Sci-tech literature question-answering method and system based on self-feedback iterative optimization

Abstract

The application provides a scientific and technological literature question-answering method and system based on self-feedback iterative optimization, which comprises the following steps: receiving a user question, generating an initial response by using a retrieval enhancement generation strategy; taking the initial response as an iterative input, performing multidimensional evaluation on the iterative input to obtain a structured feedback with multiple specific feedbacks; extracting information requirements from the structured feedback, performing supplementary retrieval according to the information requirements to obtain supplementary information; performing progressive information fusion on the supplementary information and the iterative input according to the priority of the specific feedbacks to obtain optimized content; verifying and correcting the references in the optimized content to obtain an iterative output; judging whether the iterative output meets an iterative termination condition, if yes, taking the iterative output as a final response, otherwise, inputting the iterative output as an iterative input to perform iterative optimization until a final response is obtained. The application can greatly improve the answering quality of academic answers.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, and in particular to a scientific literature question-answering method and system based on self-feedback iterative optimization. Background Technology

[0002] In applications such as scientific research, academic surveys, and professional consultation, users need systems capable of synthesizing multiple scientific documents to answer complex academic questions. With the development of Large Language Models (LLMs), applying LLMs to scientific document question answering systems has become a trend.

[0003] Large language models, as a significant breakthrough in the field of artificial intelligence, have demonstrated powerful capabilities in tasks such as text generation, question answering systems, and knowledge reasoning. However, with the continuous expansion of application scenarios, especially in demanding fields such as scientific research and professional consulting, the limitations of traditional large language models are becoming increasingly apparent. Traditional large language models primarily rely on parameterized knowledge learned during the pre-training phase. When faced with queries requiring up-to-date information or iterative input knowledge, they are prone to problems such as outdated knowledge, inaccurate information, or the inability to provide verifiable references. Therefore, numerous improvements have been made to traditional large language models, resulting in improved large language models, including:

[0004] (1) Retrieval Enhancement Generation System

[0005] Typical RAG (Retrieval-Augmented Generation) systems, such as the scheme proposed by Lewis et al. in 2020 (Lewis et al., "Retrieval-augmented generation for knowledge-intensive NLP tasks", NeurIPS 2020), adopt a two-stage "retrieval-generation" architecture. The system first retrieves relevant documents from an external knowledge base based on the user query, and then inputs the retrieval results together with the original query into the generative model to produce the final answer.

[0006] The limitations of this scheme are: the retrieval process is static, leading to missing information; the quality of generated content is unstable; and it cannot handle inconsistencies between the retrieved information and the generated content.

[0007] (2) Self-improvement methods based on reinforcement learning

[0008] The Self-Refine method proposed by Madaan et al. (Madaan et al., "Self-refine: Iterative refinement with self-feedback", arXiv 2023) allows the model to self-evaluate the generated content and propose improvement suggestions, then regenerates the content based on this feedback. This method demonstrates improved performance on multiple text generation tasks.

[0009] The shortcomings of this scheme are: information retrieval is prone to be missing; retrieval quality is unstable; and the accuracy and verifiability of academic citations are insufficient.

[0010] (3) Multi-agent cooperative optimization method

[0011] The multi-agent debate framework proposed by Du et al. ("Improving factuality and reasoning in language models through multiagent debate", arXiv 2023) improves the quality of answers through debates between multiple agents. This method involves different agents generating initial answers, and then reaching a consensus through multiple rounds of debate.

[0012] The limitations of this approach include: the need to run multiple large model instances, resulting in high computational costs; unstable retrieval quality; and limited effectiveness when handling tasks requiring extensive expertise.

[0013] (4) Iterative retrieval generation method

[0014] The Active Retrieval Augmented Generation proposed by Jiang et al. (EMNLP 2023) attempts to perform multiple retrievals during the generation process. This method determines when to perform additional retrievals based on preset triggering conditions, and then continues the generation process.

[0015] The problem with this approach is that the conditions for triggering a search are relatively fixed, and the judgment of the search timing depends on preset rules, resulting in a lack of flexibility and accuracy.

[0016] (5) Iterative Input Question Answering System

[0017] The PaperQA2 system developed by Skarlinski et al. (Skarlinski et al., "Language agents achieve superhuman synthesis of scientific knowledge", arXiv 2024) is specifically designed for scientific literature question answering tasks and employs a hierarchical document processing and citation generation strategy.

