Multi-stage chain-oriented intelligent analysis method and device for JRAS documents

Abstract

The application discloses a multi-stage chain intelligent analysis method and device for JRAS documents, and a server method comprises the following steps: loading a preset structured JRAS document library and a preset oriental studies vertical knowledge graph; performing hierarchical retrieval according to the preset structured JRAS document library and the preset oriental studies vertical knowledge graph, so as to screen out a hierarchical document set associated with a research topic; performing paragraph-level extraction on documents in the hierarchical document set, and using a pre-tuned large language model to perform structured interpretation on the extracted paragraphs, to obtain an interpretation result; preprocessing the interpretation result according to dimensions of scholars, classics and civilizations, to generate a multi-dimensional structured knowledge base with traceability; and generating a structured report, a visual graph and a traceability archive based on the multi-dimensional structured knowledge base. By using the application, systematic and traceable analysis of tens of thousands of JRAS documents is realized, the distribution of computing power is optimized by using quantitative correlation degree scoring, and the problem of deep correlation in the vertical field of oriental studies is solved.

Description

Technical Field

[0001] This application relates to the field of natural language processing technology, and in particular to a multi-stage chaining intelligent analysis method and device for JRAS documents. Background Technology

[0002] In the field of Oriental Studies, especially for academic treasures like the *Journal of the Royal Asiatic Society* (JRAS; formerly *Transactions of the Royal Asiatic Society*, TRAS before 1834; for consistency, hereinafter referred to as JRAS), which contains a vast amount of multilingual, unstructured historical documents, scholars often face enormous challenges in systematically tracing their historical development. For example, when scholars attempt to comprehensively extract and integrate scattered information about specific texts and civilizational developments from tens of thousands or even more JRAS documents spanning hundreds of years, they must rely on manual, piecemeal reading and manual correlation, which is not only inefficient but also makes it difficult to guarantee the systematicity and completeness of the analytical results.

[0003] To address the limitations of manual processing, scholars have attempted to use large language models (such as ChatGPT). However, this approach is limited by the context window length of large models, making it impossible to achieve end-to-end, systematic, and global analysis of tens of thousands of documents. Furthermore, the generated content lacks a precise source tracing mechanism to specific paragraphs in the original documents, and coupled with the influence of AI illusions, it is difficult to meet the stringent standards of rigor and verifiability required for academic research.

[0004] On the other hand, the current system lacks vertical optimization for the field of Oriental studies. It has neither built a professional ontology model adapted to "scholar lineage - text version - civilization context" nor can it effectively deal with the technical challenges unique to Oriental studies literature, such as multilingual entity alignment. As a result, the analysis results of existing tools are often superficial and can at most play a low-level and crude auxiliary role, unable to support in-depth professional academic research.

[0005] Furthermore, the current system lacks a layered perception of the depth of the topic, and cannot distinguish between direct evidence documents, indirect references and potentially related documents, resulting in mixed search results and frequent errors. Scholars need to manually sift through massive amounts of noise to find high-value documents. At the same time, the existing AI interpretation process treats all searched documents equally, performing full-text processing and high-computational structured interpretation, which not only causes a huge waste of computing resources, but also reduces the signal-to-noise ratio of the final knowledge integration due to interference from low-value documents. Summary of the Invention

[0006] This application provides a multi-stage chaining intelligent analysis method and apparatus for JRAS literature. To provide a basic understanding of some aspects of the disclosed embodiments, a brief summary is given below. This summary is not intended as a general commentary, nor is it intended to identify key / important components or describe the scope of protection of these embodiments. Its sole purpose is to present some concepts in a simple form as a prelude to the detailed description that follows.

[0007] In a first aspect, embodiments of this application provide a multi-stage intelligent analysis method for JRAS literature, applied to the server side, the method including: In response to a client's request for intelligent chain analysis carrying research topics, a pre-defined structured JRAS literature repository and a pre-defined Oriental studies vertical knowledge graph are loaded. The pre-defined structured JRAS literature repository is constructed based on original JRAS PDF documents containing multilingual content. The pre-defined Oriental studies vertical knowledge graph's ternary ontology model includes scholar entities, classic entities, civilization entities, and pre-defined relation types. The pre-defined Oriental studies vertical knowledge graph is a graph-structured knowledge base with clear domain boundaries and semantic constraints, constructed based on the ternary ontology model and through entity recognition and relation extraction of JRAS documents. A hierarchical search is conducted based on a pre-defined structured JRAS literature database and a pre-defined Oriental studies vertical knowledge graph to select hierarchical literature sets that are relevant to the research topic. Paragraph-level extraction is performed on the documents in the hierarchical document set, and a pre-tuned large language model is used to perform structured interpretation of the extracted paragraphs to obtain the interpretation results; The interpretation results are preprocessed according to the dimensions of scholars, classics, and civilizations to generate a multi-dimensional structured knowledge base with traceability; based on the multi-dimensional structured knowledge base, a structured report, a visualization map, and traceability archives are generated and sent to the client.

[0008] This embodiment uses hierarchical retrieval to precisely select hierarchical literature sets related to the research topic from a pre-defined structured JRAS literature database, focusing the analysis scope from the entire set to a subset. Then, paragraph-level extraction further narrows the processing object from the entire document to core text blocks, completely avoiding the risk of context overflow from large models and achieving systematic and batch processing of tens of thousands of documents. Based on this, a traceable archive is finally generated, establishing an immutable mapping link between AI-automated analysis and original documentary evidence, meeting the stringent standards of academic research for verifiable and retrievable conclusions.

[0009] This embodiment utilizes a pre-constructed vertical knowledge graph of Oriental studies, whose ternary ontology model specifically defines scholars, classics, civilizations, and their various core relationships, providing a structured container for domain knowledge. Simultaneously, a large language model fine-tuned based on Oriental studies corpora is employed for structured interpretation, enabling precise adaptation to entity recognition and relation extraction tasks within the domain. Combined with relational reasoning retrieval and cross-linguistic entity alignment using the graph in hierarchical retrieval, it achieves the transformation from unstructured documents to automatic linking across the entire scholar-classic-civilization domain, supporting in-depth and professional interpretation of Oriental studies topics.

[0010] In some embodiments of this application, a hierarchical search is performed based on a pre-defined structured JRAS literature database and a pre-defined Oriental studies vertical knowledge graph to filter out hierarchical literature sets related to the research topic, including: Based on scholar entities, classic text entities, civilization entities, and predefined relation types, semantic analysis and decomposition of the research topic are performed to obtain the combination of core entities and relations, resulting in a triple structure in the format of entity-relation-entity; the core entities correspond to scholar entities, classic text entities, and civilization entities; Based on the combination of core entities and relationships, a progressive retrieval from precise to generalized is performed on the pre-defined structured JRAS document database to obtain the first candidate document set for precise retrieval, the second candidate document set for semantic expansion retrieval, and the third candidate document set for relational reasoning retrieval. For each document in the retrieved candidate document set, the relevance score is quantified based on its semantic matching degree with the research topic, keyword co-occurrence frequency, cross-citation count, and the length of the association path in the knowledge graph; the quantified relevance score is used to guide paragraph-level extraction and structured interpretation. Based on the relevance score and a preset threshold range, each document is divided into high-relevance documents, medium-relevance documents, and low-relevance documents, resulting in a stratified document set related to the research topic.

[0011] In the above embodiments, a ternary ontology-based approach is used to decompose natural language topics into structured triples, thereby accurately identifying the core of the topic. Subsequently, a three-tiered progressive retrieval process is executed, progressing from exact matching to semantic extension and relational reasoning, systematically covering the range of literature from direct to potential. Finally, a multi-dimensional quantitative scoring model is introduced to transform fuzzy relevance into precise numerical scores, automatically classifying literature into high, medium, and low relevance categories. This process transforms the previously manual noise reduction and screening work into a directly output, clearly structured hierarchical literature set, enabling scholars to quickly focus on high-value literature groups.

[0012] In the above embodiments, the quantitative relevance score is used as a guiding parameter for downstream processing. The system can dynamically configure differentiated processing strategies based on the relevance level of the documents. This on-demand computing power allocation mechanism based on hierarchical results effectively filters out low-value interference while ensuring the depth of core document analysis. This not only significantly saves computing resources but also significantly improves the signal-to-noise ratio and academic rigor of the final multidimensional structured knowledge base by eliminating noisy data.

[0013] In some embodiments of this application, a progressive retrieval from precise to generalized is performed on a pre-defined structured JRAS document database based on a combination of core entities and relationships, including: In the pre-defined structured JRAS document database, Boolean logic combination is used to perform precise retrieval using the decomposed core entities and their synonyms as keywords, resulting in the first candidate document set for the first layer of precise retrieval. From the pre-set vertical knowledge graph of Oriental studies, obtain the extended attributes of the core entities. The extended attributes include the scholar's research field, the theme of the classics, and the core characteristics of the civilization. In the pre-defined structured JRAS literature database, semantic generalization retrieval is performed using extended attributes as keywords to obtain the second candidate literature set for the second-level semantic extended retrieval. Based on the triplet structure, one-hop or multi-hop relational reasoning is used to search for and study potential entities that are indirectly related to the research topic. In the pre-defined structured JRAS literature database, a search is conducted using latent entities as keywords to obtain the third candidate literature set for third-level relational reasoning retrieval.

[0014] In this embodiment, a three-layer progressive retrieval mechanism is used to ensure accurate targeting of core literature while systematically mining potentially related literature by leveraging the extended attributes and multi-hop reasoning capabilities of the knowledge graph. This effectively solves the problem of flattened traditional retrieval results and achieves full coverage from direct evidence to indirect references and then to potential connections, significantly reducing the cost for scholars to manually sift through massive amounts of noise to find high-value literature.

[0015] In some embodiments of this application, the research topic corresponds to a core entity; paragraph-level extraction is performed on the documents in the hierarchical document set, and a pre-tuned large language model is used to perform structured interpretation of the extracted paragraphs to obtain the interpretation results, including: Obtain the relevance level of each document to be processed in the hierarchical document set. The relevance level is used to indicate whether each document to be processed is a high-relevance document, a medium-relevance document, or a low-relevance document. If the relevance level indicates that each document to be processed is a highly relevant document, a fine-grained scan is performed on the full text of each document to locate all sentences containing core entities and their synonyms and near-synonyms, and these are merged into paragraphs to be interpreted; or, If the relevance level indicates that each document to be processed is of medium relevance, locate the chapter or paragraph containing the core entity in the full text of each document, and use that chapter or paragraph as a whole as the paragraph to be interpreted; or, If the relevance level indicates that each document to be processed is a low-relevance document, locate the paragraph in the full text of each document to be processed that first mentions any core entity, and use that paragraph as the paragraph to be interpreted. Based on the paragraph to be interpreted and the pre-tuned large language model, the interpretation result is generated.

