A plug-in semantic enhancement RAG method for formula symbol-intensive technical documents

Abstract

Existing retrieval enhancement methods suffer from problems such as inconsistent symbol meanings, difficulty in fine-grained retrieval of formula and diagram evidence, and insufficient cross-location association in complex technical document applications. This invention studies a plug-in semantic enhancement RAG method for formula- and symbol-intensive technical documents: First, the multimodal parsing results of the technical document are obtained. Based on this, a global symbol table containing symbols, physical meanings, associated formulas, and contextual evidence is constructed according to the candidate symbol content. Then, referring to the construction results of the global symbol table, structured evidence metadata is generated for formulas, images, and tables respectively, and bound to the original document fragments to form an enhanced document representation. Finally, without changing the existing retrieval method, a dual-path retrieval based on the parallel use of enhanced and original indexes is constructed, and the final answer is generated and selected through position alignment and evidence fusion.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and intelligent document processing technology, and in particular to document-side semantic enhancement, multimodal evidence organization, and retrieval enhancement generation technology for formula-symbol-intensive technical documents. Background Technology

[0002] Complex technical documents are widely found in fields such as radar, communications, aerospace, electronic information, and automatic control. They serve as crucial carriers of technical principles, parameter definitions, design processes, algorithm models, and experimental results. Compared to general business documents, some technical documents are characterized by a large number of formulas, dense mathematical symbols, frequent reuse of specialized variables, and interrelationships between images, tables, and text descriptions; these can be termed formula-symbol-intensive technical documents. In such documents, technical knowledge is typically not expressed solely in continuous natural language but is distributed throughout formula derivations, variable definitions, parameter descriptions, graphs, data tables, figure captions, and text descriptions. The same mathematical symbol may be defined earlier, reused in subsequent formulas, and further associated with image indicators, table results, or technical conclusions. Therefore, effective utilization of such documents requires not only recognizing the text content but also maintaining the semantic connections between symbols, formulas, multimodal evidence, and their source locations.

[0003] Traditional technical documents are primarily managed and used in the form of paper manuals, electronic documents, portable document formats, and interactive electronic technical manuals (IETMs). Paper documents are intuitive and easy to archive, but suffer from low retrieval efficiency, difficulty in updating and maintaining, and limited cross-chapter association capabilities. Ordinary electronic documents can improve information location efficiency through table of contents navigation and keyword matching, but when user questions involve variable definitions, formula derivations, cross-page evidence, or relationships between multiple diagrams, manual searching and comprehensive analysis across different chapters is still required. IETMs offer structured directories, workflow navigation, multimedia presentations, and interactive operations, but their information organization typically relies on pre-designed directories, tags, relationships, or business processes, lacking the ability to automatically understand and dynamically associate the semantic relationships between formulas, symbols, and diagrams implicit within technical documents.

[0004] In recent years, with the development of large language models and Retrieval-Augmented Generation (RAG) technologies, intelligent retrieval and question answering based on complex technical documents have become an important direction for improving the efficiency of technical data utilization. Compared with methods that rely solely on the internal parameter knowledge of large language models, RAG can use specified technical documents as the basis for answers, thereby improving the domain adaptability, factual consistency, and traceability of the answers. However, most existing RAG methods use natural language text fragments or complete pages as the basic retrieval objects. When applied to formula- and symbol-intensive technical documents, there are still significant limitations. For text-block-based RAG methods, document segmentation may separate symbol definitions from the subsequent formula usage locations, resulting in the loss of the technical meaning of the same symbol across different retrieval units; at the same time, key technical information in formulas, images, and tables is difficult to fully express using only ordinary text similarity. For page-level multimodal retrieval methods, although page layout and visual content can be preserved, when using complete pages as the retrieval unit, it may still be difficult to explicitly reveal the fine-grained evidential semantics such as the function of formulas, the meaning of variables, image observation indicators, and table conclusions within the page.

[0005] Furthermore, semantic relationships in formula- and symbol-intensive technical documents often exhibit cross-location and cross-modal characteristics. For example, the same symbol may appear repeatedly on different pages or in different chapters, and its specific meaning needs to be determined by combining the definition, the formula it appears in, and the context of the chapter. The technical meaning of the variables involved in a formula may come from previous definitions, while the technical results produced by the formula may be reflected in subsequent images or tables. If existing RAG methods only index each document fragment independently, they are prone to problems such as semantic drift of symbols, unclear formula functions, insufficient expression of multimodal evidence, and missing cross-location evidence associations, thereby affecting the completeness of search results and the accuracy of generated answers.