[0018] The limitations of this system are: it relies heavily on preprocessed document sets, which can easily lead to missing information; and its consistency in handling complex multi-document problems needs to be improved.

[0019] The limitations of these existing technologies are specifically manifested in the following ways:

[0020] 1. The overall quality of information in the scientific literature question-and-answer system is unstable.

[0021] In existing technologies, traditional LLMs employ a single-generation strategy, while RAG systems use a two-stage "retrieval-generation" processing model. Self-Refine methods, however, have relatively simple feedback mechanisms and cannot guarantee the stability of retrieval quality. This results in significant deficiencies in the academic answers generated by the systems in terms of information completeness, logical argumentation, and citation accuracy, failing to meet the needs of complex academic queries. For example, they may omit key research directions or lack a clear structure.

[0022] 2. Information loss caused by static retrieval strategies in scientific literature retrieval

[0023] Traditional RAG systems perform a one-time retrieval based on an initial query, failing to dynamically adjust to knowledge gaps during the generation process. For example, the ActiveRAG iterative retrieval method has fixed triggering conditions, lacking flexibility. This is particularly evident in interdisciplinary research and reviews of emerging technologies, leading to incomplete information coverage and impacting the depth of question answering. For instance, when searching for "ethical challenges of artificial intelligence in biomedicine," ethical literature may be easily overlooked.

[0024] 3. Insufficient accuracy and verifiability of academic citations

[0025] Traditional LLMs often suffer from issues such as fabricated citations, inconsistent content, and non-standard formatting when generating content. For example, PaperQA2's iterative input system lacks a deep citation verification mechanism, and existing self-improvement methods are not optimized for citation accuracy. This directly affects academic integrity and credibility, with core challenges including "identifying statements that need to be cited, matching authentic literature paragraphs, verifying content consistency, and handling multi-source attribution," thus limiting the system's application in academic writing, research consulting, and other scenarios.

[0026] 4. Consistency and logical issues in the fusion of information from multiple documents

[0027] Existing multi-agent collaborative methods struggle to effectively integrate differing viewpoints and methodological conflicts from multiple sources; iterative input systems lack consistency when processing complex multi-document synthesis. This results in the system's inability to reasonably weigh contradictory conclusions when integrating literature from different periods and teams, a lack of a unified analytical perspective, and the potential for logical conflicts or redundancy between new and existing information, making it difficult to form a coherent and rigorous comprehensive answer.

[0028] Therefore, the scientific literature question-answering system based on the large language model still needs further optimization. Summary of the Invention

[0029] To address the aforementioned problems in existing technologies, this invention provides a scientific literature question-answering method and system based on self-feedback iterative optimization, thereby improving the quality of academic answers from multiple perspectives.

[0030] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0031] In a first aspect, the present invention provides a scientific literature question-answering method based on self-feedback iterative optimization, comprising:

[0032] Step S1: Receive user questions and generate an initial response using a search-enhanced generation strategy;

[0033] Step S2: Using the initial response as the iterative input, perform multi-dimensional evaluation on the iterative input to obtain structured feedback with multiple specific feedbacks;

[0034] Step S3: Extract information requirements from the structured feedback, and perform supplementary retrieval based on the information requirements to obtain supplementary information;

[0035] Step S4: Based on the priority of the specific feedback, progressively fuse the supplementary information and the iterative input to obtain optimized content;

[0036] Step S5: Verify and correct the references in the optimized content to obtain the iterative output;

[0037] Step S6: Determine whether the iteration termination condition is met based on the iteration output. If yes, use the iteration output as the final response; otherwise, input the iteration output into step S2 as the iteration input for iterative optimization until the final response is obtained.

[0038] The beneficial effects of this invention are as follows: By generating structured feedback through multi-dimensional evaluation of iterative input, dynamic supplementary retrieval is performed based on the information needs in the structured feedback. Finally, the results of the dynamic supplementary retrieval and iterative input are progressively integrated, and the cited content is verified and corrected. This achieves continuous self-improvement of academic answers, thereby significantly improving the quality, stability, and professional level of academic Q&A, significantly improving the comprehensiveness and precision of literature retrieval, ensuring the logical consistency and academic rigor of the content in each information integration process, and significantly improving the credibility and practicality of the system in scenarios such as academic writing assistance, research report generation, and professional consultation, thus greatly improving the quality of academic answers.

