Document fusion retrieval system and method thereof

Abstract

The invention provides a document fusion retrieval system and a document fusion retrieval method. The document fusion retrieval system comprises a storage device and a processor. The processor executes the plurality of modules and receives a natural language instruction. The plurality of modules comprise a multi-path recall module, an automatic merging retrieval module and a sentence window retrieval module. The multi-path recall module recall the candidate document block set from the index database according to the natural language instruction, and the automatic merging retrieval module performs merging and filtering processing on the candidate document block set to generate filtered candidate blocks. A sentence window retrieval module extracts text before and after the filtered candidate blocks to generate an expanded context. And generating a retrieval result according to the context through a large language model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of complex document processing and information retrieval processes, and in particular to a document fusion retrieval system and method thereof. BACKGROUND

[0002] With the digital transformation of enterprises, due to the fact that the documents of enterprises are scattered in different storage media and lack of unified data standards and models, it is difficult to achieve effective data management. Especially when it is necessary to manage and process unstructured data, it is difficult to effectively process complex document content containing tables, multi-column layout or long text. As a result, users face many obstacles when looking for and obtaining the required documents, and data cannot effectively serve business needs. Therefore, how to establish a system that can process complex text and provide accurate recall and management of unstructured data has become the development direction of improving the efficiency of enterprise data utilization. SUMMARY

[0003] The present application is a document fusion retrieval system and method thereof, which can automatically perform fusion retrieval, automatic merge retrieval and sentence window retrieval according to an index database and a natural language instruction to generate a retrieval result.

[0004] According to an embodiment of the present application, the document fusion retrieval system of the present application includes a storage device and a processor. The storage device stores a plurality of modules and an index database. The processor is coupled to the storage device. The processor executes the plurality of modules and receives a natural language instruction. The plurality of modules include a multi-path recall module, an automatic merge retrieval module and a sentence window retrieval module. The multi-path recall module recalls a candidate document block set from the index database according to the natural language instruction, and the automatic merge retrieval module performs merge and filtering processing on the candidate document block set to generate a filtered candidate block. The sentence window retrieval module extracts the context before and after the filtered candidate block to generate an expanded context. The retrieval result is generated according to the context by a large language model.

[0005] According to an embodiment of the present application, the document fusion retrieval method of the present application includes the following steps: receiving a natural language instruction; recalling a candidate document block set from an index database according to the natural language instruction by a multi-path recall module; performing merge and filtering processing on the candidate document block set by an automatic merge retrieval module to generate a filtered candidate block; extracting the context before and after the filtered candidate block by a sentence window retrieval module to generate an expanded context; and generating a retrieval result according to the context by a large language model.

[0006] Based on the above, the document fusion retrieval system and method thereof of the present application can retrieve relevant indexes according to a natural language instruction, merge and expand the candidate blocks of the corresponding indexes, and generate a retrieval result based on a large language model.

[0007] So that the foregoing features and advantages of the present application can be understood in detail, a brief description of embodiments of the present application is provided below with reference to the attached drawings. BRIEF DESCRIPTION OF DRAWINGS

[0008] Figure 1 is a schematic diagram of a document fusion retrieval system of an embodiment of the present application;

[0009] Figure 2 is a flowchart of a document fusion retrieval method of an embodiment of the present application;

[0010] Figure 3 is a schematic diagram of a document fusion retrieval system and modules of an embodiment of the present application;

[0011] Figure 4 is a flowchart of a multi-dimensional semantic based relevant passage recall of an embodiment of the present application;

[0012] Figure 5 is a flowchart of a document parsing and generating a comprehensive index set of an embodiment of the present application.

[0013] BRIEF DESCRIPTION OF DRAWINGS

[0014] 100: a document fusion retrieval system;

[0015] 110: a processor;

[0016] 120: a storage device;

[0017] 121: a multi-path recall module;

[0018] 122: an automatic merge retrieval module;

[0019] 123: a sentence window retrieval module;

[0020] 124: a comprehensive document parsing module;

[0021] 125: a semantic enhancement module;

[0022] 126: an index database;

[0023] 400: a natural language instruction;

[0024] 401: a block;

[0025] 402: a natural language index;

[0026] 403: a spoken language index;

[0027] 404: a keyword index;

[0028] 405: a retrieval answer;

[0029] S210-S250, S410-S440, S510-S560: steps. DETAILED DESCRIPTION

[0030] Reference will now be made in detail to the exemplary embodiments of the present application, examples of which are illustrated in the accompanying drawings. Wherever possible, the same reference numbers are used in the drawings and the description to refer to the same or like parts.