[0006] Some existing methods explicitly organize document information by modeling graph structures, knowledge graphs, or complex cross-modal relationships. However, these methods typically require additional entity construction, relation extraction, graph structure maintenance, or dedicated retrieval process design, resulting in high modification costs when integrating with existing RAG systems. On the other hand, generating uniform text summaries or general metadata only for images, tables, or formulas is insufficient to fully reflect the specific technical semantics of different evidence types, and cannot effectively address issues such as the cross-location reuse of mathematical symbols and the lack of physical meaning in formula variables.

[0007] Therefore, it is necessary to introduce a plug-in document-side semantic enhancement mechanism, compatible with the existing RAG architecture, to address the content characteristics of formula-symbol-intensive technical documents. This mechanism would perform document-level semantic association on technical symbols reused across different locations, generate searchable semantic representations for different types of evidence such as formulas, images, and tables, and maintain the source association between the enhanced semantic information and the original document evidence. Through this approach, the searchability, semantic consistency, and traceability of formulas, symbols, and multimodal technical evidence can be improved without significantly altering the existing retrieval and generator structures. This addresses the shortcomings of existing RAG methods in processing formula-symbol-intensive technical documents. Summary of the Invention

[0008] This invention constructs a plug-in semantic enhancement RAG method for formula-symbol-intensive technical documents. The core content of this invention is attached. Figure 1 As shown, the main components include: In the document storage and index building stages, firstly, based on the parsing results of text, formulas, images, tables, and captions in complex technical documents, a global symbol table based on formulas and their context is constructed. When the same symbol has multiple meanings, the global default semantics or locally limited semantics of the symbol are determined according to explicit definitions, chapter scope, and source location, providing a basis for symbol interpretation for subsequent evidence semantic enhancement, enabling variables, indicators, or symbols involved in formulas, images, and tables to be mapped to their corresponding technical or physical meanings. On this basis, semantic enhancement fields are set for formulas, images, and tables respectively, generating structured evidence semantics containing information such as formula roles, variable meanings, image indicators, and table conclusions. Then, the above enhanced semantics are associated with the original document units according to their source location to form an enhanced semantic index. In the retrieval stage, the system simultaneously uses the enhanced semantic index and the original document index for dual-path retrieval. The enhanced path recalls fine-grained semantic evidence of symbols, formulas, images, and tables, while the original path preserves the original page, paragraph, and layout context, and aligns and merges the retrieval results according to their source location to generate a candidate evidence set and the final answer. This method can improve the evidence retrievability, semantic consistency, and answer traceability of formula-symbol-intensive technical documents without changing the underlying retrieval and generator structure.

[0009] The main steps of this invention are as follows: Figure 2 As shown, it mainly includes steps 200 to 250.

[0010] Step 200: Obtain the technical document to be processed, perform multimodal parsing on the document, extract document units such as text, formulas, images, tables and figure captions, and record the page number, number and source location information corresponding to each document unit.

[0011] Specifically, the technical document can be a PDF file, a Word file, a technical report, a system manual, a professional paper, a test report, or other electronic documents containing professional knowledge. After parsing the input document, the system obtains several multimodal document units, each of which includes at least the unit content, unit type, and source location.

[0012] Let the original technical document be The set of multimodal document units obtained from the parsing can then be represented as:

[0013]

[0014] in, Indicates the first Each document unit This indicates the total number of document units. Each document unit... This can be further expressed as:

[0015]

[0016] in, Indicates the content of a document unit. Indicates the document unit type, Indicates the source location of a document unit within the original document. This indicates the context information adjacent to the document unit.

[0017] The document unit type It must include at least one of the following: text, formulas, images, tables, and captions. The source location is... It must include at least one of the following: page number, paragraph number, formula number, figure number, table number, or page coordinates.

[0018] Step 210: Extract candidate symbols from formulas and their associated text, figure captions and table fields, gather contextual evidence of the same symbol from different sources, and construct a global symbol table after symbol normalization and semantic ambiguity elimination.