[0039] Optionally, step S2 includes:

[0040] Using the initial response as iterative input, the iterative input is subjected to content organization quality assessment, information coverage quantification assessment, and citation accuracy assessment to obtain structured feedback with multiple specific feedbacks;

[0041] Among them, the content organization quality assessment function E org (y t ,q) is:

[0042]

[0043] In the formula, y t Let q represent the user question, α, β, and γ represent the corresponding weights, and Cohe(y) represent the iterative input. t Stru(y) represents the coherence scoring function. t Rele(y) represents the structural quality scoring function. t ) represents the relevance scoring function, n represents the total number of sentences, sim(s) i ,s i+1 ) represents the semantic similarity between the i-th sentence and the (i+1)-th sentence in the iterative input, N valid N represents the number of valid paragraphs. total Indicates the total number of paragraphs. This represents the average paragraph integrity score;

[0044] Among them, the information coverage metric evaluation function E cov (y t ,q) is:

[0045]

[0046] In the formula, Topics(q) represents the total topic of user question q, and topic1, topic2, and topic3 ​​are also included. m This represents the 1st, 2nd, and mth topics after decomposition.i This represents the i-th topic. Indicates whether the i-th topic was input in the t-th iteration. cover, Indicates an indicator function, This indicates the depth of discussion surrounding a topic in the responses. Indicates the topic Number of related sentences Indicate topic The expected length of the discussion;

[0047] Among them, the accuracy evaluation function E cite (y t C t ,P t )for:

[0048]

[0049] In the formula, C t P represents a collection of references. t Let c represent the set of documents retrieved in the t-th iteration. i Represents the i-th reference, Exis(c i ) represents the existence verification function, Cons(c i ,source(c i ) represents the i-th reference c i With source document source(c i Content consistency score, c i .source indicates a reference to c i The source document, c i .position indicates a reference to c i In the source document, valid_positions represents the range of valid citation positions. Cons(c i ,p j ) represents the i-th reference c i With the j-th retrieved document p j The consistency verification function, BLEU stands for bilingual evaluation substitution, extract(p j ,c i .position) indicates retrieving document p j c i The original text extracted at the .position position, c i .content indicates a reference to c i The original text quoted in the document;

[0050] The structured feedback F t The generating function of the i-th specific feedback for:

[0051] ;

[0052] In the formula, E represents the feedback generating function. i (y t ) represents the i-th evaluation dimension relative to the iterative input y. t Quality rating This represents the quality threshold for evaluation dimension i;

[0053] Each feedback It consists of the following three components:

[0054]

[0055] In the formula, issue i This represents the problem identified in the i-th specific feedback. This represents the problem identification function, suggestion. i This refers to the issue. i directions for improvement This indicates an improved recognition function, info_need i Indicates directions for improvement. i The specific information required is specified by extr_requ, which represents the information request function.

[0056] Optionally, step S4 includes:

[0057] Prioritize the information based on its importance, urgency, and feasibility, and then merge the supplementary information and iterative input in stages, including structural adjustment, content supplementation, and reference correction, to obtain optimized content.

[0058] Among them, priority sorting function for:

[0059]

[0060] In the formula, α, β, and γ represent the corresponding weights, This represents the importance evaluation function. Indicates the urgency assessment function, This represents the feasibility assessment function. Indicates the current quality loss. Indicates the maximum possible mass loss. E represents the quality threshold for evaluation dimension i. i (y t ) represents the i-th evaluation dimension relative to the iterative input y.t Quality rating Indicates the number of relevant search documents. Indicates the minimum number of documents required;

[0061] Among them, optimized content y t+1 The fusion generation function is:

[0062]

[0063] In the formula, Indicates a reference to the fusion function, y content F represents the intermediate response after content supplementation. citation This indicates feedback from the reference class. This indicates supplementary information including a supplementary document collection. This represents the content fusion function, y struct F represents the intermediate response after structural supplementation. coverage This indicates content-based feedback. F represents the structure fusion function. struct This indicates a structured feedback. Let y represent the probability distribution function. new This represents the new text generated during the current fusion phase, y current f represents the pre-processed text generated during the current fusion phase. i P represents the feedback at the current integration stage. relvant P(y) represents a subset of supplementary documents related to the current fusion phase. new |y current ,f i ,P relevant ) indicates the preceding text y in the current fusion phase. current Feedback f i and supplementary document subset P relevant Generate new text y under common constraints new The probability, w j Represents the new text y new The j-th word in; w <j P(w) represents all words preceding the j-th word in the new text. j |w <j y current ,f i ,P relevant ) indicates all the preceding words w <j And the pre-text y of the current fusion phase current Feedback f i and supplementary document subset P relevant Generate j words w under the common constraints j The probability of.