[0016] In this embodiment, a differentiated paragraph extraction strategy is dynamically configured based on the relevance level of the documents. Specifically, a fine-grained full-text scan is performed on highly relevant documents to maximize information capture, relevant chapters are located for medium-relevance documents to balance efficiency and coverage, and only the first-mentioned paragraphs are extracted from low-relevance documents for rapid filtering. This achieves intelligent allocation of computing resources to high-value documents, avoids the risk of context spillover from large models, and significantly improves the efficiency of structured interpretation and the signal-to-noise ratio of final knowledge integration.

[0017] In some embodiments of this application, the interpretation result is generated based on the paragraph to be interpreted and a pre-tuned large language model, including: The maximum token length of the text block is assigned to the paragraph to be interpreted based on the relevance level of each document to be processed; the maximum token length of the text block assigned to documents with high relevance is less than the maximum token length of the text block assigned to documents with medium and low relevance. The paragraph to be interpreted is split into one or more text blocks according to the maximum token length of the assigned text block; Each text block is input into a pre-tuned large language model, and interpretation task instructions of different complexities are configured according to the relevance level of the document to which it belongs, so as to obtain structured interpretation sub-results. All sub-interpretations from the same document are merged and integrated to form a comprehensive interpretation result for that document; among them, The comprehensive interpretation results should include at least the identified entities, the relationships between entities, the core viewpoints, and the original text evidence fragments used as supporting evidence, and should be associated with the document's identifier and the original text's location information.

[0018] In this embodiment, by dynamically configuring differentiated text block lengths and interpretation task complexity based on the document relevance level, smaller granularity blocks and high-precision instructions can be used for highly relevant documents to capture deep semantics, while larger blocks and basic instructions can be used for documents with medium to low relevance to balance efficiency. While effectively controlling the context overflow and computational consumption of large models, it achieves accurate structured extraction from micro evidence to macro viewpoints and ensures that each conclusion is anchored to the original document location.

[0019] In some embodiments of this application, structured reports, visual maps, and source tracing archives are generated based on a multidimensional structured knowledge base, including: Newly discovered entities and relationships in the multidimensional structured knowledge base are injected into a pre-defined Oriental studies vertical knowledge graph. Based on the injected entities and relationships, the knowledge graph's built-in graph computing reasoning engine performs multi-hop reasoning along predefined relationship paths to mine and create the associations between entities, thus obtaining the latest knowledge graph. Based on the latest knowledge graph, knowledge content is automatically extracted and organized according to preset template dimensions to generate structured reports; The visualization engine is invoked, and based on the latest knowledge graph, a visual graph is generated with nodes in the format of scholar-classic-civilization and edges with preset relationship types and inferred connections. The visual graph supports node zooming, clicking to view details, and highlighting of relationship paths. By traversing all conclusion entries in the multidimensional structured knowledge base, extracting their associated document identifiers and original text location information, and organizing them in document order to generate a source archive, the source archive enables each conclusion in the structured report and visualization map to be traced back to a specific paragraph in the original document.

[0020] This embodiment utilizes a closed-loop mechanism of dynamic injection, graph reasoning, and multidimensional generation to automatically output structured reports containing scholar lineages, textual transmission, and civilizational contexts based on a visualized graph. It also provides visualized graphs that support interactive exploration and archives with complete traceability chains. This transforms discrete multidimensional knowledge into verifiable and interactive academic results, thereby achieving a leap from data integration to deep knowledge discovery.

[0021] In some embodiments of this application, a preset structured JRAS literature database and a preset Oriental studies vertical knowledge graph are generated according to the following steps: Obtain the original PDF JRAS documentation containing multilingual content; The original JRAS PDF documents are parsed and OCR recognized to obtain the processed text, metadata and translation results; The processed text, metadata, and translation results are stored in the FTS5 full-text search engine database to form a standardized structured document library containing original text location information, which serves as the preset structured JRAS document library. Define a ternary ontology model containing scholar entities, classic entities, and civilization entities, and multiple relation types; scholar entities include master-disciple relationship attributes, classic entities include version transmission attributes, and civilization entities include spatiotemporal range attributes. Based on a ternary ontology model and multiple relation types, a standardized structured literature database is used to generate a pre-defined vertical knowledge graph of Oriental studies.

[0022] In some embodiments of this application, a pre-defined Oriental studies vertical knowledge graph is generated based on a ternary ontology model, multiple relation types, and a standardized structured literature database, including: Identify scholar names, classic titles, and civilization names from standardized structured literature databases; For multilingual document scenarios, a unified identification and alignment of scholar names, classic titles, and civilization names is achieved through a pre-constructed name mapping table and semantic similarity calculation strategy to obtain an entity set; Based on multiple relation types, the associations between entities are extracted from the document texts of a standardized structured document database, and multiple relation instances are constructed. Entity sets and multiple types of relationship instances are injected into a graph database, and pattern matching is performed using a ternary ontology model to generate a preset Oriental studies vertical knowledge graph containing basic relationships.

[0023] This embodiment solves the problem of unstructured documents being difficult to systematically associate by constructing a standardized multilingual JRAS structured literature database and defining a ternary ontology and exclusive attributes based on knowledge in the field of Oriental studies. It generates a vertical knowledge graph using cross-lingual entity alignment and relation extraction techniques. This realizes the transformation from original multilingual documents to a structured knowledge network with deep attributes, laying a reliable data foundation for subsequent accurate hierarchical retrieval and in-depth academic interpretation.

[0024] In some embodiments of this application, a pre-tuned large language model is generated according to the following steps: Collect professional literature, classical commentaries, academic papers and historical journal texts in the field of Oriental studies to construct a domain-specific training corpus containing multiple languages; The domain-specific training corpus is cleaned and labeled to form a structured training dataset of four tuples: entity-relationship-opinion-evidence. Load the base big language model and fine-tune it with instructions based on the four-tuple structured training dataset so that the base big language model can be adapted to the tasks of entity recognition, relation extraction and paragraph-level structured interpretation of scholar entity, classic entity and civilization entity. The base language model after fine-tuning the instructions is verified. After successful verification, it is deployed to the server and configured to only receive paragraph-level input from hierarchical document sets and output interpretation results, thus obtaining the pre-fine-tuned language model.

[0025] This embodiment constructs a dedicated corpus for Oriental studies that includes multiple languages ​​and fine-tunes instructions based on quadruples. This enables the large model to accurately adapt to entity recognition, relation extraction, and structured interpretation capabilities covering vertical fields of scholars, classics, and civilizations. Thus, while avoiding the illusion of general models, it achieves end-to-end transformation from paragraph-level input to structured output with traceability, providing a high-precision dedicated AI engine for in-depth academic analysis of large-scale JRAS literature.

[0026] Secondly, embodiments of this application provide a multi-stage intelligent analysis device for JRAS literature, the device comprising: The document library and graph loading module is used to respond to the intelligent analysis request sent by the client carrying research topics, and load the preset structured JRAS document library and the preset Oriental Studies vertical knowledge graph. The preset structured JRAS document library is built based on JRAS original PDF documents containing multilingual content. The preset Oriental Studies vertical knowledge graph has a ternary ontology model that includes scholar entities, classic entities, civilization entities and preset relation types. The preset Oriental Studies vertical knowledge graph is a graph structure knowledge library with clear domain boundaries and semantic constraints, built based on the ternary ontology model and through entity recognition and relation extraction of JRAS documents. The hierarchical retrieval module is used to perform hierarchical retrieval based on a preset structured JRAS literature database and a preset Oriental studies vertical knowledge graph to filter out hierarchical literature sets related to the research topic. This includes: performing semantic analysis and decomposition on the research topic based on the scholar entity, classic text entity, civilization entity, and preset relation types to obtain a combination of core entities and relations, resulting in a triple structure in the format of entity-relation-entity; the core entities correspond to the scholar entity, classic text entity, and civilization entity; and performing a progressive retrieval from precise to generalized based on the combination of core entities and relations in the preset structured JRAS literature database to obtain the first layer of precise... The system retrieves a first candidate document set, a second candidate document set through semantic expansion, and a third candidate document set through relational reasoning. For each document in the retrieved candidate document set, a relevance score is quantified based on its semantic matching degree with the research topic, keyword co-occurrence frequency, cross-citation count, and the length of the association path in the knowledge graph. The quantified relevance score is used to guide the paragraph-level extraction and the structured interpretation. Based on the relevance score and a preset threshold range, each document is divided into high-relevance documents, medium-relevance documents, and low-relevance documents, resulting in a hierarchical document set related to the research topic. The structured interpretation module is used to extract paragraphs from documents in the hierarchical document set and use a pre-tuned large language model to perform structured interpretation of the extracted paragraphs to obtain interpretation results. The analysis result generation module is used to preprocess the interpretation results according to the dimensions of scholars, classics, and civilizations to generate a multi-dimensional structured knowledge base with traceability. Based on the multi-dimensional structured knowledge base, a structured report, a visualization map, and a traceability archive are generated and sent to the client.

[0027] The above description is merely an overview of the technical solutions of the embodiments of this application. In order to better understand the technical means of the embodiments of this application and to implement them in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the embodiments of this application more obvious and understandable, specific implementation methods of this application are described below. Attached Figure Description

[0028] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0029] Figure 1 This is a schematic diagram of the process flow for a multi-stage intelligent analysis method for JRAS literature provided in an embodiment of this application; Figure 2 This is a schematic diagram of a client interface for submitting a request, provided in an embodiment of this application; Figure 3 This is a schematic diagram illustrating the process of generating a preset structured JRAS literature database and a preset Oriental studies vertical knowledge graph, as provided in an embodiment of this application. Figure 4 This is a schematic diagram of a hierarchical retrieval result provided in an embodiment of this application; Figure 5 This is a schematic diagram of a hierarchical retrieval process provided in an embodiment of this application; Figure 6 This is a schematic diagram of a relation extraction method provided in an embodiment of this application; Figure 7 This is a schematic diagram illustrating a core concept provided in an embodiment of this application; Figure 8 This is a schematic diagram illustrating a structured interpretation process provided in an embodiment of this application; Figure 9 This is a schematic diagram of a traceability archive in an application scenario provided by an embodiment of this application; Figure 10 This is a schematic diagram of the structure of a multi-stage linked intelligent analysis device for JRAS literature provided in an embodiment of this application; Figure 11 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0030] The following description and accompanying drawings fully illustrate specific embodiments of this application to enable those skilled in the art to practice them.