[0019] Specifically, as shown in the attached document Figure 3 As shown, the system extracts candidate symbols from formula units, definitional text, figure captions, table fields, and adjacent paragraphs. These candidate symbols include variable symbols, subscript symbols, superscript symbols, physical quantity names, abbreviations, and parameter identifiers appearing in the technical documentation.

[0020] The candidate symbol set can be represented as:

[0021]

[0022] in, Indicates the first One candidate symbol, This indicates the total number of candidate symbols.

[0023] For each candidate symbol The system collects its contextual evidence. The contextual evidence includes at least one of the following: the formula in which the symbol appears, adjacent text segments, figure captions, table fields, chapter titles, page number positions, and positions preceding or following it.

[0024] Furthermore, a global symbol table is constructed based on candidate symbols and their contextual evidence. The system performs normalization processing on the candidate symbols, merging duplicate symbols, synonyms, and symbol variations caused by typesetting differences. The normalization processing includes, but is not limited to, removing layout noise, standardizing subscript and superscript representations, standardizing Greek letter representations, standardizing capitalization rules, and standardizing mathematical symbol formats.

[0025] For candidate symbols The system relies on its contextual evidence. Generate symbolic semantic descriptions And record the location of the symbol's origin. Construct a global symbol table :

[0026]

[0027] in, Indicates candidate symbols, The semantic description of the candidate symbols. Indicates the source location of the candidate symbol. This represents contextual evidence for candidate symbols.

[0028] The semantic description At a minimum, it should include one of the following: the physical meaning of the symbol, its unit, the formula it belongs to, the associated index, and the applicable context. For example, in radar documentation, the symbol R could represent range, range map, or radar representation; the symbol v could represent velocity; and the symbol SNR could represent signal-to-noise ratio. The system determines the specific meaning of a symbol based on contextual evidence to reduce semantic drift caused by symbol reuse in long documents.

[0029] When the same symbol has multiple possible meanings, the system prioritizes the earliest explicit definition in the document as the global default semantic description and retains subsequent context as supporting evidence; when the same symbol has significantly different meanings in different chapters, the system establishes a locally limited semantic description based on the chapter scope and source location.

[0030] Step 220: Based on the evidence types of formulas, images, and tables, the corresponding semantic enhancement fields are used for processing to generate evidence semantic entries containing formula roles, variable meanings, observation indicators, key values, and conclusion information.

[0031] Specifically, as shown in the attached document Figure 4 As shown, the system organizes formulas, images, tables, captions, and adjacent short text fragments into local evidence units. The set of evidence units can be represented as:

[0032]

[0033] in, Indicates the first One unit of evidence, Indicates the total number of evidence units.

[0034] For each unit of evidence The system identifies its evidence type. :

[0035]

[0036] in, Indicates formulaic evidence, Indicates visual evidence, Indicates tabular evidence, This refers to textual evidence.

[0037] The identification of evidence types can be accomplished through rule matching, layout parsing, document structure parsing, or machine learning classification methods. For example, units containing mathematical expressions or formula numbers are identified as formula evidence; units containing graphs, structure diagrams, flowcharts, or experimental diagrams are identified as image evidence; and data units containing row and column structures are identified as tabular evidence.

[0038] Furthermore, the system sets different semantic enhancement fields for different evidence types to generate structured evidence metadata. For evidence units... Its structured evidence metadata Represented as:

[0039]

[0040] in, Representation and type of evidence The corresponding semantic enhancement function, This represents the local context of the evidence unit. Represents the global symbol table.

[0041] When the evidence unit is a formula, the structured evidence metadata includes at least the formula role, formula expression, involved variables, variable meanings, physical indicators, applicable conditions, and source location. For example, for radar signal processing formulas, the system can extract the variables used to represent range estimation, velocity estimation, beamforming, signal-to-noise ratio calculation, or image reconstruction, etc.

[0042] When the evidence unit is an image, the structured evidence metadata includes at least the image type, image subject, corresponding image indicators, meaning of the horizontal and vertical axes, image caption evidence, supporting conclusions, and source location. For example, for a radar experimental curve, the system can extract its corresponding detection probability, false alarm probability, range resolution, velocity resolution, signal-to-noise ratio, or error indicators.