[0064] Optionally, step S3 includes:

[0065] Information requirements are extracted from the structured feedback, and the structured feedback is supplemented and retrieved using direct expansion strategy, semantic rewriting strategy and context-aware strategy to obtain the final query set;

[0066] Among them, the information demand extraction function InfoNeed i for:

[0067]

[0068] In the formula, This represents the i-th specific feedback. Information demand conversion function, type i Keywords indicate the type of information requirement. i This refers to key information extracted from specific feedback, constraints i The constraints represent the information demand, argmax represents the set of parameters that take the maximum value, and K represents the set of selected keywords. This represents the i-th specific feedback. All entities, This indicates that keyword k is in the i-th specific feedback. TF-IDF weights in;

[0069] The function for the direct extension strategy is:

[0070] ;

[0071] In the formula, For the i-th information request, the direct extended query is generated, and extract_nouns(q) is a function that extracts phrases from the user question q;

[0072] The function for the semantic rewriting strategy is:

[0073] ;

[0074] In the formula, Let R represent the semantic rewrite query generated by the i-th information request, and let R represent the query rewriter based on the pre-trained language model.

[0075] The function for the context-aware strategy is:

[0076] ;

[0077] In the formula, This represents the context-aware query generated for the i-th information request. The context_from(y) function represents the fusion function. t) represents the iterative input y t The contextual information mentioned in the text;

[0078] The final query set is generated through a weighted fusion of direct expansion, semantic rewriting, and context-aware strategies, and its function is as follows:

[0079] ;

[0080] In the formula, This represents the final query set for the i-th information request. For the union operation, k t It represents the total number of specific feedbacks in the t-th iteration.

[0081] Optionally, step S3 further includes:

[0082] For the final query set, supplementary information is obtained by filtering through multiple rounds of parallel retrieval, document deduplication, and multi-dimensional relevance filtering.

[0083] Among them, supplementary information was obtained. The function is:

[0084]

[0085] In the formula, This represents the collection of documents after deduplication. This indicates that document p belongs to the document collection. , The correlation threshold, This is a deduplication function. Represents the original document collection. IsDuplicate(p) is the repetition threshold. i ,p j ) indicates document p i and p j A function to determine if a content is duplicated. This indicates the indicator function, sim(p) i ,p j ) indicates document p i and p j semantic similarity, query j This indicates the j-th query. Let D represent the retrieval function, D represent the knowledge base, and k represent the retrieval function. dynamic This indicates the number of most relevant documents to return, Score(p,q,F) t () represents the overall score. and Let represent the corresponding weights, and sim(p,q) represent the semantic similarity between document p and iterative input q. This indicates that document p and all feedback F t The maximum similarity is given by Quality(p), which represents the quality score of document p, w1, w2, and w3 represent the corresponding weights, A(p) represents the authority score, C(p) represents the completeness score, and R(p) represents the readability score.

[0086] Optionally, step S5 includes:

[0087] The references in the optimized content are verified and corrected through automatic reference requirement identification, multi-level reference matching algorithm, and quantitative evaluation of reference accuracy, resulting in iterative output;

[0088] The function that automatically identifies the reference requirement is:

[0089]

[0090] In the formula, CitationCandidates(y t+1 ) indicates a quotation statement, S(y t+1 ) indicates optimized content y t+1 The set of sentences in the set, where RequiresCitation(s) represents a function to determine whether sentence s needs to be cited, F s (s) indicates a factual rating. S represents the factual threshold. s (s) represents the specificity score. Indicates a specificity threshold. Indicates an indicator function, Indicates parameters In the factual classification model, E(s) represents the set of entities in sentence s, N(s) represents the set of numerical information in sentence s, and W(s) represents the total vocabulary set in sentence s. Indicates the number of elements in the set;