[0031] It should be understood that the described embodiments are merely some, not all, of the embodiments in this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application.

[0032] In the following description, when referring to the accompanying drawings, the same numbers in different drawings denote the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0033] In the description of this application, it should be understood that the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances. Furthermore, in the description of this application, unless otherwise stated, "multiple" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship.

[0034] The *Proceedings of the Royal Asiatic Society* is the most important academic publication platform of the Royal Asiatic Society. For two centuries, it has accompanied the Society's development, accumulating a wealth of research findings in British Oriental studies. Looking at the history of British Oriental studies, it is undoubtedly the founding journal, its development a microcosm of the history of British Oriental studies. Many articles in the early *Proceedings* were milestones and pioneering achievements in British Oriental studies, and even in Western Oriental studies, serving as the origin of numerous Oriental studies and laying the foundation for the research tone, fields, and characteristics of British Oriental studies. For two centuries, the *Proceedings* has continuously adapted itself to better promote and adapt to the development of Oriental studies. Today, it remains one of the most authoritative journals in the field of Oriental studies, carrying the genealogy and development history of British Oriental studies, and is an invaluable historical resource for understanding the history of British Oriental studies. Therefore, the academic history of the *Proceedings* cannot be ignored when analyzing the history of British Oriental studies.

[0035] By leveraging digital intelligence and humanities, we systematically and comprehensively integrate and absorb over two centuries of Western Oriental academic history, and deeply explore the Oriental significance of classical Chinese texts, such as historical records, Confucian classics, Buddhist texts, and literary classics, to investigate the underlying logic and future development possibilities of Chinese Oriental studies, so as to better serve the needs of national strategies.

[0036] Building upon this foundation, the system further delves into the Oriental significance of Chinese historical classics, Confucian classics, Buddhist classics, and other literary works. Through differentiated paragraph extraction and structured interpretation using a large language model, it automatically extracts the deep connections between "classics-reflection-civilization" from unstructured ancient texts. The system not only verifies the citations and misinterpretations of Chinese classics by Western academia but also uncovers the intrinsic value of Chinese indigenous literature in the mutual learning between civilizations. Thus, under the dual perspectives of "Western learning spreading eastward" and "Chinese learning spreading westward," it clarifies the underlying logic of the development of Chinese Oriental studies and provides calculable, verifiable, high-standard, and end-to-end digital and humanistic support for building a development path that combines an international perspective with an independent Chinese knowledge system.

[0037] The "Scholar-Classic-Civilization" vertical knowledge graph of Oriental studies constructed in this solution is the cornerstone for comprehensively integrating two centuries of Western Oriental scholarship. Through hierarchical retrieval and structured interpretation of tens of thousands of JRAS documents, the system not only accurately captures the core viewpoints and lineages of representative Western Orientalists and Sinologists such as Henry Yul, Paul Pelliot, and Édouard Chavannes, but also utilizes cross-language entity alignment technology to deeply analyze JRAS documents, primarily in English. Simultaneously, it identifies and connects content in numerous Asian and European languages ​​used in these documents, while also recognizing that the authors of JRAS documents are Orientalists from around the world. This allows for the mapping of the academic exchange network and historical lineage of global Oriental studies, revealing the academic debates within JRAS documents and the historical context of research on Oriental civilization. All of this is systematically linked into a unified knowledge graph network, achieving a digital reconstruction and panoramic understanding of the evolution of Western Orientalist methodologies and research paradigms.

[0038] Furthermore, the technical solution of this application overcomes the prevalent mindset and technical bias in existing technologies, which hold that it is impossible to achieve automated, systematic, and in-depth intelligent analysis of massive, multilingual, and unstructured professional historical documents while ensuring academic rigor. Traditional views tend to rely on expert interpretation or superficial processing by general AI tools. This application, through an innovative hierarchical retrieval-differentiated interpretation-vertical graph architecture, combines the powerful semantic understanding capabilities of large models with the structured constraints of knowledge graphs and the professional design of vertical domain ontologies, successfully solving the following technical challenges: 1) It breaks through the bottleneck of the limited context window of general large language models, which cannot handle the systematic analysis of large-scale documents; 2) It fills the knowledge representation gap in general digital humanities tools in vertical fields such as Oriental studies, which lacks professional ontologies and cannot achieve deep integration of "scholar-classic text-civilization"; 3) It eliminates the defects of AI processing results lacking reliable traceability and failing to meet the needs of academic verification, achieving efficient automated analysis of tens of thousands of documents while ensuring the transparency of the analysis process and the verifiability of the conclusions. Therefore, this application is neither a simple combination of existing technologies, but can produce synergistic and enhanced technical effects, achieving unexpected technological progress in the intersection of digital humanities and artificial intelligence, and these are problems that the field has been eager to solve.

[0039] The following will be combined with the appendix Figure 1 - Appendix Figure 9 This application provides a detailed description of the multi-stage intelligent linking analysis method for JRAS literature provided in its embodiments. This method can be implemented using a computer program and can run on a von Neumann-based multi-stage intelligent linking analysis device for JRAS literature. This computer program can be integrated into applications or run as a standalone tool application.

[0040] Please see Figure 1 This document provides a flowchart illustrating a multi-stage intelligent analysis method for JRAS literature, applicable to the server side. (Example:) Figure 1 As shown, the method in this application embodiment includes the following steps: S101, in response to the client's request for intelligent chain analysis carrying research topics, loads a preset structured JRAS literature library and a preset Oriental studies vertical knowledge graph. The preset structured JRAS literature library is constructed based on JRAS original PDF documents containing multilingual content. The preset Oriental studies vertical knowledge graph's ternary ontology model includes scholar entities, classic entities, civilization entities, and preset relation types. The preset Oriental studies vertical knowledge graph is a graph structure knowledge library with clear domain boundaries and semantic constraints, constructed based on the ternary ontology model and through entity recognition and relation extraction of JRAS documents. The client is the terminal interface through which users initiate analysis requests. It receives research topics in natural language and displays the final analysis results. The linked intelligent analysis request carries a specific research topic and triggers a multi-stage linked processing flow on the server. The research topic is a natural language query entered by the user that requires system analysis; its content implicitly contains the scholar-text-civilization triad relationship to be analyzed. The pre-built structured JRAS literature database is a pre-constructed, high-speed searchable digital literature dataset. Its data source is the original PDF document of the *Journal of the Royal Society of Asiatic Studies* (JRAS), containing content in Chinese, English, Sanskrit, and other languages, constructed through parsing, OCR recognition, metadata indexing, and full-text indexing. The pre-built Oriental Studies vertical knowledge graph is a pre-constructed domain knowledge base. It uses a specially defined ternary ontology model to model knowledge in the Oriental Studies domain. This model includes three core entities: scholars, texts, and civilizations, as well as pre-defined relationship types connecting these entities (such as annotation, reflection, and transmission). The ternary ontology model is the schema layer of a knowledge graph, which defines the data structure of basic concepts (entity types), attributes, and relationships between concepts (relationship types) within the domain.

[0041] For example, the scholar entity is a node type in the ontology model representing Oriental studies researchers, containing attributes such as name, nationality, birth and death dates, and teacher-student relationships. The classic text entity is a node type in the ontology model representing Oriental studies research objects (books, manuscripts, etc.), containing attributes such as book title, author, date of composition, and version distribution. The civilization entity is a node type in the ontology model representing a specific historical and cultural region, containing attributes such as civilization name, spatiotemporal scope, and core characteristics. Preset relationship types are edge types defined in the ontology model used to connect different entity types, specifying the allowed interaction logic between entities (e.g., "scholar-annotation-classic text").

[0042] For example Figure 2As shown, in response to the user typing the research topic "Henry Yuer's annotations on Marco Polo's Travels and his contribution to the study of Yuan Dynasty China" in the input box of the client's web interface and clicking the start analysis button, a chain intelligent analysis request is triggered. This request is sent to the server. After receiving the request, the server loads the preset structured JRAS literature database and the preset Oriental Studies vertical knowledge graph from the storage system. The pre-defined structured JRAS database includes FTS5 full-text index data of 34,996 JRAS documents published between 1860 and 1930, including the original English text and some Chinese and Sanskrit translations and annotations. The pre-defined Oriental Studies vertical knowledge graph is constructed based on a ternary ontology model, which defines scholar entities (such as "Henry Yull" with the attribute "British Orientalist"), classic entities (such as "The Travels of Marco Polo" with the attribute "13th-century travelogue"), civilization entities (such as "Yuan Dynasty China" with the attribute "1271-1368"), as well as pre-defined relation types such as "annotations", "reflections", and "influences".

[0043] In some embodiments of this application, the specific process of generating a preset structured JRAS literature database and a preset Oriental studies vertical knowledge graph includes: acquiring original JRAS PDF documents containing multilingual content; performing PDF parsing and OCR recognition on the original JRAS PDF documents to obtain processed text, metadata, and translation results; storing the processed text, metadata, and translation results in the FTS5 full-text search engine database to form a standardized structured literature database containing original text location information, which serves as the preset structured JRAS literature database; defining a ternary ontology model and multiple relation types, including scholar entities, classic entities, and civilization entities, based on common knowledge and / or basic research methods in the field of preset Oriental studies; scholar entities include teacher-student relationship attributes, classic entities include version transmission attributes, and civilization entities include spatiotemporal range attributes; and generating a preset Oriental studies vertical knowledge graph based on the ternary ontology model, multiple relation types, and the standardized structured literature database.

[0044] Among them, the presupposition of common knowledge in the field of Oriental studies refers to the widely recognized basic classification system and consensus content in the discipline of Oriental studies. For example, the core analytical objects of Oriental studies are usually divided into three categories: scholars (Orientalist / Sinologist / Indologist, etc.), classic texts (classical documents, commentaries, archaeological inscriptions, etc.), and civilizations (ancient East Asian / Central Asian / South Asian civilizations and cultural forms, etc.). Known typical academic relationship types include "scholar-mentor / influence-scholar", "scholar-textual research / translation and annotation-classical texts", "classical texts-record / reflection-civilization", "civilization-exchange / conflict-civilization", etc.

[0045] Among them, the basic research methods refer to the conventional research methods commonly used in the collation and research of Oriental studies literature and are well known to those in the field. For example, the source tracing and collation method of locating the original documents by author's name, publication year, volume and issue and checking the citations; the method of aligning terms and comparing versions based on multilingual variants (such as Sanskrit-Chinese-English comparison); and the map-based approach of inferring the scope of academic influence based on lineage and transmission paths of classics (i.e. the practice of traditional humanities scholars manually drawing "academic lineage maps" or "classical transmission maps").

