A trusted academic question answering method and device based on focused rag

Abstract

The application discloses a kind of based on FocusedRAG's trusted academic question and answer method and device, belong to artificial intelligence, information retrieval field.Step includes: download academic paper data as original data set from public data source;The document content in original data set is split along section boundary, and the explicit retrieval intent of user query is rewritten using large language model;Design query adaptive signal-to-noise separation mechanism filters out the chapter relevant to query as retrieval space;In selected section, expand query adaptive signal-to-noise separation mechanism to obtain sentence level signal-to-noise separation threshold, then design sentence selection mechanism to obtain candidate evidence interval, based on the cumulative correlation score of candidate evidence interval sorting, finally under the constraint of retrieval token budget, candidate evidence interval is greedily selected to form retrieval context;Retrieval context and user query are input into large language model to generate final answer.The application can establish accurate attribution basis, significantly improve the accuracy and credibility of question and answer.

Description

Technical Field

[0001] This invention belongs to the fields of artificial intelligence and information retrieval, and specifically relates to a trustworthy academic question-answering method and device based on FocusedRAG. Background Technology

[0002] Retrieval-Augmented Generation (RAG) is a technique that combines large language models with external knowledge bases to generate more accurate and reliable text. The core of RAG technology lies in retrieving relevant information from external knowledge bases to provide contextual support to the large language model, thereby improving the quality and accuracy of the generated content. In the field of academic paper processing, the application value of RAG technology is particularly prominent. For example, in academic research assistance, RAG can provide researchers with precise knowledge support by quickly retrieving and analyzing relevant literature, accelerating the research process; in intelligent question-answering systems, RAG can generate high-quality question-and-answer results based on academic documents, meeting the needs of academic users for in-depth information. This not only improves the efficiency of academic research but also promotes the intelligent development of knowledge discovery and academic communication.

[0003] Existing RAG (Research and Analysis) technologies face significant challenges in processing academic question answering. Credible academic question answering requires not only accurate responses but also precise evidence attribution to ensure interpretability and build user trust. Traditional keyword-based or vector-based RAG technologies are susceptible to noise interference, which can crowd out or dilute genuine evidence, leading to "contextual contamination" and reducing attribution accuracy. Furthermore, they fail to adaptively focus on high-value areas, ignoring a crucial characteristic of academic evidence distribution: spatial concentration. Specifically, academic writing conventions dictate that content with a specific intent often concentrates in localized functional areas; the inability to adaptively focus on these areas severely reduces the accuracy of evidence attribution and overall credibility of academic question answering. Summary of the Invention

[0004] This invention addresses the technical problems of existing RAG technology in academic question-answering scenarios, such as contextual pollution and low accuracy of evidence attribution. Through multi-dimensional mechanism innovation and process optimization, it provides a credible academic question-answering method and device based on FocusedRAG, which achieves precise compression of the retrieval space, efficient extraction of high-value evidence, and establishes a traceable evidence attribution basis, significantly improving the accuracy and credibility of academic question-answering.

[0005] This invention provides a trusted academic question-answering method based on FocusedRAG, comprising the following steps:

[0006] S1. Data Acquisition: Obtain academic paper data from publicly available data sources as the raw dataset;

[0007] S2. Preprocessing: The document content in the original dataset is segmented along the chapter boundaries, and the user query is rewritten using a large language model to obtain the rewritten query.

[0008] S3. Search Space Compression: Construct a query adaptive signal-to-noise separation mechanism, and select chapters related to the query as the search space based on the rewritten query;

[0009] S4. Evidence Interval Filtering: Within the retrieval space, an extended query adaptive signal-to-noise separation mechanism is used to obtain sentence-level signal-to-noise boundary thresholds. Candidate evidence intervals are obtained from the retrieval space through a sentence selection mechanism guided by relevance density. The candidate evidence intervals are sorted based on their cumulative relevance scores. Under the constraint of retrieval token budget, the candidate evidence intervals are selected optimally to form the retrieval context.