[0091] The function for the multi-level reference matching algorithm is:

[0092]

[0093] In the formula, D candidate (s) represents the coarse-grained matching value, P t+1 This represents the set of documents retrieved in round t+1. This indicates that document p belongs to the retrieved document set P. t+1 sim(s,p) represents the semantic similarity between sentence s and document p. The value represents the coarse matching threshold; BestMatch(s,p) represents the fine matching value; argmax represents selecting the paragraph with the highest score; BERT-Score(s,π) represents the BERT semantic similarity score between sentence s and paragraph π; w represents the words in sentence s; w' represents the words in paragraph π; h w and h w' This represents the BERT embedding vectors of words w and w', where max represents the maximum value. Indicates cosine similarity;

[0094] The function for quantifying the accuracy of citations is:

[0095]

[0096] In the formula, CQ(c,s,p) is the comprehensive citation quality score of citation c, sentence s, and document p; w1, w2, and w3 represent the corresponding weights; CC represents the content consistency score; PA represents the positional accuracy score; and CR represents the context relevance score. The expressions indicate relational judgments: `content(c)` refers to the claimed content in document `c`, and `extract(p)` extracts the original content from document `p`. This indicates the indicator function, where position(c) represents the annotation position referencing c, and ValidPositions(p) is the set of valid positions for document p. The semantic similarity is represented by context(s), which represents the content of sentence s in the optimized content, and surrounding(p,c) which represents the content before and after the reference position in document p.

[0097] Optionally, the function ShouldStop(t) for determining the iteration termination condition in step S6 is:

[0098]

[0099] In the formula, QM represents that all dimensions meet the standard. This indicates that the improvement is not significant and the improvement trend is converging, where t is the current iteration number, T max E is the preset maximum number of iterations. i (y t ) represents the score of the i-th evaluation dimension, min i E i (y t The lowest score across all evaluation dimensions. Let represent the minimum threshold for the i-th evaluation dimension, SI represent the magnitude of improvement, and IR represent the magnitude of improvement in the current iteration. Indicates the improvement amplitude threshold. w represents the trend convergence threshold.i The corresponding weights are represented by CT, the trend improvement effect is represented by T(t), the average improvement rate in the current iteration is represented by Δt, and the iteration sliding window is represented by E. overall (y k ) represents the overall quality score for the kth round.

[0100] Optionally, step S6 further includes:

[0101] During the iteration process, a quality degradation detection algorithm is used for detection. If quality degradation is detected, a graded rollback strategy is adopted.

[0102] The function of the quality degradation detection algorithm is:

[0103] ;

[0104] In the formula, QD(t) represents the quality degradation detection result, and E overall (y t Eoverall(y) represents the overall quality score for round t. t-1 ) represents the overall quality score in round t-1. Indicates the degradation threshold;

[0105] The backoff decision function for the tiered backoff strategy is as follows:

[0106]

[0107] In the formula, RS(t) represents the result of the rollback strategy, LR represents local rollback, and GR represents global rollback. This represents the best historical quality score up to round t-1, and "otherwise" indicates other scores. rollback Indicates the content after rollback, y t-1 Let y represent the content of round t-1. best This indicates the round in which the best historical quality score was achieved. This indicates the iteration number corresponding to the best historical iteration output.

[0108] Optionally, step S6 further includes:

[0109] During the iteration process, the time complexity of the entire iteration process is calculated, and parallelization optimizations are performed based on the time complexity to parallelize the feedback generation, retrieval operation, and verification process.

[0110] The time complexity function is:

[0111] ;

[0112] In the formula, Let |Q| represent the time complexity function, |D| represent the knowledge base size, |P| represent the number of documents retrieved, |F| represent the average document complexity, and |F| represent the number of feedbacks. M represents the complexity of the fusion operation.

[0113] The function for parallelizing the feedback generation is:

[0114] ;

[0115] In the formula, Indicates scores across different dimensions. Represents a parallel computation function. Representing different evaluation functions;

[0116] The function for parallelizing the retrieval operation is as follows:

[0117] ;

[0118] In the formula, P new This indicates the addition of a new collection of documents for retrieval. Q represents the parallel retrieval function. additional This represents a dynamically generated supplementary query set, where D represents the knowledge base;

[0119] The function for parallelizing the verification process is as follows:

[0120] ;

[0121] In the formula, VR represents the newly added set of search documents. Represents a parallel verification function, C t Let P represent the set of references in round t. t This represents the corresponding collection of documents.