[0046] Specifically, the process of generating a pre-defined vertical knowledge graph of Oriental studies based on a ternary ontology model, multiple relation types, and a standardized structured literature database includes: identifying scholar names, classic text names, and civilization names from the standardized structured literature database; for multilingual literature scenarios, uniformly identifying and aligning scholar names, classic text names, and civilization names using a pre-constructed name mapping table and semantic similarity calculation strategy to obtain an entity set; extracting the relationships between entities from the literature texts of the standardized structured literature database according to multiple relation types to construct multiple relation instances; injecting the entity set and multiple relation instances into a graph database, combining the ternary ontology model to complete pattern matching, and generating a pre-defined vertical knowledge graph of Oriental studies containing basic relational relationships.

[0047] Herein, the JRAS original PDF documents of multilingual content refer to the original academic documents published by the Journal of the Royal Asiatic Society (JRAS) in different historical periods, which contain multiple languages such as English, Chinese, Sanskrit, and Arabic, are stored in PDF (Portable Document Format), and include scanned image versions and searchable text versions. PDF parsing is a technical process of extracting text streams, fonts, coordinates and metadata (such as title, author, publication year) from PDF files; OCR (Optical Character Recognition) is a technical means of performing image preprocessing, layout analysis and character recognition on scanned image version PDFs to convert them into machine-readable text. The FTS5 full-text search engine database refers to the FTS5 (Full Text Search 5) virtual table module in the SQLite database engine, which is used to build high-performance full-text indexes. The name mapping table refers to a pre-constructed mapping dictionary of multilingual entity aliases to standard entity IDs, which is used to solve the problem of different names for the same entity in documents in different languages (such as "Yule" and "裕尔", "Marco Polo" and "马可·波罗"). The semantic similarity calculation strategy is an algorithm strategy that uses a word vector model (such as Word2Vec) or an edit distance algorithm to calculate the semantic correlation of name strings in different languages to assist entity alignment when there is no clear mapping. The multi-class relation instance is a directed edge that exists specifically in the knowledge graph and connects two entities. It is composed of subject entity-relation type-object entity, and is a specific fact extracted from the text of the standardized document library through rules or models (such as [Henry Yule]-[annotate]-[[《马可·波罗游记》](@replace=10001)]).

[0048] In the above embodiment, by constructing a standardized JRAS structured document library rich in multilingual content, defining a ternary ontology and exclusive attributes based on domain knowledge of oriental studies, and using cross-lingual entity alignment and relation extraction technologies to generate a vertical knowledge graph, this embodiment solves the problem that unstructured documents are difficult to systematically associate, realizes the transformation from original multilingual content documents to a structured knowledge network with deep attributes, and lays a data foundation for subsequent accurate hierarchical retrieval and in-depth academic interpretation.

[0049] In a possible implementation, approximately 50,000 multilingual original PDF documents of JRAS published between 1860 and 1950 are obtained from the official archives of JRAS and partner institutions, covering original English texts and their corresponding Chinese, French, and German translations. PDF parsing and OCR recognition are performed on the PDF documents. The text stream is directly extracted from searchable PDFs; for 19th-century scanned PDFs, a deep learning-based OCR model (such as CRNN) is used for high-precision recognition to ensure the recognition rate of non-Latin characters (such as Chinese characters). The metadata (title, author, publication year, language) of each document is extracted, the recognized text is segmented by natural paragraphs, and an original text positioning information index of "document ID-page number-paragraph number" is established. The processed text, metadata and machine translation results (for non-English documents) are stored in a SQLite library configured with an FTS5 full-text search engine database, forming a standardized structured document library that supports Boolean search and phrase matching.

[0050] Then, a triple ontology model is defined based on domain expert knowledge. The scholar entity includes the mentor-student relationship attribute (such as supervisor); the classic work entity includes the version circulation attribute (such as translated from version); the civilization entity includes the spatiotemporal scope attribute (such as start year - end year). Multiple types of relationship types are defined, such as annotates, cites, and researches. A NER model is used to identify scholar names (such as H. Yule), classic work names (such as The Book of Ser Marco Polo), and civilization names (such as Yuan Dynasty) from the document library. For the multilingual document scenario, a name mapping table (such as Yule->裕尔) is used for unified identification; for aliases not included, a semantic similarity calculation strategy (such as BERT-based cross-language semantic matching) is adopted for alignment to generate a unique entity set. Based on the predefined multiple types of relationship types, a relationship extraction model (such as a joint extraction model based on a pre-trained language model) is used to extract facts from the document text, and construct multiple types of relationship instances (for example, [Yule]-[annotates]-[Marco Polo's Travels] is extracted from the sentence "Yule’s notes on Marco Polo..."). The cleaned entity set and multiple types of relationship instances are batch-injected into the Neo4j graph database, which strictly follows the triple ontology model to complete schema matching, and finally generates a preset Oriental studies vertical knowledge graph containing about 100,000 entities and 300,000 relationships.

[0051] For example Figure 3 as shown in, Figure 3This is a schematic diagram illustrating the process of generating a pre-defined structured JRAS literature database and a pre-defined Oriental studies vertical knowledge graph provided in this application. First, original JRAS PDF documents with multilingual content are acquired and subjected to PDF parsing and OCR recognition to obtain the corresponding processed text, metadata, and translation results. Then, these processing results are stored in the FTS5 full-text search engine database, thus forming a standardized structured literature database containing original text location information. Simultaneously, based on common knowledge and / or basic research methods in the field of Oriental studies, a ternary ontology model including scholar entities, classic text entities, and civilization entities, as well as multiple relation types, are defined. Finally, combining the aforementioned standardized structured literature database with the above-mentioned ternary ontology model and relation types, entities are identified and aligned, and relationships between entities are extracted, thereby generating a pre-defined Oriental studies vertical knowledge graph containing basic relational relationships.

[0052] S102, perform hierarchical retrieval based on the pre-set structured JRAS literature database and the pre-set Oriental studies vertical knowledge graph to select hierarchical literature sets related to the research topic; In some embodiments of this application, the specific process of performing hierarchical retrieval based on a preset structured JRAS literature database and a preset Oriental studies vertical knowledge graph to select a hierarchical literature set associated with the research topic includes: semantic analysis and decomposition of the research topic based on scholar entities, classical text entities, civilization entities, and preset relation types to obtain a combination of core entities and relations, resulting in a triplet structure in the format of entity-relation-entity; the core entities correspond to scholar entities, classical text entities, and civilization entities; and a progressive retrieval from precise to generalized based on the combination of core entities and relations is performed on the preset structured JRAS literature database to obtain... The first layer consists of a precise search for the first candidate document set, the second layer consists of a semantic expansion search for the second candidate document set, and the third layer consists of a relational reasoning search for the third candidate document set. For each document in the retrieved candidate document set, a relevance score is quantified based on its semantic matching degree with the research topic, keyword co-occurrence frequency, cross-citation count, and the length of the association path in the knowledge graph. The quantified relevance score is used to guide paragraph-level extraction and structured interpretation. Based on the relevance score and a preset threshold range, each document is divided into high-relevance documents, medium-relevance documents, and low-relevance documents, resulting in a hierarchical document set related to the research topic.

[0053] The entity-relationship-entity triple structure is the basic data structure for representing facts in knowledge graphs, in the form of (Subject, Predicate, Object). In this application, Subject and Object are limited to three types of entities: "scholar," "classic texts," and "civilization," while Predicate is limited to a predefined relation type, used to precisely define the boundaries of the research topic.

[0054] In the above embodiments, a ternary ontology-based approach is used to decompose natural language topics into structured triples, thereby accurately identifying the core of the topic. Subsequently, a three-tiered progressive retrieval process is executed, progressing from exact matching to semantic extension and relational reasoning, systematically covering the range of literature from direct to potential. Finally, a multi-dimensional quantitative scoring model is introduced to transform fuzzy relevance into precise numerical scores, automatically classifying literature into high, medium, and low relevance categories. This process transforms the previously manual noise reduction and screening work into a directly output, clearly structured hierarchical literature set, enabling scholars to quickly focus on high-value literature groups.

[0055] In the above embodiments, the quantitative relevance score is used as a guiding parameter for downstream processing. The system can dynamically configure differentiated processing strategies based on the relevance level of the documents. This on-demand computing power allocation mechanism based on hierarchical results effectively filters out low-value interference while ensuring the depth of core document analysis. This not only significantly saves computing resources but also significantly improves the signal-to-noise ratio and academic rigor of the final multidimensional structured knowledge base by eliminating noisy data.

[0056] In some embodiments of this application, the specific process of performing a progressive retrieval from precise to generalized based on the combination of core entities and relationships in a preset structured JRAS literature database includes: performing a precise retrieval in the preset structured JRAS literature database using the decomposed core entities and their synonyms as keywords, obtaining a first candidate literature set for the first layer of precise retrieval; obtaining extended attributes of the core entities from a preset Oriental studies vertical knowledge graph, including the scholar's research field, the theme of the classics, and the core characteristics of the civilization; performing a semantic generalization retrieval in the preset structured JRAS literature database using the extended attributes as keywords, obtaining a second candidate literature set for the second layer of semantic extension retrieval; performing one-hop or multi-hop relational reasoning based on the triple structure to search for potential entities that are indirectly related to the research topic; and performing a retrieval in the preset structured JRAS literature database using potential entities as keywords, obtaining a third candidate literature set for the third layer of relational reasoning retrieval.

[0057] One-hop or multi-hop relationship reasoning is a path query algorithm based on graph theory. In a knowledge graph, a hop refers to traversing an edge; one-hop reasoning refers to a direct association (A→B), while multi-hop reasoning refers to an indirect association achieved through intermediate nodes (A→B→C). This process is used to discover potential entities that have indirect logical connections with core issues. The preset threshold range is a pre-defined numerical range used to discretize continuous quantitative association scores into different levels (e.g., high, medium, low), thereby guiding subsequent processing strategies.

[0058] In this embodiment, a three-layer progressive retrieval mechanism is used to ensure accurate targeting of core literature while systematically mining potentially related literature by leveraging the extended attributes and multi-hop reasoning capabilities of the knowledge graph. This effectively solves the problem of flattened traditional retrieval results and achieves full coverage from direct evidence to indirect references and then to potential connections, significantly reducing the cost for scholars to manually sift through massive amounts of noise to find high-value literature.