[0010] S5. Answer Generation: Input the retrieval context and user query into the large language model, which then generates academic question-and-answer answers based on the initial question content.

[0011] Furthermore, the preprocessing module specifically includes:

[0012] Extract chapter titles from the content of each document in the original dataset. Based on these detected chapter titles, divide the corresponding document content into chapter units. The content in each chapter is further divided into independent sentences.

[0013] The user query and all detected chapter titles are input into the large language model as context information, and the query is rephrased to generate a rewritten query.

[0014] Furthermore, the query adaptive signal-to-noise separation mechanism specifically includes:

[0015] The rewritten query and the target object are input into the pre-trained intent recognition model to obtain the intent relevance score between the target object and the rewritten query. The intent relevance scores of all target objects and the rewritten query are aggregated to obtain the intent relevance distribution.

[0016] The intent relevance distribution is sorted in descending order. Then, based on each candidate boundary position, the target object is divided into signal class and noise class in turn, and the inter-class variance of each division is calculated.

[0017] Identify the optimal boundary location by finding the maximum inter-class variance. Based on the optimal boundary position and The average intent relevance score is calculated as the optimal threshold for dividing the intent relevance score of the target object.

[0018] Chapters with intent relevance scores higher than the optimal threshold are designated as signal chapters for subsequent sentence selection.

[0019] Furthermore, the extended query adaptive signal-to-noise separation mechanism specifically includes:

[0020] The rewritten query and all sentences in the signal chapter are input into the query adaptive signal-to-noise separation mechanism. A pre-trained intent recognition model is used to calculate the intent relevance score between the rewritten query and each sentence, and the sentence-level intent relevance distribution is obtained.

[0021] Gaussian smoothing is applied to the sentence-level intent relevance distribution to obtain a smoothed intent relevance distribution.

[0022] By reusing operations S32, S33, and S34 on the smoothed intent relevance distribution, the sentence-level optimal boundary threshold is obtained.

[0023] Furthermore, the sentence selection mechanism guided by relevance density specifically includes:

[0024] Calculate the intention-related surplus:

[0025] For the smoothed intent relevance score of a sentence in the signal chapter, calculate its intent relevance surplus relative to the sentence-level optimal threshold.

[0026] Construct a set of candidate evidence intervals:

[0027] For each starting position that satisfies the intention relevance surplus being greater than zero, the ending position that maximizes the cumulative surplus is identified using the maximum subarray principle. All intervals with positive cumulative surplus are collected to form a candidate interval set. Topological filtering is performed on each signal chapter separately, and the filtering results of all signal chapters are aggregated to form a candidate evidence interval set.

[0028] Building the search context:

[0029] The candidate evidence intervals are sorted in descending order of cumulative relevance to obtain a priority sequence. Then, a greedy selection is performed on the priority sequence to obtain the final retrieval context.

[0030] Furthermore, greedy choices specifically include:

[0031] Retrieving token budget Under the constraints, perform a greedy selection on the priority sequence:

[0032]

[0033] In the formula, The optimal boundary position is... Indicates priority sequence The number of tokens;

[0034] Before The content of each interval is aggregated into a selected search context. The intervals in the selected search context are then reordered according to their original document positions to form the final search context.

[0035] Furthermore, the answer generation module specifically includes:

[0036] Receive the final retrieval context output in step S4 and user query As input, the final answer is generated using a large language model.