[0031] Figure 1 is a schematic diagram of a document fusion retrieval system according to an embodiment of the present application. Referring to Figure 1 , the document fusion retrieval system 100 includes a processor 110 and a storage device 120. The processor 110 is coupled to the storage device 120. The storage device 120 stores a multi-recall module 121, an automatic merge retrieval module 122, and a sentence window retrieval module 123. In this embodiment, the document fusion retrieval system 100 can be, for example, set in a cloud server for users to connect and execute relevant business service functions of different application programming interfaces (APIs) also set in the cloud server. The cloud server can be, for example, a software as a service (SaaS) server, and the application programming interfaces correspond to SaaS applications, but the present application is not limited thereto. Alternatively, in an embodiment, the document fusion retrieval system 100 can be set in an enterprise resource planning (ERP) system run by an on-premise server in an enterprise to integrate and process enterprise documents from multiple sources through different application programming interfaces.

[0032] In the embodiment, the processor 110 may, for example, include a central processing unit (CPU), or other programmable general purpose or special purpose microprocessors (Microprocessor), digital signal processors (DSP), application specific integrated circuits (ASIC), programmable logic devices (PLD), other similar processing circuitry, or a combination of these devices. The storage device 120 may, for example, include a memory and / or a database, wherein the memory may, for example, be a non-volatile memory (NVM). The storage device 120 may, for example, store programs, modules, systems or algorithms related to the implementation of the embodiments of the present application, for the processor 110 to access and execute to realize the related functions and operations described in the embodiments of the present application. In the embodiment, the multi-channel recall module 121, the automatic merging retrieval module 122, the sentence window retrieval module 123, the comprehensive document parsing module 124 and the semantic enhancement module 125 may, for example, be implemented in a program language such as JSON (JavaScript Object Notation), Extensible Markup Language (XML) or YAML, but the present application is not limited thereto.

[0033] In the embodiment, the user may, for example, execute the document fusion retrieval system 100 through a personal computer device, and input a natural language instruction to the document fusion retrieval system 100 to execute corresponding document retrieval and retrieval result generation. The document fusion retrieval system 100 may, for example, automatically execute the multi-channel recall module 121, the automatic merging retrieval module 122, the sentence window retrieval module 123, the comprehensive document parsing module 124 and the semantic enhancement module 125 according to the natural language instruction to automatically generate the corresponding retrieval result. In the embodiment, the user may, for example, execute the document fusion retrieval system 100 to realize the effect of accurate recall of document paragraphs by combining keyword retrieval, hierarchical structure retrieval and sentence window retrieval, and to generate the retrieval result of the corresponding natural language instruction based on the recalled text. Moreover, the document fusion retrieval system 100 may, for example, adopt a specific weight division method to ensure that the multi-dimensional information can be comprehensively considered to quickly and accurately recall the document paragraphs most relevant to the user query (i.e., the natural language instruction).

[0034] Figure 2 is a flowchart of a document fusion retrieval method according to an embodiment of the present application. Referring toFigure 1 and Figure 2 The document fusion retrieval system 100 can perform the following steps S210-S250 to automatically generate the retrieval result corresponding to the natural language instruction. In the present embodiment, the user can input the natural language instruction to the document fusion retrieval system 100 to obtain the retrieval result generated by the document fusion retrieval system 100. In step S210, the processor 110 receives the natural language instruction. In step S220, the candidate document block set is recalled from the index database according to the natural language instruction.

[0035] In an embodiment, the index database stores comprehensive index groups. In other words, the index database is a vector database storing multiple index groups. The comprehensive index groups include natural language joint index, question group joint index, and word group joint index, that is, the vector representations of natural language blocks, question groups, and feature word groups are stored in the index database. The natural language joint index is the joint index of original standard blocks and natural language blocks, which records the correspondence between the two. In an embodiment, the index can be a hash table, the key is the original block identifier, and the value is the content and position of the natural language block. Similarly, the question group joint index is the joint index of the naturalized question group and the original standard block. The word group joint index is the joint index of the group and the original standard block. The candidate document block set is all blocks related to the natural language instruction recalled from the natural language block, the naturalized question group, and the feature word group through the related index in the comprehensive index group.

