of cultural corpus data are improved. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0014] Figure 1 This is an embodiment of the present invention: a method for generating corpus data in the cultural field; The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0015] 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 a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0016] It should be noted that if the embodiments of the present invention involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of the components in a specific posture. If the specific posture changes, the directional indicators will also change accordingly.

[0017] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the use of "and / or" or "and / or" throughout the text includes three parallel solutions. For example, "A and / or B" includes solution A, solution B, or a solution where both A and B are satisfied simultaneously. Furthermore, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0018] This invention provides a method for generating corpus data in the cultural field, aiming to solve the technical problems of existing corpora having a single language style and low semantic accuracy.

[0019] Example 1 like Figure 1 A flowchart illustrating a method for generating corpus data in the cultural domain is shown. To address the technical problems of corpus data exhibiting limited expression and a lack of diverse linguistic styles, this embodiment provides a method for generating corpus data in the cultural domain, comprising the following steps: S1 Constructs a set of triples based on textual data from the cultural domain; First, based on data texts from the cultural field (including but not limited to historical figures, festival customs, intangible cultural heritage, ancient literature, ethnic rituals, cultural landmarks, etc.), a set of triplets is constructed.

[0020] Specifically, it includes: S11. Obtain and clean the text data in the cultural field; The text cleaning process includes removing noisy content (such as advertisements and meaningless repetitions), segmenting and sentence-by-sentence, Chinese word segmentation and part-of-speech tagging, providing clean and parsable language input for entity recognition.

[0021] S12 After the cultural entity recognition of the cleaned text is completed, the cultural entities are normalized. S121 uses the NER (Named Entity Recognition) model to identify cultural entities in the cleaned text.

[0022] Cultural entity recognition covers a variety of cultural named entities, such as: people, festivals, cultural relics, places, concepts, etc. Each recognition result includes: entity name, location of occurrence, entity category.

[0023] Because the same entity may have different spellings, problems such as redundant entity recognition, repetitive knowledge structures, and ambiguous relation extraction arise, affecting the quality of subsequent knowledge graph construction and the consistency of corpus generation. Therefore, after named entity recognition is completed, this invention introduces an entity normalization mechanism. Through methods such as vector merging based on semantic similarity, entities with different expressions but the same semantics are uniformly represented as standardized terms, forming a unique set of cultural entities.

[0024] S122 The method based on semantic similarity is used to normalize the identified cultural entities.

[0025] To eliminate ambiguity and repetition caused by inconsistent entity representation methods, this invention employs an entity normalization method including: Among them, the semantic similarity method is one of the key normalization mechanisms in this invention, which is applicable to the automatic merging of non-standard spellings such as aliases, colloquialisms, and abbreviations of cultural entities.

[0026] Let the set of candidate cultural entities be: ; The system will select each candidate entity Perform word vector representation based on embedding It also calculates the semantic similarity between any two entities based on cosine similarity: ; in: : Semantic vectors generated based on embedding for cultural entities; The closer the value is to 1, the more similar the semantics.

[0027] By setting a similarity threshold δ∈[0.85,0.95], a value is selected as the entity normalization criterion based on the actual needs of the project. Then For the same entity, merge them. If there is a standard entry in the cultural dictionary, the merged entity will be named according to the standard name. If there is no standard name, the entity will be named according to the most frequent or shortest expression.

[0028] We obtain a unified cultural entity set E: ; Each of these belongs to a specific type of cultural entity (such as "festival", "cultural relic", "person").

[0029] For example: The original text is "On the Dragon Boat Festival, people commemorate the patriotic poet Qu Yuan. The Dragon Boat Festival is widely celebrated in southern China." After NER model analysis, the result is: (Dragon Boat Festival, Qu Yuan, Southern China) By using normalization methods, variants such as synonyms, aliases, and abbreviations are unified into standard terms: Original entity Standardized entity name illustrate Dragon Boat Festival Dragon Boat Festival "Duanwu Festival" is another name for "Dragon Boat Festival". Southern China Southern region Standardize the description of geographical scope (optional operation) Finally, the standardized entity set is deduplicated to obtain the final unique entity set E: E = {Dragon Boat Festival, Qu Yuan, Southern Regions} S13 constructs a set of triples T by extracting the relationships R between entities e in the text: ; The specific method is as follows: S131 This invention classifies texts based on differences in structure and language style, and selects appropriate relation extraction methods. By first classifying the text and then selecting different extraction methods, the accuracy and diversity of the extraction results can be improved: for texts with standardized formats and clear relations, rule-based or syntactic dependency-based extraction methods are used; for colloquial or natural narrative texts, pre-trained language models are used for supervised relation prediction; and for short-sentence tag texts, Open Information Extraction (OpenIE) methods such as word vector similarity or co-occurrence analysis are used to extract implicit relations, thereby constructing a complete entity relation set R. ; S132 Obtaining the entity set and entity relationship set R Then, iterate through all possible entity pairs. Based on the entity relationships identified in the relation extraction stage Combine them into structured triples ( h , r , t ),in r The semantic relationship label between the two entities is set (e.g., "originating from", "commemorating", "belonging to"). If multiple relationships exist between a pair of entities, multiple triples can be constructed; if there is no significant relationship, the entity pair will not be included in the construction.