[0037] This invention also provides a trusted academic question-answering device based on FocusedRAG, comprising:

[0038] Data acquisition module: Acquires academic paper datasets containing fine-grained evidence annotations from public data sources;

[0039] Preprocessing module: Extracts text content, detects the lines containing chapter titles and divides the document along chapter boundaries, further segments the content of each chapter by sentence, and uses a large language model combined with all chapter titles to rewrite the user query to clarify the search intent;

[0040] Search space compression module: Design an adaptive signal-to-noise separation mechanism to divide the entire document content into signal chapters that are relevant to the query and noise chapters that are irrelevant to the query, thereby compressing the search scope from the entire document space to the chapter space relevant to the query;

[0041] Evidence interval filtering module: Based on the selected signal chapter content, the query adaptive signal-to-noise separation mechanism is first extended to the sentence level to establish the sentence-level signal-to-noise separation threshold; then, a relevance density-guided sentence selection mechanism is designed to identify continuously occurring high-relevance sentence regions as candidate evidence intervals, and the cumulative relevance scores of each candidate evidence interval are sorted. Finally, under the constraint of retrieval token budget, the sorted candidate evidence intervals are greedily selected to form the final retrieval context.

[0042] Answer generation module: Receives the final retrieval context and user query as input, uses a large language model to generate the final answer and output it.

[0043] The present invention also provides a computer-readable storage medium storing a computer program that can be executed by a processor to implement the FocusedRAG-based trusted academic question-answering method as described above.

[0044] Beneficial effects of this invention:

[0045] This invention preserves the original natural text structure of academic papers and clarifies the retrieval intent through document segmentation and query rewriting, laying the foundation for accurate attribution of credible academic questions and answers. Through the query adaptive signal-to-noise separation mechanism, the chapter-level and sentence-level screening criteria are dynamically determined according to the query characteristics, compressing the search space to a region highly relevant to the query at the retrieval source, fundamentally solving the problem of context pollution and significantly improving attribution accuracy.

[0046] By using a sentence selection mechanism guided by relevance density, we prioritize the identification and extraction of consecutively occurring highly relevant sentence regions, effectively suppressing the misleading influence of isolated high-scoring interference fragments, and further improving attribution accuracy and evidence validity.

[0047] Furthermore, by combining the query adaptive signal-to-noise separation mechanism with the relevance density-guided sentence selection mechanism, the limited retrieval token budget can be prioritized for allocation to real evidence, thereby establishing an accurate attribution basis and significantly improving the accuracy and credibility of academic question answering. Attached Figure Description

[0048] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0049] Figure 1 This is a schematic diagram of the data preprocessing workflow;

[0050] Figure 2 To query the flowchart of the adaptive signal-to-noise separation mechanism;

[0051] Figure 3 A schematic diagram of the sentence selection mechanism guided by relevance density. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] Example 1

[0054] In this embodiment, the present invention provides a trustworthy academic question-answering method based on FocusedRAG, the specific steps of which include:

[0055] S1. Data Acquisition:

[0056] In this embodiment, the QASPER dataset with complete evidence attribution annotations is selected. These datasets not only contain full-text PDF files of papers and question-answer pairs, but more importantly, they provide specific original text fragments for answering questions. Such annotated data containing precise evidence sources is the foundation for building and validating a credible academic question-answering system, and can support subsequent accurate attribution evaluation of the retrieved context.

[0057] S2. Preprocessing

[0058] The dataset is preprocessed with document content and user queries, including two parallel operations: document segmentation and query rewriting.

[0059] S21. Document Segmentation

[0060] First, the MinerU PDF parser is applied to extract text content from the PDF document. Then, lines starting with the hash symbol (#) are detected as chapter titles to identify the document's natural boundaries. Based on these detected chapter titles, the document is divided into chapter units, and the content within each chapter is further segmented into independent sentences.

[0061] S22. Query Rewrite

[0062] User queries The chapter titles detected in step S201 are input into the large language model as contextual information. The large language model is prompted to rephrase the query to make the retrieval intent explicit while maintaining semantic equivalence, thereby generating the rewritten query. .

[0063] S3. Search space compression

[0064] An adaptive signal-to-noise separation mechanism is applied at the chapter level to compress the retrieval space, reducing the entire document-wide retrieval space to query-related signal chapters. Specifically, this step first concatenates each chapter title with its content to form a chapter representation. Then, the rewritten query and the chapter representations of all chapters are input into the adaptive signal-to-noise separation mechanism to obtain the optimal threshold. Finally, chapters with intent relevance scores higher than this optimal threshold are designated as signal chapters for subsequent sentence selection.