[0059] In one possible implementation, the user inputs the research topic on the client side as: "Henry Yule's commentary on *The Travels of Marco Polo* and his contribution to the study of Yuan Dynasty China." The system identifies the core concepts in the topic. The scholar entity is Henry Yule, the text entity is *The Travels of Marco Polo*, the civilization entity is Yuan Dynasty China, the relation type is commentary and contribution, and the result is the generated triplet structure (Henry Yule, commentary, *The Travels of Marco Polo*) and (Henry Yule, contribution, Yuan Dynasty China). Based on the decomposed triplets, the system performs a three-level search in a pre-defined structured JRAS literature database: The first-level precise retrieval process is as follows: Strategy: The system uses the core entities and their synonyms (such as "Yule", "Marco Polo", "The Book of Ser Marco Polo") as keywords to perform precise Boolean logic combinations for retrieval.

[0060] Result: 8 articles were found (e.g.) Figure 4 (As shown). Example reference: JRAS 1866, Vol. 3: 'Yule'snotes on Marco Polo'.

[0061] The second-level semantic expansion retrieval process is as follows: Strategy: The system obtains the extended attributes of core entities from the preset Oriental Studies vertical knowledge graph.

[0062] Henry Yull -> Research areas: Indian history, Central Asian history; The Travels of Marco Polo -> Themes: Silk Road, Travel Notes; Yuan Dynasty China; The system uses these extended attributes ("Yuan Dynasty China", "Silk Road", "Travelogue Notes") as keywords for semantic generalization retrieval.

[0063] Result: 32 articles were hit (e.g.) Figure 4 (As shown). Example literature: JRAS 1895, Vol. 27: 'Medieval travels to China'.

[0064] The third-level relational reasoning and retrieval process is as follows: Strategy: The system performs one-hop or multi-hop relation reasoning based on the triple structure.

[0065] Reasoning path 1 (one jump): Henry Yull -> Notes -> The Travels of Marco Polo; Reasoning path 2 (multiple jumps): Henry Yull -> Paul Pelliot -> Paul Pelliot (assuming that Pelliot and Paul Pelliot were scholars who studied this field at the same time); The system uses inferred "potential entities" (such as "Cordie" and "Pelliot") as keywords for retrieval.

[0066] Result: 302 articles were hit (e.g.) Figure 4 (As shown). Example reference: JRAS 1927, Vol. 59: 'Pelliot's corrections to Cordier'.

[0067] For example Figure 5 As shown, Figure 5 This is a schematic diagram of a hierarchical retrieval process provided in this application. First, the input research topic is semantically analyzed and decomposed to extract combinations of core entities and relationships, constructing a triplet structure in the format of entity-relationship-entity. Second, based on this triplet structure, a progressive retrieval is performed on a preset literature database, specifically divided into three levels: the first level uses Boolean logic for precise retrieval to obtain directly matching core entity documents; the second level combines extended attributes in the knowledge graph for semantic generalization retrieval to broaden the scope; and the third level uses one-hop or multi-hop relational reasoning to mine indirectly related potential entities and retrieve them accordingly. Subsequently, all candidate documents obtained from the above three levels are comprehensively evaluated, and their relevance is quantitatively scored based on their semantic matching degree with the research topic, keyword co-occurrence frequency, cross-citation count, and the length of the association path in the knowledge graph. Finally, based on the scoring results and a preset threshold range, the documents are divided into three relevance levels: high, medium, and low, thus obtaining a structured hierarchical literature set.

[0068] S103, paragraph-level extraction is performed on the documents in the hierarchical document set, and a pre-tuned large language model is used to perform structured interpretation of the extracted paragraphs to obtain the interpretation results; Among them, the research topics correspond to core entities. The pre-fine-tuned large language model is a specialized model obtained by using annotated data from a specific domain (such as Orientalism) to perform secondary training on a general basic large language model (such as GPT, Llama, etc.).

[0069] In this embodiment, the specific process of extracting paragraphs from documents in a hierarchical document set and using a pre-tuned large language model to perform structured interpretation of the extracted paragraphs to obtain the interpretation results includes: obtaining the relevance level of each document to be processed in the hierarchical document set, where the relevance level indicates whether each document to be processed is a high-relevance document, a medium-relevance document, or a low-relevance document; when the relevance level indicates that each document to be processed is a high-relevance document, performing a fine-grained scan in the full text of each document to locate all sentences containing core entities and their synonyms and near-synonyms, and merging them into paragraphs to be interpreted; or, when the relevance level indicates that each document to be processed is a medium-relevance document, locating the chapters or paragraphs containing core entities in the full text of each document to be processed, and using the entire chapter or paragraph as the paragraph to be interpreted; or, when the relevance level indicates that each document to be processed is a low-relevance document, locating the paragraph that first mentions any core entity in the full text of each document to be processed, and using that paragraph as the paragraph to be interpreted; and generating interpretation results based on the paragraphs to be interpreted and the pre-tuned large language model.

[0070] In this embodiment, a differentiated paragraph extraction strategy is dynamically configured based on the relevance level of the documents. Specifically, a fine-grained full-text scan is performed on highly relevant documents to maximize information capture, relevant chapters are located for medium-relevance documents to balance efficiency and coverage, and only the first-mentioned paragraphs are extracted from low-relevance documents for rapid filtering. This achieves intelligent allocation of computing resources to high-value documents, avoids the risk of context spillover from large models, and significantly improves the efficiency of structured interpretation and the signal-to-noise ratio of final knowledge integration.

[0071] Specifically, the process of generating interpretation results based on the paragraph to be interpreted and the pre-tuned large language model includes: assigning a maximum token length to the paragraph to be interpreted according to the relevance level of each document; assigning a maximum token length to a text block for documents with high relevance to the document to be interpreted that is less than the maximum token length assigned to documents with medium or low relevance to the document; splitting the paragraph to be interpreted into one or more text blocks according to the assigned maximum token length; inputting each text block into the pre-tuned large language model and configuring interpretation task instructions of different complexities according to the relevance level of the document to which it belongs, so as to obtain structured interpretation sub-results; merging and integrating all interpretation sub-results from the same document to form a comprehensive interpretation result for the document; wherein, the comprehensive interpretation result includes at least the identified entities, relationships between entities, core viewpoints, and supporting original text evidence fragments, and is associated with the document's identifier and original text location information.

[0072] The maximum token length for a text block is the upper limit of characters / words set for a single request (Prompt) or a single data block when using a large language model.

[0073] In this embodiment, by dynamically configuring differentiated text block lengths and interpretation task complexity based on the document relevance level, smaller granularity blocks and high-precision instructions can be used for highly relevant documents to capture deep semantics, while larger blocks and basic instructions can be used for documents with medium to low relevance to balance efficiency. While effectively controlling the context overflow and computational consumption of large models, it achieves accurate structured extraction from micro evidence to macro viewpoints and ensures that each conclusion is anchored to the original document location.

[0074] In one possible implementation, after the previous layer of retrieval, 57 articles are selected, including those with high, medium, and low relevance. The system first obtains the relevance level of each of these 57 articles and then adopts different extraction strategies: Scenario A: Processing highly relevant documents: Scenario: Suppose a document, "JRAS 1890, Vol. 22", is determined to be highly relevant.

[0075] Operation: The system performed a fine-grained scan of the document. Because Cordier vehemently criticized Yule's translation in the article, the system scanned the entire text sentence by sentence, finding not only "Cordier" and "Yule," but also their synonyms / near-synonyms (such as "Cordier" and "Yule's notes"). The system merged all sentences containing these relevant entities to form the paragraph to be interpreted.

[0076] Scenario B: Processing of documents with moderate relevance: Scenario: Suppose a paper, "JRAS 1885, Vol. 17", discusses the geography of Yuan Dynasty China and is classified as having a medium degree of relevance.

[0077] Operation: The system no longer searches the entire text sentence by sentence, but instead locates the chapter or paragraph in the text that specifically discusses "Yuan Dynasty China" (such as Section 3 "The Yuan Dynasty"), and inputs the entire chapter as the paragraph to be interpreted.

[0078] Scenario C: Processing low-relevance documents: Scenario: Suppose a document, "JRAS 1915, Vol. 47", mainly discusses Southeast Asian navigation, and only mentions "influenced by "The Travels of Marco Polo" in the opening introduction, and is judged to have low relevance.

[0079] Operation: The system only locates the paragraph in the entire text that first mentions "Marco Polo", then stops and extracts only this paragraph as the paragraph to be interpreted.

[0080] Then, based on the relevance level of the documents, the system assigns different maximum token lengths to the extracted paragraphs and configures instructions accordingly: Highly relevant literature: Assign shorter token lengths (e.g., 512 tokens). Because the content is highly relevant, it requires highly complex interpretation task instructions (e.g., "Please strictly extract entities and relationships, and provide original text evidence to refute the viewpoint").

[0081] For documents with medium to low relevance: allocate longer token lengths (e.g., 2048 tokens). Since only the main idea needs to be captured, configure low-complexity interpretation task instructions (e.g., "Please summarize whether this paragraph involves core entities").

[0082] Finally, the system inputs the segmented text blocks into a pre-tuned large language model. After the system completes the interpretation, it generates a structured table.

[0083] Entity Recognition Column: The model identifies "Henry Yule" and provides supporting evidence in the Original Evidence Column: "British Orientalist, known for translating and annotating 'The Travels of Marco Polo'", and notes "JRAS 1866, Vol. 3" in the Source Document Column.

[0084] Relationship extraction / core viewpoints, for example Figure 6 and Figure 7 As shown, the model not only lists the entities but also implies their relationships. For example, the relationship between Cordier (entity) and Yule (entity) is "inheritance and development / criticism" (interpreted from the description of evidence), and the relationship between Yuan Dynasty China (entity) and The Travels of Marco Polo (entity) is "historical background".

[0085] Finally, all interpretation sub-results from the same document are merged and integrated to form a comprehensive interpretation result for that document.

[0086] For example Figure 8 As shown, Figure 8This is a schematic diagram of a structured interpretation process provided in this application. First, the system obtains the relevance level of the document (high, medium, low) and executes differentiated preprocessing strategies based on this level: for highly relevant documents, a fine-grained scan is performed to extract sentences containing core entities and their synonyms / near-synonyms, which are then merged into paragraphs to be interpreted; for medium-relevance documents, chapters or paragraphs containing core entities are directly located as content to be interpreted; for low-relevance documents, only paragraphs that mention the core entity for the first time are located as paragraphs to be interpreted. Then, the system assigns different maximum token lengths to the paragraphs to be interpreted based on the document's relevance level (shorter lengths for high relevance, longer lengths for medium and low relevance), and splits the paragraphs into one or more text blocks according to this length. Next, these text blocks are input into a pre-tuned large language model, and interpretation task instructions of different complexities are configured according to the document's relevance level, thereby obtaining structured interpretation sub-results. Finally, the system merges and integrates all interpretation sub-results from the same document to form a comprehensive interpretation result containing entities, relationships, viewpoints, and evidence.