[0036] In step S230, the processor 110 executes the automatic merging retrieval module 122 to perform merging and filtering processing on the candidate document block set to generate filtered candidate blocks. Specifically, the automatic merging retrieval module 122 retrieves all document blocks in the selected candidate document block set, and filters out duplicate document blocks and merges document blocks related to each other, thereby improving the accuracy of recall.

[0037] In step S240, the processor 110 executes the sentence window retrieval module 123 to extract the context before and after the filtered candidate blocks to generate an expanded context. In other words, the sentence window retrieval module 123 retrieves the original document at the sentence level, and expands the filtered candidate blocks based on the original document context, thereby improving the details and context coherence of the recall.

[0038] In step S250, the retrieval result is generated according to the context by the large language model. That is, the processor 110 inputs the expanded context and the natural language instruction into the large language model to obtain the retrieval result corresponding to the natural language instruction. In this way, the document fusion retrieval system and the method thereof can automatically retrieve relevant documents and generate corresponding retrieval answers according to the natural language instruction.

[0039] Figure 3is a schematic diagram of a document fusion retrieval system and modules of an embodiment of the present invention. Reference is made to Figure 1 and Figure 3 The storage device 120 also stores a comprehensive document parsing module 124, a semantic enhancement module 125, and an index database 126. Figure 4 is a flowchart of a multi-dimensional semantic relevant passage recall of an embodiment of the present invention. Reference is made to Figure 1 , Figure 3 and Figure 4 The embodiment details the multi-dimensional semantic relevant passage recall. At step S410, the processor 110 executes the multi-pass recall module 121 to perform fusion retrieval by the multi-pass recall module 121. Specifically, the multi-pass recall module 121 first parses the natural language instruction 400 to generate keywords and semantic vectors, and then the multi-pass recall module 121 retrieves the index database 126 according to the keywords and semantic vectors, and further matches corresponding indexes from the comprehensive index set in the index database 126. The comprehensive index set includes a natural language index 402 (i.e., a natural language joint index), a spoken language index 403 (i.e., a question set joint index), and a keyword index 404 (i.e., a keyword set joint index).

[0040] That is, the multi-pass recall module 121 performs fusion retrieval combining keyword retrieval and semantic retrieval according to the keywords and semantic vectors. For example, the natural language instruction 400 is “2023Q3 revenue growth reasons”, and the keywords parsed by the multi-pass recall module 121 are “revenue growth reasons”. In this way, the multi-pass recall module 121 matches the feature keyword set index (i.e., the keyword index 404) in the comprehensive index set by keyword retrieval, and further recalls document blocks containing “revenue growth” or “cost optimization” from the multiple blocks 401. Moreover, the multi-pass recall module 121 performs semantic retrieval to match the semantic vectors of the natural language blocks and the spoken language question sets, and further recalls relevant document blocks from the multiple blocks 401. The blocks 401 can include original standard blocks, natural language blocks, formatted document blocks, spoken language question sets, and keyword feature sets. Then, the multi-pass recall module 121 inputs all document blocks recalled according to the indexes as a candidate document block set to the automatic merging retrieval module 122.

[0041] In an embodiment, the multi-pass recall module 121 performs a comprehensive search on the keywords and semantic vectors by a sparse search algorithm and a semantic search algorithm to obtain relevant indexes. The sparse search algorithm can be term frequency-inverse document frequency (tf-idf) or Best Matching 25 (BM25). The semantic search algorithm is, for example, a vector-based similarity search algorithm. The multi-pass recall module 121 can use at least two sparse search algorithms, each having a first weight value and a second weight value, which can be represented by the following equation (1), wherein w1 and w2 are the weights of the tf-idf and BM25 search algorithms, respectively, and can be adjusted according to the actual application scenario.

[0042] Score fusion = w1 Score tf-idf + w2 Score BM25 …(1)

[0043] At step S420, the processor 110 performs the automatic merge search module 122 to perform merging and filtering processing on the candidate document block set. First, the automatic merge search module 122 divides the candidate document block set into a multi-level block having a hierarchical structure. Specifically, the hierarchical structure includes at least one parent block and a plurality of child blocks corresponding thereto. In this way, the automatic merge search module 122 merges the child blocks corresponding to the same parent block in the candidate document block set, and filters the duplicate blocks in the candidate document block set based on vector similarity, thereby generating filtered candidate blocks. In other words, the automatic merge search module 122 divides the candidate document block set into a hierarchical structure, and performs search on the smallest leaf block (and child block). When multiple leaf blocks refer to the same parent block (i.e., parent block), the leaf blocks are merged, and the parent block is taken as the new search result. The merging result can be represented by the following equation (2), wherein Mergeparent represents the merged parent block, and Refi represents the number of times that the ith child block refers to the parent block.