[0043] When the evidence unit is a table, the structured evidence metadata includes at least the table fields, comparison objects, key values, indicator names, optimal results, table conclusions, and source location. For example, for an algorithm performance comparison table, the system can extract the accuracy, recall, computational complexity, and experimental conclusions of different algorithms.

[0044] When the evidence unit is text, the structured evidence metadata includes at least subject keywords, technical objects, technical functions, constraints, concluding statements, and source location.

[0045] Step 230: Convert the global symbol table and evidence semantic entries into searchable semantic fields, and establish associations with the corresponding original document units based on the source location information to form an enhanced document representation and an enhanced semantic index that runs parallel to the original index.

[0046] Specifically, the system binds original text fragments, formulas, images, tables, and other content with their corresponding semantic enhancement information to form an enhanced document representation. :

[0047]

[0048] in, Represents a fragment of the original document or a multimodal fragment. This indicates semantic enhancement information associated with the fragment. This indicates the source location of the fragment within the original document. This indicates the total number of enhanced representation units.

[0049] The semantic enhancement information It must include at least one of the symbol semantic descriptions from the global symbol table and evidence semantic enhancement metadata. (Based on source location) This enhances the document representation and maintains a correspondence with the original document pages, paragraphs, formulas, images, and tables.

[0050] Furthermore, a text-based retrieval index field is constructed. The system serializes the global symbol table and structured evidence metadata into a text-based index field to facilitate integration with existing text vector search engines or hybrid search engines. The text-based index field may include symbol description fields, formula semantic fields, image semantic fields, table semantic fields, and evidence source fields.

[0051] For example, the textual index field of formulaic evidence can be represented as: "Formula role: distance-velocity relationship calculation; variables involved: R represents distance, v represents velocity; physical indicators: distance resolution, velocity resolution; source location: formula n on page p."; the textual index field of image evidence can be represented as: "Image type: experimental curve; horizontal axis: signal-to-noise ratio, vertical axis: detection probability; supporting conclusion: the proposed method has better detection performance under low signal-to-noise ratio conditions; source location: figure n on page p."

[0052] By using the above methods, formulas, images, and tables that are originally difficult to use directly by text search engines are transformed into searchable semantic descriptions, thereby enhancing search recall capabilities.

[0053] Step 240: Receive user queries, perform dual-path retrieval based on the enhanced semantic index and the original document index respectively, and match, associate and deduplicate the retrieval results of the two paths according to the source location information to construct a candidate evidence set.

[0054] Specifically, as shown in the attached document Figure 5 As shown, the user enters a query. Then, the system represents the enhanced documents respectively. and the original document representation Perform a search:

[0055]

[0056]

[0057] in, This indicates the results of enhanced path retrieval. This represents the results of the original path retrieval. Enhanced paths are used to recall fine-grained evidence containing symbolic semantics, formulaic semantics, image semantics, and table semantics; original paths are used to preserve original page, original paragraph, and layout context information.

[0058] Furthermore, the system aligns the search results based on their source location to construct a candidate evidence set:

[0059]

[0060] in, This represents a set of evidence indicating that the enhanced path and the original path share a common source location. This represents the set of joint evidence after deduplication.

[0061] By employing a dual-path retrieval system, the system can simultaneously utilize enhanced semantic evidence and original contextual evidence, thereby improving evidence coverage and response reliability in complex technical issues.

[0062] Step 250: Generate multiple candidate answers based on different candidate evidence sets, determine the final answer based on the relevance, consistency and evidence completeness of the user query and candidate answers, and simultaneously output the original document source supporting the answer.

[0063] Specifically, for the candidate evidence set Each candidate evidence set in System call generates model Generate candidate answers :

[0064]

[0065] in, This indicates a user query. Represents the candidate evidence set. This represents the candidate answer generated based on the candidate evidence set.

[0066] The system will then process the user's query. and multiple candidate answer input answer selection module To get the final answer :

[0067]

[0068] in, This indicates the final output answer. The answer selection module can be implemented using a large language model, a rule selector, or a fusion ranking model. The answer selection module does not need access to the standard answer or external supervision information; it only judges consistency, completeness, and relevance of evidence based on the user's question and candidate answers.