[0065] Specifically, the implementation process of this query adaptive signal-to-noise separation mechanism is as follows:

[0066] S31. Calculate the intention relevance score

[0067] Rewrite query With the target object (All chapters indicate) The input is fed into the pre-trained intent recognition model Qwen3-Reranker to obtain the target object. Rewrite query Intent relevance score The intent relevance distribution is obtained by aggregating the intent relevance scores of all target objects and the rewritten query. ,in For the number of target objects, This represents the intent relevance score of the nth target object;

[0068] S32. Calculate the optimal boundary threshold.

[0069] First, the intention-related distribution is... Sort in descending order to obtain ,in Indicates the first Larger scores; then based on each candidate boundary position The target objects are divided into signal classes in sequence. and noise And calculate the inter-class variance for each partition:

[0070]

[0071] in, Represents the variance between classes. and These represent the proportions of objects assigned to the signal class and the noise class, respectively. Indicates the number of target objects. Indicates the candidate boundary position. , and These represent the average intent relevance scores for the signal class and the noise class, respectively. This indicates the degree of separation between the two classes in the score space, while This indicates that the weighting is based on the number of objects contained in each class. Indicates multiplication;

[0072] After calculating the inter-class variance for all candidate boundary locations, the optimal boundary location is identified by finding the one with the largest inter-class variance. :

[0073]

[0074] In the formula, Indicates the optimal boundary position. This represents a function that finds the location corresponding to the maximum inter-class variance. Represents the variance between classes;

[0075] Based on the optimal boundary position and The intention relevance score is calculated as the average intention relevance score:

[0076] in, The average intent relevance score serves as the optimal threshold for dividing the target object's intent relevance score. Indicates the optimal boundary position The intention relevance score at the location.

[0077] S4. Evidence Interval Screening

[0078] Within the selected signal section, the query adaptive signal-to-noise separation mechanism is first extended to adapt to sentence-level signal-to-noise separation, and then the relevance density-guided sentence selection mechanism is applied to construct the retrieval context.

[0079] Specifically, the rewritten query and all sentences in the signal section are first input into the query adaptive signal-to-noise separation mechanism. The pre-trained intent recognition model Qwen3-Reranker is then used to calculate the intent relevance score between the rewritten query and each sentence, thereby obtaining the sentence-level intent relevance distribution. Subsequently Perform Gaussian smoothing, that is, for sentences in the signal section... Intent relevance score , application kernel radius Gaussian smoothing kernel is used to calculate the smoothed score:

[0080]

[0081] In the formula, The Gaussian kernel weights are set to [0.25, 0.5, 0.25], satisfying... , This represents the smoothed fraction. Indicates the relevance score of the intent;

[0082] After processing all sentences, a smoothed intent relevance distribution is obtained. Subsequently, the query-adaptive signal-to-noise separation mechanism based on... Calculate the sentence-level optimal threshold ;

[0083] Then, the sentence-level optimal threshold is used. The sentence selection mechanism, guided by the relevance density, inputs a smoothed intent relevance score of all sentences in the signal section to construct the retrieval context. .

[0084] The implementation process of the relevance density-guided sentence selection mechanism is as follows:

[0085] Calculate the intention-related surplus

[0086] Regarding the sentences in the signal chapter Smoothed Intent Relevance Score Calculate its relative to the sentence-level optimal boundary threshold. Intentional relevance surplus :

[0087]

[0088] In the formula, Sentence Smoothed intent relevance score This represents the optimal threshold for sentence-level delimitation. Indicates a surplus in relevance to intent;

[0089] Construct a set of candidate evidence intervals

[0090] For each satisfying starting position The termination position that maximizes the accumulated surplus is identified using the maximum subarray principle:

[0091]

[0092] In the formula, Indicates the maximum termination position. Indicates as of the date The global maximum cumulative surplus at each position;

[0093] Then, all intervals with positive cumulative surplus are collected to form a candidate interval set:

[0094]

[0095] In the formula, Represents the set of candidate intervals. This represents the cumulative surplus at the maximum termination position. Indicates a surplus in relevance to intent;

[0096] Then, for each signal chapter Perform topological filtering separately: If there is at least one multi-sentence interval with a positive cumulative surplus within the chapter, then only the multi-sentence interval is retained. Otherwise, retain all single-sentence intervals. :

[0097]

[0098] In the formula, Indicates a range of multiple sentences. Indicates a single sentence interval. This indicates the filtering results for chapter numbers;

[0099] Next, the filtering results from all signal sections are aggregated to form a set of candidate evidence intervals:

[0100]

[0101] in, This indicates all signal chapters. This represents the set of candidate evidence intervals.

[0102] Building a search context

[0103] According to the set of candidate evidence intervals Cumulative correlation of each interval Arrange in descending order to obtain the priority sequence. ;

[0104] Then retrieve the token budget Under the constraints, a greedy selection is performed on the above priority sequence:

[0105]

[0106] in Representing an interval The number of tokens, This indicates the retrieval token budget, with a value of 500;

[0107] The greedy selection process has two termination conditions: either the token retrieval budget is exhausted or the priority sequence is overwritten; eventually, the previous... The content of each interval is aggregated into the selected search context:

[0108]

[0109] In the formula, Indicates the first One candidate evidence interval;

[0110] Select search context The intervals in the search results are reordered according to their original document positions to form the final search context. .

[0111] S5, Answer Generation

[0112] This step receives the retrieval context output from step S4. and user query As input, the final answer is generated using a large language model.

[0113] Example 2

[0114] To better demonstrate the practical technical effectiveness of this invention in academic question-answering tasks, it was applied to a real-world academic question-answering scenario to verify the accuracy of evidence attribution and the precision of answers. Table 1 compares the evidence retrieval quality and generated answer quality of the FocusedRAG method in this invention with several existing RAG baseline methods on three academic question-answering datasets (QASPER, QASA, and M3SciQA). The evaluation metrics include precision, recall, F1 score, and LLM-based rating. Precision represents the proportion of genuine evidence in the retrieved content, recall represents the proportion of genuine evidence retrieved, F1 score is the harmonic mean of precision and recall, and LLM-based rating is used to evaluate the semantic quality of answers generated based on a large language model.

[0115] As shown in Table 1, FocusedRAG achieved state-of-the-art performance across all datasets. Regarding evidence retrieval quality, FocusedRAG improved precision by 4.9%, 10.9%, and 40.0% on QASPER, QASA, and M3SciQA, respectively, compared to the strongest baseline method. This demonstrates that the query-adaptive signal-to-noise separation mechanism effectively improves the accuracy of evidence attribution by physically isolating interfering segments. In terms of recall, FocusedRAG improved recall by 12.3% and 47.7% on QASA and M3SciQA, respectively, proving that the relevance density-guided sentence selection mechanism achieves more comprehensive evidence coverage with a limited retrieval token budget by constructing continuous evidence intervals. In terms of the quality of generated answers, FocusedRAG improved the F1 score by 9.4%, 13.0%, and 16.6% respectively compared to the strongest baseline, and improved the LLM-based semantic evaluation score by 10.1%, 16.2%, and 15.2% respectively. This demonstrates that FocusedRAG provides accurate evidence support for large language models by constructing a clean context at the retrieval source through two innovative mechanisms, thereby significantly improving the accuracy and credibility of the answers.