[0030] Since the entity pair relations are the set of all entity pair relations, it is not necessarily appropriate to combine all relations and entity pairs. Therefore, it is necessary to filter the combined triples.

[0031] S133 uses the pre-trained model Bert to predict the confidence of triples and retains combinations with higher confidence.

[0032] Specifically, when text / sentences are input into the BERT model, they are first segmented into a sequence of tokens.

[0033] For example, input the text / sentence: "The Dragon Boat Festival is a traditional festival established to commemorate Qu Yuan" into the BERT model.

[0034] The sentence was segmented into the following token sequence: [CLS] [E1]Dragon Boat Festival[ / E1] is a traditional festival established to commemorate [E2]Qu Yuan[ / E2]. [SEP] in: - [CLS] and [SEP] are special tokens used to mark the beginning and end of a sentence.

[0035] - [E1]...[ / E1] marks the main entity, [E2]...[ / E2] marks the secondary entity.

[0036] The model then maps the entire token sequence into a vector sequence, with each token (including [CLS]) corresponding to a vector representation.

[0037] The vector corresponding to [CLS] is often used as a semantic representation of the entire sentence.

[0038] During the relation extraction phase, the BERT model predicts the relationship R between entity pairs (primary entity, secondary entity). And output the relationship distribution. For example: (Dragon Boat Festival, Qu Yuan) → {Commemoration: 0.91, Belongs to: 0.03, Unrelated: 0.06} This yields the triplet: That is, (Dragon Boat Festival, Commemoration, Qu Yuan), and record its confidence level as 0.91. If the confidence threshold is set to 0.85, the confidence level of this triple exceeds the threshold, so it is retained.

[0039] If the confidence level of (Dragon Boat Festival, belongs to, Qu Yuan) and (Dragon Boat Festival, has no relation to, Qu Yuan) is below the threshold, then the requirement is not met.

[0040] After the above filtering is completed, the final output is the complete set of triples: ; This process ensures that entity information extracted from unstructured text is transformed into structured cultural knowledge representations, facilitating subsequent template filling and corpus generation.

[0041] S2 generates multi-style corpus sets from triple sets using basic templates and large language models respectively; The basic templates include both manually defined and automatically inductive methods, resulting in different template sets M. Triples then generate a corpus set Stemplate based on these different templates. ; .

[0042] The manual definition is based on a set of predefined templates, either manually designed or automatically summarized, according to common semantic relationship types and text expression rules in the cultural domain. These templates are used to transform structured knowledge triples (entity-relation-entity) into fluent and natural sentences, providing a stable text construction framework for the generation stage.

[0043] For example, based on common semantic relationship types in the cultural field (such as “originating from”, “commemorating”, “inherited from”, “invented from”, “located in”, etc.), several basic templates for language expressions are designed manually for these relationships.

[0044] For example, regarding the relationship "originating from", several basic templates can be designed: M 1 = Text1: "{ h Originating from { t} M 2=Text2: "{ t}yes{ h The place of origin of}. M 3=Text3: "{ h} originates from { t The traditional culture of}. For example, for the "commemoration" of a relationship, several basic templates can be designed: M 4 = Text4: "People set up { h} in memory of { t} M 5 = Text5: "{ h} is to commemorate { t Traditional festivals that arose from this. M 6 = Text6: "{ t The spirit of} through {h Thus, it was passed down. The { in these templates h} and{ t} represent the head and tail entities in the triple, respectively, and can be replaced during the corpus generation stage.

[0045] Automatic inductive methods include clustering using common grammatical patterns (such as structure extraction based on syntactic dependency trees); N-gram high-frequency sentence pattern extraction; and verb-preposition phrase configuration induction (such as "developed from" or "appeared during the period of").