[0122] Secondly, the present invention provides a scientific and technological literature question-answering system based on self-feedback iterative optimization, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the scientific and technological literature question-answering method based on self-feedback iterative optimization of the first aspect.

[0123] The technical effects of the self-feedback iterative optimization-based question-and-answer system for scientific and technological literature provided in the second aspect are described in the relevant description of the self-feedback iterative optimization-based question-and-answer method for scientific and technological literature provided in the first aspect. Attached Figure Description

[0124] Figure 1 This is a schematic diagram of the main process of the scientific literature question-answering method based on self-feedback iterative optimization according to an embodiment of the present invention;

[0125] Figure 2This is a schematic diagram of the structure of a scientific literature question-answering system based on self-feedback iterative optimization according to an embodiment of the present invention.

[0126] Explanation of reference numerals in the attached figures:

[0127] 1: A scientific literature question-answering system based on self-feedback iterative optimization;

[0128] 2: Processor;

[0129] 3: Memory. Detailed Implementation

[0130] To better understand the above technical solutions, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the present invention can be understood more clearly and thoroughly, and that the scope of the present invention can be fully conveyed to those skilled in the art.

[0131] Example 1

[0132] In the field of scientific literature question answering, traditional large language models have many problems that need to be addressed. This embodiment uses a multi-dimensional self-feedback evaluation mechanism, a dynamic retrieval mechanism, a multi-level citation verification and attribution mechanism, and a progressive information fusion algorithm to form an intelligent system capable of continuous self-improvement, thereby significantly improving the accuracy of scientific literature question answering. Please refer to the discussion below for details.

[0133] Please refer to Figure 1 Scientific literature question-answering methods based on self-feedback iterative optimization include:

[0134] Before implementing the other steps of this embodiment, it is necessary to initialize and mathematically model the system architecture of this embodiment, which mainly consists of four core functional modules: generation module, self-feedback evaluation module, dynamic retrieval module, and information fusion module. These correspond to steps S1, S2, S3, and S4-S5 in this embodiment, respectively. This design not only ensures the scalability of the system but also facilitates customized optimization for different application scenarios.

[0135] Simultaneously, the mathematical model foundation for the system used in this embodiment is established. Let the input query be q∈Q, where Q represents the set of all possible queries. The external knowledge base is defined as D={d1,d2,…,d…} N}, where each document d i It includes text content and metadata information. The system's goal is to generate an optimal response. Make it fit within the predefined set of quality evaluation functions E={E1,E2,…,E kTo achieve optimal performance.

[0136] Define the iterative optimization process as a Markov decision process, with the state space as follows:

[0137] S t =(q,y t ,P t C t ,F t );

[0138] In the formula, y t It is the response of the t-th iteration. It is the collection of documents currently retrieved. It is the corresponding collection of references. It is a self-feedback set.

[0139] Step S1: Receive user questions and generate an initial response using a search-enhanced generation strategy.

[0140] In this embodiment, a traditional retrieval enhancement generation strategy is adopted in the initial stage to provide a baseline for subsequent iterative optimization. The design of this stage considers the balance between retrieval quality and computational efficiency, and ensures high-quality initial responses by optimizing the retrieval strategy.

[0141] The initial retrieval process can be represented as:

[0142] ;

[0143] In the formula, P0 is the initial set of documents to be retrieved, and Retrieve(q,D,k0) represents the initial retrieval function with inputs of query q, knowledge base D, and the number of documents to be retrieved k0. This means finding the largest subset P of size k0 in the knowledge base D. sim(q,p) is the semantic similarity function between query q and document p, typically using cosine similarity based on a vector space model.

[0144] ;

[0145] In the formula, v q and v p These are the embedding vector representations of query q and document p, respectively.

[0146] The process of building the initial suggestion template involves query restructuring and context organization. At this point, the suggestion template construction function is defined as:

[0147]

[0148] In the formula, prompt0 represents the structured text input to the large language model. This indicates a prompt template construction function whose input is query q and initial search document set P0. [INSTRUCTION] represents the generation instructions. This indicates a text concatenation operation. [CONTEXT] represents a context identifier. format(p) converts document p into a standardized format. This means that documents p converted to a standardized format are concatenated to form the initial retrieval document set P0.