[0044]

[0045] At step S430, the processor 110 executes the sentence window retrieval module 123 to perform sentence window retrieval. Specifically, the sentence window retrieval module 123 uses cosine similarity search to search the semantic similarity of each sentence based on the filtered candidate blocks to expand the context window to k sentences of text before and after, k being a positive integer. In an embodiment, k is 5, but the present case should not be limited thereto. The sentence retrieval algorithm can be as shown in equation (3), in which Contextsentence is the context window of the target sentence, Sentence target is the target sentence, and k is the size of the context window. In other words, the sentence window retrieval module 123 extracts k sentences of text before and after the candidate block. For example, the natural language instruction 400 is "How is the sales of product line A?", the matched recall colloquial question group is "P001", and the expanded context sentence is "Product line A launched a promotion activity in Q3 2023, with a 20% year-on-year increase in sales, driving overall revenue growth." At step S440, the processor 110 executes a large language model (LLM) to summarize and generate a retrieval answer 405 according to the context window and the natural language instruction 400.

[0046] Context sentence = [Sentence target ± k]…(3)

[0047] In another embodiment, the processor 110 combines the recall candidate document block set, the filtered candidate blocks, and the context based on a weighted fusion method to update the retrieval result, wherein the weighted fusion method includes a first weight value, a second weight value, and a third weight value corresponding to the recall candidate document block set, the filtered candidate blocks, and the context, respectively. The specific weight distribution formula (4) is as follows, in which Final Score is the final fusion score, Step1, Step2, and Step3 are the scores of steps S410, S420, and S430, respectively. The score of each step is calculated by the respective retrieval algorithm, for example, step S410 can use the BM25 algorithm to obtain the score, step S420 can be the score obtained based on the automatic merging retriever, and step S430 can be the score obtained by performing sentence window retrieval.

[0048] Final Score = 0.6 · Score Step1 + 0.3 · Score Step2 + 0.1 · Score Step3 …(4)

[0049] In the present embodiment, the weight coefficients are 0.6, 0.3 and 0.1 respectively, but the weights of steps S410, S420 and S430 can be set according to the importance and accuracy in the recall process, so the present case should not be limited thereto. In this way, the contributions of different retrieval strategies are balanced by setting the weights to optimize the overall recall effect, and to ensure that the algorithm can more accurately recall relevant information while reducing the interference of irrelevant or redundant information.

[0050] Regarding the establishment of the comprehensive index set, the comprehensive document parsing module 124 performs document preprocessing and semantic enhancement processing on the original document to generate the comprehensive index set, and stores the comprehensive index set in the index database 126. Figure 5 is a flowchart of document parsing and generation of a comprehensive index set according to an embodiment of the present application. Referring to Figure 1 , Figure 3 and Figure 5 , the present embodiment details the execution flow of document parsing and generation of a comprehensive index set. In an embodiment, the comprehensive document parsing module 124 includes a preprocessing module, an adaptive table conversion module and a dynamic adjustment table module, and the processor 110 executes the preprocessing module to perform document parsing in step S510. Specifically, the preprocessing module performs data cleaning operations on the original document to remove noise information in the original document. Moreover, after removing the noise information from the original document, the preprocessing module generates structured table data to improve the accuracy of subsequent semantic retrieval or model processing. The noise information can be format errors, irrelevant characters or semantically disordered text.

[0051] In step S520, the processor 110 executes the adaptive table conversion module to intelligently identify the cells that can be merged. That is, the adaptive table conversion module establishes a cell mapping table according to the structured table data generated in step S510, and judges whether the structured table has a cross-cell attribute to update the cell mapping table. For example, the cell mapping table records the position information (row index and column index) of each cell in the HTML table, and whether these cells participate in merging. Moreover, the cell mapping table also records the start position and end position of the merged cells. Then, the adaptive table conversion module converts the structured table data into a format marked document according to the cell mapping table. The format marked document can be a Markdown table.