[0046] By constructing semantically driven templates based on the two types of cultural knowledge relationships and integrating artificial language design with real corpus induction, a basic structure for cultural content control, style control and language security is provided for subsequent language model generation.

[0047] To achieve the diverse transformation from structured knowledge to natural language text, a large language model is used to perform diversity sampling to generate a corpus. S LM Large language models, including ChatGPT, DeepSeek, and Qwen3, can generate multiple semantically consistent but expressively varied natural language sentences by calling different large model functions and inputting different triples. These sentences are used to construct a cultural corpus with rich style and natural language. ; This indicates the sampling parameter configuration: Top-k, Top-p, Temperature. These are classic configuration parameters for large models. The parameter configurations and parameter sets are shown in the table below: Model Typical values ​​or settings ChatGPT (GPT-4 / GPT-4.5) - temperature 0.2 ~ 1.2 - top-p ∈ [0.7, 1.0], top-k=10 - style: "academic tone", "colloquial expression", etc. DeepSeek - top-k=20 - Style: "Use idioms", "Explain in a few words" - temperature: 0.8 ~ 1.0 - top-p ∈ [0.6, 0.8] Qwen3 (A Thousand Questions on General Principles) - Style: "Playing the Expert", "Please Imitate a Child" - Temperature: 0.7 ~ 1.1 - Top-k=50, - Top-p=0.9 Finally, the corpora generated from the basic template and the large model are merged to generate a multi-style corpus set. S gen ( h , r , t ),in ; S3 generates a basic corpus set by performing style transformation on multi-style corpora based on manually defined rules; To improve the diversity of corpus styles, a set of style tags with clear structure, stable expression, effective control, and strong extensibility were selected based on a combination of scenario-oriented, user-segmented, corpus-inductive, and manual verification methods. These style tags were defined as control variables. The generating function is expanded to:

[0048] This function indicates that it will generate multi-style corpus text by processing a set of triples using a base template and a large language model, by introducing different style tags. Φ Generate basic text corpus.

[0049] This mechanism not only enhances the diversity and style adaptability of the generated corpus, but also helps the system achieve more natural and credible cultural expression in various scenarios such as education, communication, and question answering.

[0050] Using triples: Taking (Dragon Boat Festival, commemoration, Qu Yuan) as an example, the following corpus was generated using different style tags: Style tag (Φ) Example of generated corpus Academic Style The Dragon Boat Festival is one of China's traditional festivals, established to commemorate the patriotic poet Qu Yuan. Children's style Did you know? We celebrate the Dragon Boat Festival to commemorate a great ancient hero named Qu Yuan! Classical Chinese style The Dragon Boat Festival is a tradition to commemorate Qu Yuan; it has been passed down for thousands of years and remains a popular folk custom. colloquial style The Dragon Boat Festival is actually to commemorate a person from ancient times named Qu Yuan, which is quite interesting! Concise science popularization style The Dragon Boat Festival is a festival to commemorate Qu Yuan, and it is celebrated in China on the fifth day of the fifth lunar month every year. Short video script style Do you know why we eat zongzi (sticky rice dumplings) and race dragon boats during the Dragon Boat Festival? It's actually to commemorate an ancient figure—Qu Yuan. Social media style #DragonBoatFestival# Commemorating Qu Yuan is not just about the festival itself, but also about cultural confidence! S4. Construct cultural corpus data from the basic corpus that simultaneously satisfies both structural alignment and semantic consistency checks.

[0051] Errors that may occur in corpora generated by large models mainly fall into two categories. The first is structural inconsistency errors, where the generated natural language text fails to accurately reproduce the structural elements of the original knowledge triples. For example, generating the statement "The Dragon Boat Festival is established by people to commemorate a historical figure" from the triple (Dragon Boat Festival, Commemoration, Qu Yuan), while semantically related, lacks a clear referent for the tail entity "Qu Yuan," or replaces the relation with a vague expression, resulting in entity omission or relation mismatch—a structural deviation. The second is semantic misinterpretation errors, where the text appears structurally correct, but the semantic understanding or direction is inconsistent with the original triples, even leading to factual deviations. For example, generating the statement "Qu Yuan wrote poems to celebrate the Dragon Boat Festival" from the same triple, while containing all entities and relational words in the triple, suffers from subject-object reversal, reversed causal relationships, or logical inconsistencies, severely misinterpreting the original meaning. These types of problems are difficult to detect with simple keyword matching and require methods such as semantic vector similarity or factual knowledge comparison for identification.