[0087] In this embodiment, the specific process of generating a pre-tuned large language model includes: collecting professional literature, classical commentaries, academic papers, and historical JRAS journal texts in the field of Oriental studies to construct a domain-specific training corpus containing multiple languages; cleaning and annotating the domain-specific training corpus to form a structured training dataset of entity-relation-opinion-evidence quadruples; loading the base large language model and fine-tuning it based on the quadruple structured training dataset to adapt the base large language model to entity recognition, relation extraction, and paragraph-level structured interpretation tasks of scholar entity-classical entity-civilization entity; validating the fine-tuned base large language model, deploying it to the server after successful verification, and configuring it to only receive paragraph-level input from the hierarchical literature set and output interpretation results to obtain the pre-tuned large language model.

[0088] This embodiment constructs a dedicated corpus for Oriental studies that includes multiple languages ​​and fine-tunes instructions based on quadruples. This enables the large model to accurately adapt to entity recognition, relation extraction, and structured interpretation capabilities covering vertical fields of scholars, classics, and civilizations. Thus, while avoiding the illusion of general models, it achieves end-to-end transformation from paragraph-level input to structured output with traceability, providing a high-precision dedicated AI engine for in-depth academic analysis of large-scale JRAS literature.

[0089] S104 preprocesses the interpretation results according to the dimensions of scholars, classics, and civilizations to generate a multi-dimensional structured knowledge base with traceability; based on the multi-dimensional structured knowledge base, it generates a structured report, a visualization map, and traceability archives, and sends them to the client.

[0090] In some embodiments of this application, the specific process of generating structured reports, visualization maps, and source archives based on a multidimensional structured knowledge base includes: injecting newly discovered entities and relationships from the multidimensional structured knowledge base into a preset Oriental studies vertical knowledge graph; based on the injected entities and relationships, using the graph computing reasoning engine built into the knowledge graph, performing multi-hop reasoning along predefined relationship paths to mine and create relationships between entities, thus obtaining the latest knowledge graph; based on the latest knowledge graph, automatically extracting and organizing knowledge content according to preset template dimensions to generate a structured report; calling the visualization engine to generate a visualization map based on the latest knowledge graph, with nodes in the format of scholar-classic-civilization and edges in the preset relationship types and inferred relationships; the visualization map supports node scaling, clicking to view details, and highlighting relationship paths; traversing all conclusion entries in the multidimensional structured knowledge base, extracting their associated document identifiers and original text location information, and organizing them according to document order to generate source archives; the source archives enable each conclusion in the structured report and visualization map to be traced back to a specific paragraph of the original document.

[0091] Furthermore, generating a source tracing archive and executing the source tracing process specifically includes: for each conclusion entry in the multidimensional structured knowledge base, extracting its unique document identifier, and retrieving and retrieving the corresponding document source data from the preset structured JRAS document database based on the identifier; wherein, the document source data includes at least: the author field, publication time field, document title field, and key content abstract field containing the core argument; and structurally binding the document source data with the conclusion entry and the original text location information (including page numbers and paragraph offsets) to generate an archive record containing complete source tracing information.

[0092] Furthermore, in response to a user's request to view a viewpoint in a structured report or a specific edge in a visualization, the system calls up and displays the corresponding archival records to present the author, publication date, title, and key content data upon which the conclusion is based, thereby achieving precise backtracking of the entire chain from analytical conclusions to original documentary evidence.

[0093] One application scenario for tracing archives is as follows: Figure 9 As shown.

[0094] This embodiment utilizes a closed-loop mechanism of dynamic injection, graph reasoning, and multidimensional generation to automatically output structured reports containing scholar lineages, textual transmission, and civilizational contexts based on a visualized graph. It also provides visualized graphs that support interactive exploration and archives with complete traceability chains. This transforms discrete multidimensional knowledge into verifiable and interactive academic results, thereby achieving a leap from data integration to deep knowledge discovery.

[0095] This embodiment uses hierarchical retrieval to precisely select hierarchical literature sets related to the research topic from a pre-defined structured JRAS literature database, focusing the analysis scope from the entire set to a subset. Then, paragraph-level extraction further narrows the processing object from the entire document to core text blocks, completely avoiding the risk of context overflow from large models and achieving systematic and batch processing of tens of thousands of documents. Based on this, a traceable archive is finally generated, establishing an immutable mapping link between AI-automated analysis and original documentary evidence, meeting the stringent standards of academic research for verifiable and retrievable conclusions.

[0096] This embodiment utilizes a pre-constructed vertical knowledge graph of Oriental studies, whose ternary ontology model specifically defines scholars, classics, civilizations, and their various core relationships, providing a structured container for domain knowledge. Simultaneously, a large language model fine-tuned based on Oriental studies corpora is employed for structured interpretation, enabling precise adaptation to entity recognition and relation extraction tasks within the domain. Combined with relational reasoning retrieval and cross-linguistic entity alignment using the graph in hierarchical retrieval, it achieves the transformation from unstructured documents to automatic linking across the entire scholar-classic-civilization domain, supporting in-depth and professional interpretation of Oriental studies topics.

[0097] The following are embodiments of the apparatus described in this application, which can be used to execute the embodiments of the method described in this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the method described in this application.

[0098] Please see Figure 10 This illustration shows a schematic diagram of a multi-stage intelligent analysis device for JRAS literature provided in an exemplary embodiment of this application. This multi-stage intelligent analysis device for JRAS literature can be implemented as all or part of an electronic device through software, hardware, or a combination of both. The device 1 includes a literature database and map loading module 10, a hierarchical retrieval module 20, a structured interpretation module 30, and an analysis result generation module 40.

[0099] The document library and graph loading module 10 is used to respond to the intelligent analysis request sent by the client carrying research topics and load the preset structured JRAS document library and the preset Oriental Studies vertical knowledge graph. The preset structured JRAS document library is built based on JRAS original PDF documents containing multilingual content. The ternary ontology model of the preset Oriental Studies vertical knowledge graph includes scholar entities, classic entities, civilization entities and preset relation types. The preset Oriental Studies vertical knowledge graph is a graph structure knowledge library with clear domain boundaries and semantic constraints, built based on the ternary ontology model and through entity recognition and relation extraction of JRAS documents. The hierarchical retrieval module 20 is used to perform hierarchical retrieval based on the preset structured JRAS literature database and the preset Oriental studies vertical knowledge graph, so as to filter out the hierarchical literature set related to the research topic. The structured interpretation module 30 is used to extract paragraphs from the documents in the hierarchical document set and to perform structured interpretation on the extracted paragraphs using a pre-tuned large language model to obtain the interpretation results (the content of structured interpretation has been described in detail in the method embodiment section of this application and will not be repeated here). The analysis results generation module 40 is used to preprocess the interpretation results according to the dimensions of scholars, classics, and civilizations to generate a multi-dimensional structured knowledge base with traceability; based on the multi-dimensional structured knowledge base, it generates a structured report, a visualization map, and traceability archives, and sends them to the client.

[0100] It should be noted that the multi-stage intelligent analysis device for JRAS documents provided in the above embodiments is only illustrated by the division of the above functional modules when executing the multi-stage intelligent analysis method for JRAS documents. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the multi-stage intelligent analysis device for JRAS documents provided in the above embodiments and the multi-stage intelligent analysis method embodiments for JRAS documents belong to the same concept, and the implementation process is detailed in the method embodiments, which will not be repeated here.

[0101] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0102] This embodiment uses hierarchical retrieval to precisely select hierarchical literature sets related to the research topic from a pre-defined structured JRAS literature database, focusing the analysis scope from the entire set to a subset. Then, paragraph-level extraction further narrows the processing object from the entire document to core text blocks, completely avoiding the risk of context overflow from large models and achieving systematic and batch processing of tens of thousands of documents. Based on this, a traceable archive is finally generated, establishing an immutable mapping link between AI-automated analysis and original documentary evidence, meeting the stringent standards of academic research for verifiable and retrievable conclusions.

[0103] This embodiment utilizes a pre-constructed vertical knowledge graph of Oriental studies, whose ternary ontology model specifically defines scholars, classics, civilizations, and their various core relationships, providing a structured container for domain knowledge. Simultaneously, a large language model fine-tuned based on Oriental studies corpora is employed for structured interpretation, enabling precise adaptation to entity recognition and relation extraction tasks within the domain. Combined with relational reasoning retrieval and cross-linguistic entity alignment using the graph in hierarchical retrieval, it achieves the transformation from unstructured documents to automatic linking across the entire scholar-classic-civilization domain, supporting in-depth and professional interpretation of Oriental studies topics.

[0104] This application also provides a computer-readable medium storing program instructions that, when executed by a processor, implement the multi-stage chaining intelligent analysis method for JRAS literature provided in the above-described method embodiments.

[0105] This application also provides a computer program product containing instructions that, when run on a computer, causes the computer to execute the multi-stage chaining intelligent analysis method for JRAS documents described in the above-described method embodiments.

[0106] Please see Figure 11 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 11 As shown, the electronic device 1000 may include: at least one processor 1001, at least one network interface 1004, a user interface 1003, a memory 1005, and at least one communication bus 1002.

[0107] The communication bus 1002 is used to realize the connection and communication between these components.

[0108] The user interface 1003 may include a display screen and a camera. Optionally, the user interface 1003 may also include a standard wired interface and a wireless interface.

[0109] The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface).

[0110] The processor 1001 may include one or more processing cores. The processor 1001 connects to various parts within the electronic device 1000 using various interfaces and lines. It executes various functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in the memory 1005, and by calling data stored in the memory 1005. Optionally, the processor 1001 may be implemented using at least one hardware form selected from Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), and Programmable Logic Array (PLA). The processor 1001 may integrate one or more of the following: a Central Processing Unit (CPU), a Graphics Processing Unit (GPU), and a modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content to be displayed on the screen; and the modem handles wireless communication. It is understood that the modem may also be implemented as a separate chip, without being integrated into the processor 1001.

[0111] The memory 1005 may include random access memory (RAM) or read-only memory. Optionally, the memory 1005 may include a non-transitory computer-readable storage medium. The memory 1005 can be used to store instructions, programs, code, code sets, or instruction sets. The memory 1005 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the above-described method embodiments, etc.; the data storage area may store data involved in the above-described method embodiments, etc. Optionally, the memory 1005 may also be at least one storage system located remotely from the aforementioned processor 1001. Figure 11 As shown, the memory 1005, which serves as a computer storage medium, may include an operating system, a network communication module, a user interface module, and a multi-stage intelligent analysis application for JRAS literature.