[0052] At step S530, the processor 110 executes a dynamic table adjustment module to dynamically adjust the table layout in the formatted markup document. In an embodiment, the dynamic table adjustment module adjusts the table layout of the formatted markup document based on the structured table data, and performs a segmentation operation on the formatted markup document to generate the original standard blocks. The table layout includes column width, alignment, and blank cells. The dynamic table adjustment module presents the alignment format by appropriate space filling in the formatted markup document according to the alignment format (e.g., left alignment, center alignment, right alignment) of the cells in the original document (e.g., HTML table). The dynamic table adjustment module performs a segmentation operation on the formatted markup document based on semantics to generate the original standard blocks. That is, each block in the original standard blocks corresponds to an independent semantic unit, which can be a paragraph or a table. For example, the original document is an annual report of a company containing a sales data table, and after document preprocessing, the table is converted into Excel structured data, and the text part is segmented into multiple original standard blocks according to chapters.

[0053] At step S540, the processor 110 executes a semantic enhancement module 125 to perform natural language segmentation. Specifically, the semantic enhancement module 125 performs semantic analysis on the original standard blocks by a large language model, thereby obtaining natural language blocks and a natural language joint index. The natural language joint index is a joint index of the original standard blocks and the natural language blocks, used to record the correspondence between the two. The semantic enhancement module 125 performs segmentation according to semantic boundaries by semantic analysis of the original document, thereby preserving the original semantic structure of the original document, and thus segmenting the document content into natural language blocks that are easier to understand and facilitate subsequent retrieval. When performing natural language segmentation, the processor 110 can input a first prompt template instruction to the semantic enhancement module 125. The first prompt template instruction may, for example, be “Please summarize the following content in simple language and extract the core logic.” For example, the original standard block is “2023Q3 revenue increased by 15% year-on-year, mainly due to the increase in sales of product line A and cost control.” The natural language block is “2023Q3 revenue increased by 15%, mainly due to the increase in sales of product line A and cost optimization.” The natural language joint index of the above example is that the original block identifier “P001” corresponds to the natural language block identifier “NL001”.

[0054] At step S550, the processor 110 then performs oralization question-answer generation by the semantic enhancement module 125. Specifically, the semantic enhancement module 125 obtains an oralization question set and a question set joint index based on the natural language chunk by a large language model. The oralization question set includes a plurality of questions and a plurality of answers corresponding to each question, and the number of answers corresponding to each question is within 20 tokens. The questions in the oralization question set include a plurality of synonymous or near-synonymous user query modes. For example, the natural language chunk is “2023 third quarter revenue growth of 15%, mainly due to product line A sales increase and cost optimization.”, and the oralization question set is “How did the third quarter revenue grow?” “What is the sales situation of product line A?” or “What measures were taken for cost control?”, and the question set joint index is the index of the above oralization question set associated with the original chunk identifier chunk “P001”. When performing oralization question-answer generation, the processor 110 can input a second prompt template instruction to the semantic enhancement module 125, which may, for example, be “How can the user ask the following in an oralized way? List 3 questions”. In this step, the semantic enhancement module 125 generates at least one set of oralization question-answer pairs for each document chunk, thereby improving the user's interactive experience and the practicality and convenience of the document fusion retrieval system 100.

[0055] At step S560, the processor 110 performs multi-feature recognition generation by the semantic enhancement module 125. Specifically, the semantic enhancement module 125 extracts key words from the oralization question set to obtain a key feature word set and a word set joint index. The comprehensive index set includes a natural language joint index, a question set joint index, and a word set joint index. The key feature word set may, for example, be a noun phrase or a core verb, and the word set joint index is the joint index of the word set and the original standard chunk. When performing multi-feature recognition generation, the processor 110 inputs the oralization question set and a third prompt template instruction to the semantic enhancement module 125, which may, for example, be “Extract core keywords from the following questions for quick retrieval”. The key words (i.e., the key feature word set) may, for example, be “revenue growth”, “product line A sales”, or “cost control”. The document fusion retrieval system 100 can extract key information by extracting key words, and generate key word labels (i.e., word set joint indexes) corresponding to the key information, respectively, to facilitate quick retrieval and positioning of relevant information. In this way, the semantic enhancement module 125 enhances the retrievability of the document content, and the user can quickly find the required document paragraphs through the key word labels, thereby improving the efficiency of information retrieval.