[0116] Table 1. Comparison of Evidence Retrieval Quality and Response Generation Quality between FocusedRAG Method and Various RAG Methods

[0117]

[0118] Example 3

[0119] A trusted academic question-answering device based on FocusedRAG specifically includes:

[0120] Data acquisition module: Acquires academic paper datasets containing fine-grained evidence annotations from public data sources;

[0121] Preprocessing module: Uses a PDF parser to extract text content, then detects the lines containing chapter titles and divides the document along chapter boundaries, then further segments the content of each chapter by sentence, and uses a large language model to rewrite the user query in combination with all chapter titles to clarify the search intent, and finally sends the segmented document and the rewritten query to the chapter location module.

[0122] Search space compression module: Design an adaptive signal-to-noise separation mechanism to divide the entire document content into signal chapters that are relevant to the query and noise chapters that are irrelevant to the query, compress the search scope from the entire document space to the chapter space that is relevant to the query, and send the content of the signal chapters to the sentence selection module;

[0123] Evidence interval selection: Based on the selected signal chapter content, the query adaptive signal-to-noise separation mechanism is first extended to the sentence level to establish the sentence-level signal-to-noise separation threshold; then, a relevance density-guided sentence selection mechanism is designed to identify continuously occurring high-relevance sentence regions as candidate evidence intervals, and the cumulative relevance scores of each candidate evidence interval are sorted. Finally, under the constraint of retrieval token budget, the sorted candidate evidence intervals are greedily selected to form the retrieval context.

[0124] Answer generation module: Receives the retrieval context and user query as input, uses a large language model to generate the final answer and outputs it.

[0125] Example 4

[0126] A computer-readable storage medium storing a computer program, wherein a processor reads and executes a FocusedRAG-based trusted academic question-answering method stored in the storage medium.

[0127] The processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0128] Memory is used to store computer programs or modules. The processor implements various functions of electronic devices by running or executing computer programs and / or modules stored in memory, and by accessing data stored in memory. Memory can mainly include a program storage area and a data storage area. The program storage area can store the operating system, at least one application program required for a function, etc.; the data storage area can store data created based on terminal usage, etc. In addition, memory can also include high-speed random access memory, non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart memory cards (SMC), secure digital storage (SD) cards, flash memory cards, at least one disk storage area, flash memory devices, or other volatile solid-state storage devices to ensure data storage stability and read / write efficiency.

[0129] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A trustworthy academic question-answering method based on FocusedRAG, characterized in that, Includes the following steps: S1. Data Acquisition: Obtain academic paper data from publicly available data sources as the raw dataset; S2. Preprocessing: The document content in the original dataset is segmented along the chapter boundaries, and the user query is rewritten using a large language model to obtain the rewritten query. S3. Search Space Compression: Construct a query adaptive signal-to-noise separation mechanism, based on the rewritten query, selecting chapters relevant to the query as the search space. The specific steps are as follows: S31. Input the rewritten query and the target object into the pre-trained intent recognition model to obtain the intent relevance score between the target object and the rewritten query. Aggregate the intent relevance scores between all target objects and the rewritten query to obtain the intent relevance distribution. S32. Sort the intention relevance distribution in descending order, and then, based on each candidate boundary position, classify the target object into signal class and noise class in turn, and calculate the inter-class variance of each classification. S33. Identify the optimal boundary location by finding the maximum inter-class variance. Based on the optimal boundary position and The average intent relevance score is calculated as the optimal threshold for dividing the intent relevance score of the target object. S34. Use chapters with intent relevance scores higher than the optimal threshold as signal chapters for subsequent sentence selection; S4. Evidence Interval Filtering: Within the retrieval space, an extended query adaptive signal-to-noise separation mechanism is used to obtain sentence-level signal-to-noise boundary thresholds. Candidate evidence intervals are obtained from the retrieval space through a sentence selection mechanism guided by relevance density. The candidate evidence intervals are sorted based on their cumulative relevance scores. Under the constraint of retrieval token budget, the candidate evidence intervals are selected optimally to form the retrieval context. The extended query adaptive signal-to-noise separation mechanism specifically includes: S411. Input the rewritten query and all sentences in the signal chapter into the query adaptive signal-to-noise separation mechanism, use the pre-trained intent recognition model to calculate the intent relevance score between the rewritten query and each sentence, and obtain the sentence-level intent relevance distribution; S412. Gaussian smoothing is applied to the sentence-level intent relevance distribution to obtain the smoothed intent relevance distribution; S413. Reuse the operations of S32, S33 and S34 on the smoothed intention relevance distribution to obtain the sentence-level optimal boundary threshold; The sentence selection mechanism guided by relevance density specifically includes: S421. Calculate the intention-related surplus: For the smoothed intent relevance score of a sentence in the signal chapter, calculate its intent relevance surplus relative to the sentence-level optimal threshold. S422. Construct a set of candidate evidence intervals: For each starting position that satisfies the intention relevance surplus being greater than zero, the ending position that maximizes the cumulative surplus is identified using the maximum subarray principle. All intervals with positive cumulative surplus are collected to form a candidate interval set. Topological filtering is performed on each signal chapter separately, and the filtering results of all signal chapters are aggregated to form a candidate evidence interval set. S423. Constructing the retrieval context: The candidate evidence intervals are sorted in descending order of cumulative relevance to obtain a priority sequence. Then, a greedy selection is performed on the priority sequence to obtain the final retrieval context. S5. Answer Generation: Input the retrieval context and user query into the large language model, and the large language model generates academic question-and-answer answers for the user query.