[0069] Finally, the system outputs the final answer. Simultaneously, the system returns the original document source location supporting the answer. This source location includes at least one of the following: page number, paragraph number, formula number, figure number, table number, and evidence unit number. By outputting the evidence source, users can trace the basis of the answer, improving the credibility of Q&A results for formula-intensive technical documents. Attached Figure Description

[0070] To clearly illustrate the technical solution of this invention, all the accompanying drawings used in this specification are briefly described below. It should be noted that the following drawings are merely some embodiments of this invention, and those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0071] Appendix Figure 1 This is a schematic diagram of the system architecture of the present invention, including modules for global symbol table construction, enhanced evidence semantic representation, enhanced index construction, dual-path retrieval, and answer generation and selection.

[0072] Appendix Figure 2 This is the overall implementation process of the present invention;

[0073] Appendix Figure 3 This is a schematic diagram of the global symbol table construction process in this invention;

[0074] Appendix Figure 4 This is a schematic diagram of the semantic enhancement process for formula, image, and tabular evidence in this invention;

[0075] Appendix Figure 5 This is a schematic diagram of the dual-path retrieval process in which the enhanced path and the original path are used in parallel, as described in this invention.

[0076] Appendix Figure 6 This is an application example of the present invention in a radar technology document question-and-answer scenario. Detailed Implementation

[0077] The steps and processes of the present invention will now be described in full and clearly with reference to the accompanying drawings. Obviously, the embodiments described in this invention are merely one example application scenario of the invention, and other results based on the content of this invention without making substantial changes are within the protection scope of this invention.

[0078] Appendix Figure 6 This is an application example of the present invention in a radar technology document question-and-answer scenario. The specific processing procedure includes the following steps:

[0079] Step 300: Input a radar technology document. This document contains radar signal processing methods, range-velocity diagram explanations, radar image examples, experimental performance tables, and several formula derivations. After receiving this document, the system uses it as the original document D, and extracts multimodal content through rule matching and layout parsing to obtain text blocks, formulas, images, tables, figure captions, and page positions. The system records the page number, block number, heading level, and adjacent context for each document unit.

[0080] Step 310: Extract candidate symbols from the analysis results and construct a global symbol table. The system extracts symbols from the formula. The system identifies symbols and extracts their potential physical meanings from adjacent text. For symbols not defined in the current paragraph, the system continues searching for explicit definitions within the document's surrounding text. A global symbol table is then constructed, generating a symbol entry for each candidate symbol. For example:

[0081] Symbol: R; Meaning: Distance or distance direction representation; Unit: meter; Source: Formula n on page p; Contextual evidence: Adjacency formula and figure caption.

[0082] Symbol: v; Meaning: Radial velocity; Unit: meters per second; Source: Formula m on page p; Contextual evidence: Velocity estimation formula and distance-velocity diagram explanation.

[0083] With a global symbol table, the system can map symbols to their physical meanings in subsequent searches and question answering, instead of treating them as ordinary characters.

[0084] Step 320: Construct a set of evidence units and perform semantic enhancement according to the evidence type. The system first combines formulas, images, tables, their captions, and adjacent explanatory text into local evidence units. Specifically, for example, a distance-angle plot, its caption, text explanation, and relevant formulas are combined into an image evidence unit; a performance comparison table, its table name, column names, and text conclusions are combined into a table evidence unit. Furthermore, for formulaic evidence, the system generates formula roles, variable meanings, physical indicators, and formula expressions; for image evidence, the system generates image type, indicator meanings, and supporting conclusions; and for table evidence, the system generates comparison objects, key values, and table conclusions.

[0085] For example, for the radar range-velocity correlation formula, the system generates the following semantically enhanced result:

[0086] Evidence type: Formula; Formula role: Radar target range and velocity representation; Variables involved: R represents range, v represents radial velocity; Associated image: Range-velocity plot; Source location: Formula n on page p.

[0087] In step 330, the system associates the original document fragment with the corresponding global symbol table entry and evidence semantic enhancement metadata to form an enhanced document representation. This enhanced document representation includes both the original text and page information, as well as structured semantic information. Furthermore, a vector embedding model is used to construct the enhanced index. Simultaneously, the system retains the original document index to provide page-level context and original evidence.