[0056] To sum up, the document fusion retrieval system and method thereof can enhance the retrievability of document content according to the established index group, so that the user can quickly obtain the required document paragraph through the keyword tag, and the efficiency of information retrieval is improved. The document fusion retrieval system combines keyword retrieval, hierarchical structure retrieval and sentence window retrieval, and realizes accurate recall of the document paragraph. Moreover, the document fusion retrieval system can convert the unstructured complex document into structured data, so as to efficiently perform information retrieval and knowledge management.

[0057] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, but not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A document fusion retrieval system, characterized in that, include: Storage device, storing multiple modules and an index database; as well as The processor, coupled to the storage device, executes the plurality of modules and receives natural language instructions, wherein the plurality of modules includes a multiplexed retrieval module, an automatic merging retrieval module, and a sentence window retrieval module. The multi-path recall module recalls a set of candidate document blocks from the index database according to the natural language instructions, and the automatic merging retrieval module performs merging and filtering processing on the set of candidate document blocks to generate filtered candidate blocks. The sentence window retrieval module extracts the text before and after the filtered candidate block to generate an expanded context, and the processor generates retrieval results based on the context using a large language model.

2. The document fusion retrieval system according to claim 1, characterized in that, The sentence window retrieval module uses cosine similarity search based on the filtered candidate blocks to query the semantic similarity of each sentence, and then expands the context window to the preceding and following k sentences of text, where k is a positive integer.

3. The document fusion retrieval system according to claim 1, characterized in that, The automatic merging and retrieval module divides the candidate document block set into multi-level blocks with a hierarchical structure, wherein the hierarchical structure includes at least one parent block and its corresponding multiple child blocks. The automatic merging retrieval module merges the sub-blocks corresponding to the same parent block in the candidate document block set, and filters the duplicate blocks in the candidate document block set based on vector similarity, thereby generating the filtered candidate blocks.

4. The document fusion retrieval system according to claim 1, characterized in that, The processor updates the search results by combining the recalled candidate document block set, the filtered candidate blocks, and the context using a weighted fusion method. The weighted fusion method includes a first weight value, a second weight value, and a third weight value corresponding to the recalled candidate document block set, the filtered candidate blocks, and the context, respectively.

5. The document fusion retrieval system according to claim 1, characterized in that, The multi-path recall module parses the natural language instructions to generate keywords and semantic vectors. The multi-path recall module retrieves the index database based on the keywords and semantic vectors, and then matches the corresponding index from the comprehensive index group. The multi-path recall module recalls the candidate document block set according to the index and inputs it into the automatic merging retrieval module.

6. The document fusion retrieval system according to claim 5, characterized in that, The multi-path recall module performs a comprehensive search on the keywords and the semantic vectors using a sparse retrieval algorithm and a semantic retrieval algorithm to generate the index, wherein the sparse retrieval algorithm has a first weight value and a second weight value.

7. The document fusion retrieval system according to claim 1, characterized in that, The plurality of modules also include: The comprehensive document parsing module performs document preprocessing and semantic enhancement on the original document to generate a comprehensive index group, and stores the comprehensive index group in the index database.

8. The document fusion retrieval system according to claim 7, characterized in that, The comprehensive document parsing module includes: The preprocessing module performs data cleaning operations on the original document to remove noise information and generate structured tabular data. An adaptive table conversion module creates a cell mapping table based on the structured table data, determines whether the structured table has a cross-cell attribute, and then updates the cell mapping table. The adaptive table conversion module converts the structured table data into a formatted document based on the cell mapping table. The dynamic adjustment table module adjusts the table layout of the formatted document based on the structured table data, and performs a segmentation operation on the formatted document to generate original standard blocks.

9. The document fusion retrieval system according to claim 8, characterized in that, The plurality of modules also include: The semantic enhancement module uses a large language model to semantically analyze the original standard blocks, thereby obtaining natural language blocks and a natural language federated index. The semantic enhancement module obtains colloquial question groups and a joint index of question groups based on the natural language blocks using the large language model. The semantic enhancement module extracts keywords based on the colloquial question group to obtain key feature phrases and phrase joint indexes, wherein the comprehensive index group includes the natural language joint index, the question group joint index, and the phrase joint index.