2. The trusted academic question-answering method based on FocusedRAG according to claim 1, characterized in that, The preprocessing step S2 specifically includes: S21. Extract chapter titles from the content of each document in the original dataset. Based on these detected chapter titles, divide the corresponding document content into chapter units. The content in each chapter is further divided into independent sentences. S22. Input the user query and all detected chapter titles as context information into the large language model, reformulate the query to generate the rewritten query.

3. The trusted academic question-answering method based on FocusedRAG according to claim 1, characterized in that, The greedy choice specifically includes: Retrieving token budget Under the constraints, perform a greedy selection on the priority sequence: In the formula, The optimal boundary position is... Indicates priority sequence The number of tokens; Before The content of each interval is aggregated into a selected search context. The intervals in the selected search context are then reordered according to their original document positions to form the final search context.

4. The trusted academic question-answering method based on FocusedRAG according to claim 1, characterized in that, The answer generation process in step S5 specifically includes: The final retrieval context and user query output from step S4 are received as input, and the final answer is generated using a large language model.

5. A trusted academic question-answering device based on FocusedRAG, executing a trusted academic question-answering method based on FocusedRAG as described in any one of claims 1-4, characterized in that, include: Data acquisition module: Acquires academic paper datasets containing fine-grained evidence annotations from public data sources; Preprocessing module: Extracts text content, detects the lines containing chapter titles and divides the document along chapter boundaries, further segments the content of each chapter by sentence, and uses a large language model combined with all chapter titles to rewrite the user query to clarify the search intent; Search space compression module: Design an adaptive signal-to-noise separation mechanism to divide the entire document content into signal chapters that are relevant to the query and noise chapters that are irrelevant to the query, thereby compressing the search scope from the entire document space to the chapter space relevant to the query; Evidence interval filtering module: Based on the selected signal chapter content, the query adaptive signal-to-noise separation mechanism is first extended to the sentence level to establish the sentence-level signal-to-noise separation threshold; then, a relevance density-guided sentence selection mechanism is designed to identify continuously occurring high-relevance sentence regions as candidate evidence intervals, and the cumulative relevance scores of each candidate evidence interval are sorted. Finally, under the constraint of retrieval token budget, the sorted candidate evidence intervals are greedily selected to form the final retrieval context. Answer generation module: Receives the final retrieval context and user query as input, uses a large language model to generate the final answer and output it.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that can be executed by a processor to implement a trusted academic question-answering method based on FocusedRAG as described in any one of claims 1-4.

Images