[0052] To ensure the semantic accuracy and knowledge consistency of the generated corpus, a dual verification mechanism is proposed. Specifically, this includes: Because corpora generated through large models may produce illusions, leading to semantic inconsistencies, therefore... The generated base corpus is used to determine inconsistencies using structural alignment. Extract entity pairs and relations from statement Si. Entity relations can be extracted using rule templates or semantic dependency parsing (this extraction method is conventional and will not be elaborated here). ; Determine whether the extracted triples exist in the set of triples in the knowledge graph obtained in the original step S1: ; If the result is 1, then the corpus Si satisfies structural alignment.

[0053] To prevent semantic misinterpretation—that is, when the surface structure of the text is correct, but the semantic understanding or direction is inconsistent with the original triples, or even leads to factual deviations—a semantic consistency method is used to verify the accuracy and consistency of the corpus. The specific method is as follows: The original triples are converted into natural language summaries Ti, and then sentence vector encoding is used to compare them with the generated corpus text Si, and cosine similarity is calculated: ; Set threshold A value is manually selected: ; That is, if the result is equal to 1, then the semantic consistency of the corpus text Si corresponding to the triple satisfies the condition.

[0054] Finally, the basic corpus that simultaneously satisfies dual verification is constructed into a cultural corpus.

[0055] That is, a corpus is constructed by taking text corpus Si that simultaneously satisfies the structural alignment method and semantic consistency check.

[0056] ; The above description is merely an exemplary embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made using the contents of the present invention specification and drawings under the technical concept of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. A method for generating corpus data in the cultural domain, characterized in that, The method includes: S1. Construct a set of triples based on textual data from the cultural domain; S2. The triples generate multi-style corpus sets through the basic template and the large language model, respectively; S3. Based on manually defined rules, perform style transformation on multi-style corpora to generate a basic corpus set; S4. Construct cultural corpus data from the basic corpus that simultaneously satisfies both structural alignment and semantic consistency checks.

2. The method for generating corpus data in the cultural domain as described in claim 1, characterized in that, Step S1 includes: S11. Obtain and clean the text data in the cultural field; S12. Identify cultural entities in the cleaned text; S13. Construct a set of triples by extracting the relationships between cultural entities in the text.

3. The method for generating corpus data in the cultural domain as described in claim 2, characterized in that, Step S13 includes: S131. Extract entity pair relation sets using either supervised relation extraction or open information extraction methods based on the text type; S132. Iterate through all entity pairs and their relations, and combine them into a set of triples; S133. The confidence of triples is predicted by the pre-trained model Bert, and triples with confidence exceeding the threshold are retained.

4. The method for generating corpus data in the cultural domain as described in claim 1, characterized in that, Step S12 includes: Cultural entities in the cleaned text are identified using the NER model; semantic similarity is used to normalize these cultural entities.

5. The method for generating corpus data in the cultural domain as described in claim 4, characterized in that, The semantic similarity method includes: Cultural entities are used to generate word vectors using the embedding method, and the semantic similarity between any two entities is calculated based on cosine similarity. : ； in: : Semantic vectors generated based on embedding for cultural entities; The closer the value is to 1, the more similar the semantics. Set a phase velocity threshold δ, if ,Will If the corresponding entities are the same entity, then they are merged; If the merged entity has a standard entry in the cultural dictionary, it is named according to the standard entry; otherwise, the entity is named using the expression that appeared most frequently or was shortest before the merge.

6. The method for generating corpus data in the cultural domain as described in claim 1, characterized in that, Step S2 includes: The triples were converted into different corpora using templates and large models respectively. The basic templates were generated into different template sets through manual definition and automatic induction methods. The large models included ChatGPT, DeepSeek, and Qwen3.

7. The method for generating corpus data in the cultural domain as described in claim 1, characterized in that, The structural alignment verification method in step S4 includes: Extract triples from the basic corpus again, and determine whether the triples exist in the set of triples generated in step S1. If they do, the basic corpus satisfies the structural alignment check.

8. The method for generating corpus data in the cultural domain as described in claim 1, characterized in that, The semantic consistency verification method in step S4 includes: The triples generated in step S1 are converted into natural language summaries. The cosine similarity between the summation and the basic corpus is calculated using sentence vector encoding. If the similarity is greater than a set threshold, the basic corpus satisfies the semantic consistency check.

Images