[0088] Step 340: Receive user queries. For example, after a user inputs: "In the BEV-Fusion framework, what complementary information do radar range-angle maps and range-velocity maps provide, and how are they combined to form a radar representation?", the system performs searches on both the enhanced vector index and the original vector index based on existing retrieval methods. The enhanced path retrieval system retrieves evidence related to range-angle maps, range-velocity maps, radar representations, and the meaning of formula variables from the global symbol table and the evidence semantic enhancement index; the original path retrieval system retrieves relevant pages, figure captions, and text paragraphs from the original document.

[0089] Step 350: Construct candidate evidence sets. The system aligns the retrieval results of the enhanced path and the original path based on their source locations, forming an enhanced path evidence set, an original path evidence set, an intersection evidence set, and a union evidence set. Then, the system uses each of these evidence sets to generate candidate answers, ensuring that the answers reflect the document content from different evidentiary perspectives.

[0090] Step 360: Select and output the final answer. The answer selection module selects the answer that is semantically complete, provides the most sufficient evidence, and best matches the user's query as the final answer based on the user's question and candidate answers. In this step, this method uses a large model to autonomously score the final answer. Simultaneously, it outputs the page number, figure number, formula number, or table number corresponding to the answer.

[0091] To further illustrate the effectiveness of this invention, a set of comparative experiments are presented below. Table 1 shows the question-answering accuracy results compared to three baseline methods, and Table 2 shows the ablation experiment results conducted by this invention on the MMLongBench-Doc dataset.

[0092] Table 1 shows the question-answering accuracy across different document domains in MMLongBench-Doc (all results are percentages).

[0093]

[0094] Table 2 shows the ablation experimental results on the MMLongBench-Doc dataset (all results are percentages).

[0095]

[0096] Table 1 compares and validates the method of this invention with the basic RAG, page-level RAG, and SimpleDoc methods on long document datasets. The comparison results show that the method of this invention achieves better overall question-answering performance across different document domains. Compared to the SimpleDoc method, the accuracy of the method of this invention increases from 44.12% to 49.50% for technical papers, from 43.84% to 45.89% for research reports, and from 53.09% to 61.73% for industry documents. These results demonstrate that this invention, through document-side semantic enhancement and dual-path retrieval fusion, can improve the quality of evidence recall and the effectiveness of answer generation when formulas, symbols, charts, and text content are all involved in question answering.

[0097] Furthermore, the analysis focused on different types of evidence. Results showed that for multi-page questions requiring evidence organization across multiple pages, the question-answering accuracy of the method presented in this invention increased from 54.08% for the SimpleDoc method to 57.46%; for single-page questions, the accuracy increased from 73.24% to 75.31%. Since formula-symbol-intensive technical documents often have symbol definitions, formula derivations, charts, and text descriptions distributed across different pages, the above results demonstrate that this invention, through global symbol semantic grounding, evidence semantic enhancement, and dual-path retrieval fusion based on source location, can better maintain the semantic relevance between evidence across different locations.

[0098] Table 2 further validates the roles of different modules in this invention through ablation comparison. The results show that the improved question-answering accuracy after adding only evidence semantic enhancement to the SimpleDoc method demonstrates that generating type-specific structured semantic entries for evidence units such as formulas, images, and tables can improve the retrievalability of multimodal evidence. After further adding a global symbol table, the accuracy of multi-page questions further increased from 55.21% to 57.46%, and the accuracy of technical paper documents further increased from 47.55% to 49.50%, indicating that the global symbol table can provide technical or physical meaning support for variables in formula evidence, alleviating the semantic inconsistency problem caused by symbol reuse across locations.