[0112] exist Figure 11In the illustrated electronic device 1000, the user interface 1003 is mainly used to provide an input interface for the user and to acquire user input data; while the processor 1001 can be used to call the multi-stage chaining intelligent analysis application for JRAS literature stored in the memory 1005, and specifically perform the following operations: In response to a client's request for intelligent chain analysis carrying research topics, a pre-defined structured JRAS literature repository and a pre-defined Oriental studies vertical knowledge graph are loaded. The pre-defined structured JRAS literature repository is constructed based on original JRAS PDF documents containing multilingual content. The pre-defined Oriental studies vertical knowledge graph's ternary ontology model includes scholar entities, classic entities, civilization entities, and pre-defined relation types. The pre-defined Oriental studies vertical knowledge graph is a graph-structured knowledge base with clear domain boundaries and semantic constraints, constructed based on the ternary ontology model and through entity recognition and relation extraction of JRAS documents. A hierarchical search is performed based on a pre-defined structured JRAS literature database and a pre-defined Oriental studies vertical knowledge graph to select a hierarchical literature set related to the research topic (the content included in the structured interpretation has been described in detail in the method implementation section of this application and will not be repeated here). Paragraph-level extraction is performed on the documents in the hierarchical document set, and a pre-tuned large language model is used to perform structured interpretation of the extracted paragraphs to obtain the interpretation results; The interpretation results are preprocessed according to the dimensions of scholars, classics, and civilizations to generate a multi-dimensional structured knowledge base with traceability; based on the multi-dimensional structured knowledge base, a structured report, a visualization map, and traceability archives are generated and sent to the client.

[0113] In one embodiment, when processor 1001 performs a hierarchical search based on a preset structured JRAS literature database and a preset Oriental studies vertical knowledge graph to filter out a hierarchical literature set related to the research topic, it specifically performs the following operations: Based on scholar entities, classic text entities, civilization entities, and predefined relation types, semantic analysis and decomposition of the research topic are performed to obtain the combination of core entities and relations, resulting in a triple structure in the format of entity-relation-entity; the core entities correspond to scholar entities, classic text entities, and civilization entities; Based on the combination of core entities and relationships, a progressive retrieval from precise to generalized is performed on the pre-defined structured JRAS document database to obtain the first candidate document set for precise retrieval, the second candidate document set for semantic expansion retrieval, and the third candidate document set for relational reasoning retrieval. For each document in the retrieved candidate document set, a relevance score is quantified based on its semantic matching degree with the research topic, keyword co-occurrence frequency, cross-citation count, and the length of the association path in the knowledge graph. The quantified relevance score is used to guide paragraph-level extraction and structured interpretation. Based on the relevance score and a preset threshold range, each document is divided into high-relevance documents, medium-relevance documents, and low-relevance documents, resulting in a stratified document set related to the research topic.

[0114] In one embodiment, when processor 1001 performs a progressive retrieval from precise to generalized based on a combination of core entities and relationships on a preset structured JRAS literature database, it specifically performs the following operations: In the pre-defined structured JRAS document database, Boolean logic combination is used to perform precise retrieval using the decomposed core entities and their synonyms as keywords, resulting in the first candidate document set for the first layer of precise retrieval. From the pre-set vertical knowledge graph of Oriental studies, obtain the extended attributes of the core entities. The extended attributes include the scholar's research field, the theme of the classics, and the core characteristics of the civilization. In the pre-defined structured JRAS literature database, semantic generalization retrieval is performed using extended attributes as keywords to obtain the second candidate literature set for the second-level semantic extended retrieval. Based on the triplet structure, one-hop or multi-hop relational reasoning is used to search for and study potential entities that are indirectly related to the research topic. In the pre-defined structured JRAS literature database, a search is conducted using latent entities as keywords to obtain the third candidate literature set for third-level relational reasoning retrieval.

[0115] In one embodiment, when the processor 1001 performs paragraph-level extraction on documents in a hierarchical document set and uses a pre-tuned large language model to perform structured interpretation of the extracted paragraphs to obtain the interpretation results, it specifically performs the following operations: Obtain the relevance level of each document to be processed in the hierarchical document set. The relevance level is used to indicate whether each document to be processed is a high-relevance document, a medium-relevance document, or a low-relevance document. If the relevance level indicates that each document to be processed is a highly relevant document, a fine-grained scan is performed on the full text of each document to locate all sentences containing core entities and their synonyms and near-synonyms, and these are merged into paragraphs to be interpreted; or, If the relevance level indicates that each document to be processed is of medium relevance, locate the chapter or paragraph containing the core entity in the full text of each document, and use that chapter or paragraph as a whole as the paragraph to be interpreted; or, If the relevance level indicates that each document to be processed is a low-relevance document, locate the paragraph in the full text of each document to be processed that first mentions any core entity, and use that paragraph as the paragraph to be interpreted. Based on the paragraph to be interpreted and the pre-tuned large language model, the interpretation result is generated.

[0116] In one embodiment, when processor 1001 generates interpretation results based on a pre-tuned large language model of the paragraph to be interpreted, it specifically performs the following operations: The maximum token length of the text block is assigned to the paragraph to be interpreted based on the relevance level of each document to be processed; the maximum token length of the text block assigned to documents with high relevance is less than the maximum token length of the text block assigned to documents with medium and low relevance. The paragraph to be interpreted is split into one or more text blocks according to the maximum token length of the assigned text block; Each text block is input into a pre-tuned large language model, and interpretation task instructions of different complexities are configured according to the relevance level of the document to which it belongs, so as to obtain structured interpretation sub-results. All sub-interpretations from the same document are merged and integrated to form a comprehensive interpretation result for that document; among them, The comprehensive interpretation results should include at least the identified entities, the relationships between entities, the core viewpoints, and the original text evidence fragments used as supporting evidence, and should be associated with the document's identifier and the original text's location information.

[0117] In one embodiment, when the processor 1001 generates structured reports, visualization maps, and traceability archives based on a multidimensional structured knowledge base, it specifically performs the following operations: Newly discovered entities and relationships in the multidimensional structured knowledge base are injected into a pre-defined Oriental studies vertical knowledge graph. Based on the injected entities and relationships, the knowledge graph's built-in graph computing reasoning engine performs multi-hop reasoning along predefined relationship paths to mine and create the associations between entities, thus obtaining the latest knowledge graph. Based on the latest knowledge graph, knowledge content is automatically extracted and organized according to preset template dimensions to generate structured reports; The visualization engine is invoked, and based on the latest knowledge graph, a visual graph is generated with nodes in the format of scholar-classic-civilization and edges with preset relationship types and inferred connections. The visual graph supports node zooming, clicking to view details, and highlighting of relationship paths. By traversing all conclusion entries in the multidimensional structured knowledge base, extracting their associated document identifiers and original text location information, and organizing them in document order to generate a source archive, the source archive enables each conclusion in the structured report and visualization map to be traced back to a specific paragraph in the original document.

[0118] In one embodiment, when the processor 1001 generates a preset structured JRAS literature database and a preset Oriental studies vertical knowledge graph, it specifically performs the following operations: Obtain the original PDF JRAS documentation containing multilingual content; The original JRAS PDF documents are parsed and OCR recognized to obtain the processed text, metadata and translation results; The processed text, metadata, and translation results are stored in the FTS5 full-text search engine database to form a standardized structured document library containing original text location information, which serves as the preset structured JRAS document library. Define a ternary ontology model containing scholar entities, classic entities, and civilization entities, and multiple relation types; scholar entities include master-disciple relationship attributes, classic entities include version transmission attributes, and civilization entities include spatiotemporal range attributes. Based on a ternary ontology model and multiple relation types, a standardized structured literature database is used to generate a pre-defined vertical knowledge graph of Oriental studies.

[0119] In one embodiment, when the processor 1001 generates a preset Oriental studies vertical knowledge graph based on a ternary ontology model and a standardized structured literature database with multiple relation types, it specifically performs the following operations: Identify scholar names, classic titles, and civilization names from standardized structured literature databases; For multilingual document scenarios, a unified identification and alignment of scholar names, classic titles, and civilization names is achieved through a pre-constructed name mapping table and semantic similarity calculation strategy to obtain an entity set; Based on multiple relation types, the associations between entities are extracted from the document texts of a standardized structured document database, and multiple relation instances are constructed. Entity sets and multiple types of relationship instances are injected into a graph database, and pattern matching is performed using a ternary ontology model to generate a preset Oriental studies vertical knowledge graph containing basic relationships.

[0120] In one embodiment, when the processor 1001 generates a pre-fine-tuned large language model, it specifically performs the following operations: Collect professional literature, classical commentaries, academic papers and historical journal texts in the field of Oriental studies to construct a domain-specific training corpus containing multiple languages; The domain-specific training corpus is cleaned and labeled to form a structured training dataset of four tuples: entity-relationship-opinion-evidence. Load the base big language model and fine-tune it with instructions based on the four-tuple structured training dataset so that the base big language model can be adapted to the tasks of entity recognition, relation extraction and paragraph-level structured interpretation of scholar entity, classic entity and civilization entity. The base language model after fine-tuning the instructions is verified. After successful verification, it is deployed to the server and configured to only receive paragraph-level input from hierarchical document sets and output interpretation results, thus obtaining the pre-fine-tuned language model.

[0121] This embodiment uses hierarchical retrieval to precisely select hierarchical literature sets related to the research topic from a pre-defined structured JRAS literature database, focusing the analysis scope from the entire set to a subset. Then, paragraph-level extraction further narrows the processing object from the entire document to core text blocks, completely avoiding the risk of context overflow from large models and achieving systematic and batch processing of tens of thousands of documents. Based on this, a traceable archive is finally generated, establishing an immutable mapping link between AI-automated analysis and original documentary evidence, meeting the stringent standards of academic research for verifiable and retrievable conclusions.

[0122] This embodiment utilizes a pre-constructed vertical knowledge graph of Oriental studies, whose ternary ontology model specifically defines scholars, classics, civilizations, and their various core relationships, providing a structured container for domain knowledge. Simultaneously, a large language model fine-tuned based on Oriental studies corpora is employed for structured interpretation, enabling precise adaptation to entity recognition and relation extraction tasks within the domain. Combined with relational reasoning retrieval and cross-linguistic entity alignment using the graph in hierarchical retrieval, it achieves the transformation from unstructured documents to automatic linking across the entire scholar-classic-civilization domain, supporting in-depth and professional interpretation of Oriental studies topics.