10. The document fusion retrieval system according to claim 9, characterized in that, The colloquial question group includes multiple questions and corresponding multiple answers, wherein the number of multiple answers corresponding to each of the multiple questions is no more than 20 lexical units.

11. A document fusion retrieval method, characterized in that, include: Receive natural language commands; The candidate document block set is recalled from the index database by the multi-path recall module according to the natural language instructions; The candidate document block set is merged and filtered by the automatic merging retrieval module to generate filtered candidate blocks. The sentence window retrieval module extracts the text before and after the filtered candidate blocks to generate an expanded context. as well as Search results are generated based on the context using a large language model.

12. The document fusion retrieval method according to claim 11, characterized in that, The step of extracting the text before and after the filtered candidate block through the sentence window retrieval module to generate the expanded context includes: The sentence window retrieval module uses cosine similarity search based on the filtered candidate blocks to query the semantic similarity of each sentence, thereby expanding the context window to the preceding and following k sentences of text, where k is a positive integer.

13. The document fusion retrieval method according to claim 11, characterized in that, The steps for performing a merging process on the candidate document block set include: The automatic merging and retrieval module divides the candidate document block set into hierarchical multi-level blocks, wherein the hierarchical structure includes at least one parent block and multiple corresponding child blocks; and The automatic merging and retrieval module merges the sub-blocks corresponding to the same parent block in the candidate document block set, and filters duplicate blocks in the candidate document block set based on vector similarity, thereby generating the filtered candidate blocks.

14. The document fusion retrieval method according to claim 11, characterized in that, Also includes: The processor updates the retrieval results by combining the recalled candidate document block set, the filtered candidate blocks, and the context using a weighted fusion method, wherein the weighted fusion method includes a first weight value, a second weight value, and a third weight value corresponding to the recalled candidate document block set, the filtered candidate blocks, and the context, respectively.

15. The document fusion retrieval method according to claim 11, characterized in that, The steps of recalling the candidate document block set from the index database using the multi-path recall module according to the natural language instructions include: The natural language instructions are parsed by the multi-path recall module to generate keywords and semantic vectors; The multi-path recall module retrieves the index database based on the keywords and semantic vectors, and then matches the corresponding index from the comprehensive index group; and The candidate document block set is recalled by the multi-path recall module according to the index and input into the automatic merging retrieval module.

16. The document fusion retrieval method according to claim 15, characterized in that, The steps of retrieving the index database using the multi-path recall module based on the keywords and the semantic vector include: The multi-path recall module performs a comprehensive search on the keywords and the semantic vectors based on the sparse retrieval algorithm and the semantic retrieval algorithm to generate the index, wherein the sparse retrieval algorithm has a first weight value and a second weight value.

17. The document fusion retrieval method according to claim 11, characterized in that, Before the step of recalling the candidate document block set from the index database, the method further includes: The original document is preprocessed and semantically enhanced by the comprehensive document parsing module to generate a comprehensive index group, which is then stored in the index database.

18. The document fusion retrieval method according to claim 17, characterized in that, The steps of performing the document preprocessing on the original document include: The preprocessing module performs data cleaning on the original document to remove noise information and generate structured tabular data. An adaptive table conversion module creates a cell mapping table based on the structured table data, determines whether the structured table has cross-cell attributes, and updates the cell mapping table accordingly. The adaptive table conversion module then converts the structured table data into a formatted document based on the cell mapping table. The table layout of the formatted document is adjusted by the dynamic adjustment table module based on the structured table data, and the formatted document is segmented to generate the original standard blocks.

19. The document fusion retrieval method according to claim 18, characterized in that, The steps of performing the semantic enhancement processing on the original document to generate the comprehensive index set include: The semantic enhancement module uses a large language model to perform semantic analysis on the original standard blocks, thereby obtaining natural language blocks and natural language joint indexes; The semantic enhancement module utilizes the large language model to obtain colloquial question groups and a joint index of question groups based on the natural language blocks; and The semantic enhancement module extracts keywords from the colloquial question group to obtain key feature phrases and phrase joint indexes, wherein the comprehensive index group includes the natural language joint index, the question group joint index, and the phrase joint index.

20. The document fusion retrieval method according to claim 19, characterized in that, The colloquial question group includes multiple questions and corresponding multiple answers, wherein the number of multiple answers corresponding to each of the multiple questions is no more than 20 lexical units.

Images