[0099] Beneficial effects

[0100] This invention presents a plug-in semantic enhancement RAG method for formula-symbol-intensive technical documents. While maintaining compatibility with existing RAG architectures, it improves the searchability and comprehensibility of formulas, symbols, and diagrams in formula-symbol-intensive technical documents, helping to address the shortcomings of traditional electronic technical documentation systems and general RAG methods in intelligent question answering of complex technical documents. Specifically, by constructing a global symbol table, mathematical symbols, physical quantities, units, related formulas, and source locations in technical documents are bound together, thus mitigating the semantic drift problem caused by the significant distance between symbol definitions and usage locations in long documents. Through evidence semantic enhancement methods, formulas, images, and tables are generated with formula roles, variable meanings, physical indicators, image evidence, table fields, key values, and conclusion information, transforming multimodal non-textual evidence, which is difficult for ordinary text search engines to fully utilize, into searchable semantic representations. Finally, through a source location binding mechanism, the enhanced semantic information is mapped to the original document pages, formulas, images, and tables, enabling the system to retain the original document context and evidence traceability while retrieving enhanced semantic evidence. Experimental results further demonstrate that the proposed method does not require modification of the underlying retrieval and generator, nor does it require the construction of a complex multimodal knowledge graph. It can be integrated into existing RAG systems as a plug-in module, and features low implementation cost, strong compatibility, and good engineering feasibility.

Claims

1. A plug-in semantic enhancement retrieval and generation method for formula-symbol-intensive technical documents, characterized in that, include: The technical document to be processed is parsed to obtain text units, formula units, image units, table units, and figure caption units, and the source location information of each document unit is recorded; Candidate symbols are extracted based on formula units and their associated contexts. Contextual evidence of the same candidate symbol from different source locations is collected to construct a global symbol table containing the symbol, its semantic description, and its source location. The global symbol table is used to semantically interpret the variables, indicators, or symbols involved in formula units, image units, and table units, and structured evidence semantic entries are generated according to evidence type. The global symbol table and structured evidence semantic entries are associated with the corresponding original document units according to their source location information to form an enhanced document representation, and an enhanced semantic index is constructed in parallel with the original document index. User queries are received, and enhanced path retrieval results and original path retrieval results are obtained based on the enhanced semantic index and the original document index, respectively. The two types of search results are aligned, correlated, and merged based on the source location information to construct a candidate evidence set. Candidate answers are then generated based on the candidate evidence set, and the final answer is determined based on the relevance, consistency, and completeness of the evidence between the user query and the candidate answers.

2. The method according to claim 1, characterized in that, The construction of the global symbol table includes: standardizing candidate symbols, unifying the formats of subscripts, superscripts, mathematical characters, and symbols; collecting contextual evidence of the same candidate symbol in formulas, text, figure captions, table fields, or chapter titles; when the same symbol has multiple possible meanings, adopting the earliest explicit definition in the technical document as the global default semantic description, and retaining subsequent context as auxiliary evidence; when the same symbol has different meanings in different chapters or different source locations, establishing a locally limited semantic description based on the chapter scope and source location.

3. The method according to claim 1, characterized in that, Structured evidence semantic entries are generated based on the type of evidence. Specifically: Semantic entries for formulaic evidence include formula roles, formula expressions, involved variables, variable meanings, and technical indicators, with the variable meanings provided or corrected by a global symbol table; Semantic entries for image evidence include image type, caption evidence, coordinate meanings, observation indicators, and conclusion support information, with the observation indicators or symbols semantically interpreted based on a global symbol table; Semantic entries for tabular evidence include table fields, comparison objects, key values, indicator names, and table conclusions, with the indicator names or symbols semantically interpreted based on a global symbol table.

4. The method according to claim 1, characterized in that, The formation of the enhanced document representation includes: establishing a semantic reference relationship between the variables, indicators, or symbols involved in the structured evidence semantic entries and the corresponding symbol semantic descriptions in the global symbol table; writing or associating the symbol semantic descriptions, semantic scope, and symbol source locations to the corresponding enhanced semantic fields; when the corresponding symbol has a locally limited semantic description, selecting a matching locally limited semantic description based on the chapter scope or source location of the evidence unit; and ensuring that the enhanced document representation simultaneously contains original evidence information, evidence type semantic information, and symbol semantic information.

5. The method according to claim 1, characterized in that, Enhanced path retrieval results are used to recall fine-grained evidence containing symbolic semantics, formula semantics, image semantics, or table semantics, while original path retrieval results are used to recall original document content and page layout context. The alignment, association, and fusion include: matching the enhanced path retrieval results and the original path retrieval results based on at least one source location information, such as page number, paragraph number, formula number, figure number, table number, page coordinates, or evidence unit number; constructing the source-matched retrieval results into a set of associated evidence; and deduplicating and fusing the two types of retrieval results to form a set of candidate evidence.

Images