[0123] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program for multi-stage chaining intelligent analysis of JRAS documents can be stored in a computer-readable storage medium. When executed, the program can include the processes of the embodiments of the above methods. The storage medium for the program for multi-stage chaining intelligent analysis of JRAS documents can be a magnetic disk, optical disk, read-only memory, or random access memory, etc.

[0124] The above-disclosed embodiments are merely preferred embodiments of this application and should not be construed as limiting the scope of this application. Therefore, any equivalent variations made in accordance with the claims of this application shall still fall within the scope of this application.

Claims

1. A multi-stage intelligent analysis method for JRAS literature, characterized in that, Applied to the server side, the method includes: In response to a request from a client for intelligent chaining analysis carrying research topics, a pre-defined structured JRAS literature database and a pre-defined Oriental studies vertical knowledge graph are loaded. The pre-defined structured JRAS literature database is constructed based on original JRAS PDF documents containing multilingual content. The pre-defined Oriental studies vertical knowledge graph has a ternary ontology model that includes scholar entities, classic text entities, civilization entities, and pre-defined relationship types. The pre-defined Oriental studies vertical knowledge graph is a graph-structured knowledge base with clear domain boundaries and semantic constraints, constructed based on the ternary ontology model and through entity recognition and relationship extraction of JRAS documents. A hierarchical search is performed based on the pre-defined structured JRAS literature database and the pre-defined Oriental studies vertical knowledge graph to filter out a hierarchical literature set associated with the research topic. This includes: semantic analysis and decomposition of the research topic based on the scholar entity, classic text entity, civilization entity, and pre-defined relationship types to obtain a combination of core entities and relationships, resulting in a triplet structure in the format of entity-relationship-entity; the core entities correspond to the scholar entity, classic text entity, and civilization entity; based on the combination of core entities and relationships, a progressive search from precise to generalized is performed on the pre-defined structured JRAS literature database to obtain the first layer of precise search. The system retrieves a first candidate document set, a second candidate document set for semantic expansion retrieval, and a third candidate document set for relational reasoning retrieval. For each document in the retrieved candidate document set, a relevance score is quantified based on its semantic matching degree with the research topic, keyword co-occurrence frequency, cross-citation count, and the length of the association path in the knowledge graph. The quantified relevance score is used to guide the paragraph-level extraction and the structured interpretation. Based on the relevance score and a preset threshold range, each document is divided into high-relevance documents, medium-relevance documents, and low-relevance documents to obtain a hierarchical document set related to the research topic. Paragraph-level extraction is performed on the documents in the hierarchical document set, and a pre-tuned large language model is used to perform structured interpretation of the extracted paragraphs to obtain the interpretation results. The interpretation results are preprocessed according to the dimensions of scholars, classics, and civilizations to generate a multi-dimensional structured knowledge base with traceability; based on the multi-dimensional structured knowledge base, a structured report, a visualization map, and a traceability archive are generated and sent to the client.

2. The method according to claim 1, characterized in that, Based on the combination of the core entities and relationships, a progressive retrieval from precise to generalized is performed on the preset structured JRAS document database, including: In the preset structured JRAS document database, Boolean logic combination is used as keywords to perform precise retrieval, and the first candidate document set of the first layer of precise retrieval is obtained; From the preset Oriental Studies vertical knowledge graph, obtain the extended attributes of the core entity, including the scholar's research field, the theme of the classics, and the core characteristics of the civilization; In the preset structured JRAS document database, semantic generalization retrieval is performed using the extended attributes as keywords to obtain the second candidate document set for the second-level semantic extended retrieval. Based on the triplet structure, perform one-hop or multi-hop relation reasoning to search for potential entities that are indirectly related to the research topic; In the preset structured JRAS document database, the potential entities are used as keywords for retrieval to obtain the third candidate document set for the third-level relational reasoning retrieval.

3. The method according to claim 1, characterized in that, The research topic corresponds to a core entity; the paragraph-level extraction of documents from the hierarchical document set is performed, and a pre-tuned large language model is used to perform structured interpretation of the extracted paragraphs to obtain the interpretation results, including: Obtain the relevance level of each document to be processed in the hierarchical document set. The relevance level is used to indicate whether each document to be processed is a high-relevance document, a medium-relevance document, or a low-relevance document. If the relevance level indicates that each document to be processed is a highly relevant document, a fine-grained scan is performed on the full text of each document to locate all sentences containing the core entity and its synonyms and near-synonyms, and these sentences are merged into paragraphs to be interpreted; or, If the relevance level indicates that each document to be processed is a medium-relevance document, then the chapter or paragraph containing the core entity is located in the full text of each document to be processed, and that entire chapter or paragraph is taken as the paragraph to be interpreted; or, If the relevance level indicates that each document to be processed is a low-relevance document, the paragraph in the full text of each document to be processed that first mentions any of the core entities is located, and that paragraph is used as the paragraph to be interpreted. Based on the paragraph to be interpreted and the pre-tuned large language model, the interpretation result is generated.

4. The method according to claim 3, characterized in that, The step of generating interpretation results based on the paragraph to be interpreted and the pre-tuned large language model includes: Based on the relevance level of each document to be processed, the maximum token length of the text block to be interpreted is assigned to the paragraph to be interpreted; the maximum token length of the text block assigned to high-relevance documents is less than the maximum token length of the text block assigned to medium- and low-relevance documents. The paragraph to be interpreted is split into one or more text blocks according to the maximum token length of the assigned text block; Each text block is input into the pre-tuned large language model, and interpretation task instructions of different complexities are configured according to the relevance level of the document to which it belongs, so as to obtain structured interpretation sub-results. All sub-interpretations from the same document are merged and integrated to form a comprehensive interpretation result for that document.

5. The method according to claim 4, characterized in that, The comprehensive interpretation results include at least the identified entities, the relationships between entities, the core viewpoints, and supporting evidence fragments from the original text, and are associated with the document's identifier and original text location information.

6. The method according to claim 1, characterized in that, The generation of structured reports, visual maps, and source tracing archives based on the multidimensional structured knowledge base includes: The newly discovered entities and relationships in the multidimensional structured knowledge base are injected into the preset Oriental Studies vertical knowledge graph; Based on the injected entities and relationships, the graph computing reasoning engine built into the knowledge graph is used to perform multi-hop reasoning along predefined relationship paths to mine and create the association relationships between entities, and obtain the latest knowledge graph. Based on the latest knowledge graph, knowledge content is automatically extracted and organized according to preset template dimensions to generate a structured report; The visualization engine is invoked to generate a visualization graph based on the latest knowledge graph. The graph uses the format "scholar-classic-civilization" as nodes and the preset relationship types and inferred connections as edges. The visualization graph supports node zooming, clicking to view details, and highlighting of relationship paths. Traverse all conclusion entries in the multidimensional structured knowledge base, extract their associated document identifiers and original text location information, and organize them in document order to generate a source traceability archive; the source traceability archive enables each conclusion in the structured report and the visualization map to be traced back to a specific paragraph of the original document.

7. The method according to any one of claims 1-6, characterized in that, Generate a pre-defined structured JRAS literature database and a pre-defined Oriental studies vertical knowledge graph by following these steps: Obtain the original PDF JRAS documentation containing multilingual content; The original JRAS PDF document is parsed and OCR is performed to obtain the processed text, metadata and translation results; The processed text, metadata, and translation results are stored in the FTS5 full-text search engine database to form a standardized structured document library containing original text location information, which serves as the preset structured JRAS document library. Define a ternary ontology model comprising scholar entity, classic entity, and civilization entity, and multiple relation types; the scholar entity includes a mentorship relationship attribute, the classic entity includes a version transmission attribute, and the civilization entity includes a spatiotemporal range attribute. Based on the aforementioned ternary ontology model, multiple relation types, and the standardized structured literature database, a preset Oriental studies vertical knowledge graph is generated.

8. The method according to claim 7, characterized in that, The process of generating a pre-defined vertical knowledge graph of Oriental studies based on the ternary ontology model, multiple relation types, and the standardized structured literature database includes: Identify scholar names, classic titles, and civilization names from the standardized structured literature database; For multilingual document scenarios, a unified identification and alignment of the scholar names, classic titles and civilization names is performed through a pre-constructed name mapping table and semantic similarity calculation strategy to obtain an entity set; Based on the aforementioned multiple relationship types, the associations between entities are extracted from the document text of the standardized structured document database to construct multiple relationship instances; The entity set and the multiple relation instances are injected into the graph database, and pattern matching is performed in conjunction with the ternary ontology model to generate a preset Oriental studies vertical knowledge graph containing basic relationships.

9. The method according to any one of claims 1-6, characterized in that, Generate a pre-tuned large language model by following these steps: Collect professional literature, classical commentaries, academic papers and historical journal texts in the field of Oriental studies to construct a domain-specific training corpus containing multiple languages; The domain-specific training corpus is cleaned and labeled to form a structured training dataset of four tuples: entity-relationship-opinion-evidence. Load the base language model and fine-tune it according to the instructions based on the quadruple structured training dataset so that the base language model can be adapted to the tasks of entity recognition, relation extraction and paragraph-level structured interpretation of scholar entity, classic entity and civilization entity. The base language model after fine-tuning the instructions is verified. After successful verification, it is deployed to the server and configured to only receive paragraph-level input from the hierarchical document set and output the interpretation results, thus obtaining the pre-fine-tuned language model.

10. A multi-stage intelligent analysis device for JRAS literature implemented using the method described in any one of claims 1-9, characterized in that, The device includes: The document library and graph loading module is used to respond to a linked intelligent analysis request sent by the client carrying research topics, and load a preset structured JRAS document library and a preset Oriental studies vertical knowledge graph; the preset structured JRAS document library is based on multilingual content. Rongde The preset Oriental Studies vertical knowledge graph is constructed from JRAS original PDF documents. The ternary ontology model includes scholar entities, classic text entities, civilization entities, and preset relation types. The preset Oriental Studies vertical knowledge graph is a graph structure knowledge base with clear domain boundaries and semantic constraints, constructed based on the ternary ontology model and through entity recognition and relation extraction of JRAS documents. The hierarchical retrieval module is used to perform hierarchical retrieval based on the preset structured JRAS literature database and the preset Oriental studies vertical knowledge graph, so as to filter out the hierarchical literature set associated with the research topic. The structured interpretation module is used to extract paragraphs from the documents in the hierarchical document set and to perform structured interpretation on the extracted paragraphs using a pre-tuned large language model to obtain the interpretation results. The analysis results generation module is used to preprocess the interpretation results according to the dimensions of scholars, classics, and civilizations to generate a multi-dimensional structured knowledge base with traceability; based on the multi-dimensional structured knowledge base, it generates a structured report, a visualization map, and a traceability archive, and sends them to the client.

Images