[0149] The initial response generation process is implemented using a large language model:

[0150] ;

[0151] In the formula, This represents a large language model generation function with parameter θ, representing structured text prompt0. This model simultaneously outputs an initial response y0 and the corresponding set of references C0.

[0152] Step S2: Use the initial response as the iterative input, perform multi-dimensional evaluation on the iterative input, and obtain structured feedback with multiple specific feedbacks.

[0153] In this embodiment, the self-feedback mechanism comprehensively evaluates the quality of the current response from different perspectives through multiple specially designed evaluation functions and generates specific and actionable improvement suggestions. Specifically, step S2 includes:

[0154] Using the initial response as iterative input, the iterative input is evaluated for content organization quality, information coverage quantification, and citation accuracy, resulting in structured feedback with multiple specific feedbacks.

[0155] Content organization assessment aims to measure the logical structure and clarity of expression of the response. It calculates coherence scores by analyzing the semantic connections between sentences and evaluates structural quality through hierarchical analysis. Specifically, the content organization quality assessment function E... org (y t ,q) is:

[0156]

[0157] In the formula, y t Let q represent the user question, α, β, and γ represent the corresponding weights, and Cohe(y) represent the iterative input. t Stru(y) represents the coherence scoring function. t Rele(y) represents the structural quality scoring function. t ) represents the relevance scoring function, n represents the total number of sentences, sim(s) i ,s i+1) represents the semantic similarity between the i-th sentence and the (i+1)-th sentence in the iterative input, N valid N represents the number of valid paragraphs. total Indicates the total number of paragraphs. This represents the average paragraph integrity score.

[0158] Information coverage assessment focuses on whether the response comprehensively answers all aspects of the query. Specifically, the information coverage metric evaluation function E... cov (y t ,q) is:

[0159]

[0160] In the formula, Topics(q) represents the total topic of user question q, and topic1, topic2, and topic3 ​​are also included. m This represents the 1st, 2nd, and mth topics after decomposition. i This represents the i-th topic. Indicates whether the i-th topic was input in the t-th iteration. cover, Indicates an indicator function, This indicates the depth of discussion surrounding a topic in the responses. Indicates the topic Number of related sentences Indicate topic The expected length of the discussion.

[0161] Citation accuracy is a key indicator for applications in specialized fields. A two-layer verification mechanism was designed: existence verification and consistency verification. Existence verification ensures that each citation corresponds to a genuine source document, while consistency verification checks the degree of matching between the cited content and the original text. Specifically, the citation accuracy evaluation function E... cite (y t C t ,P t )for:

[0162]

[0163] In the formula, C t P represents a collection of references. t Let c represent the set of documents retrieved in the t-th iteration. i Represents the i-th reference, Exis(c i ) represents the existence verification function, Cons(c i ,source(c i ) represents the i-th reference c i With source document source(c i Content consistency score, ci .source indicates a reference to c i The source document, c i .position indicates a reference to c i In the source document, valid_positions represents the range of valid citation positions. Cons(c i ,p j ) represents the i-th reference c i With the j-th retrieved document p j The consistency verification function, BLEU stands for bilingual evaluation substitution, extract(p j ,c i .position) indicates retrieving document p j c i The original text extracted at the .position position, c i .content indicates a reference to c i The original text quoted in the document;

[0164] Structured Feedback F t The generating function of the i-th specific feedback for:

[0165] ;

[0166] In the formula, E represents the feedback generating function. i (y t ) represents the i-th evaluation dimension relative to the iterative input y. t Quality rating This represents the quality threshold for evaluation dimension i.

[0167] Each feedback It consists of the following three components:

[0168]

[0169] In the formula, issue i This represents the problem identified in the i-th specific feedback. This represents the problem identification function, suggestion. i This refers to the issue. i directions for improvement This indicates an improved recognition function, info_need i Indicates directions for improvement. i The specific information required is specified by extr_requ, which represents the information request function.

[0170] Step S3: Extract information requirements from the structured feedback, and perform supplementary retrieval based on the information requirements to obtain supplementary information.

[0171] In this embodiment, step S3 includes:

[0172] S31. Extract information requirements from the structured feedback, and supplement the structured feedback through direct expansion strategy, semantic rewriting strategy and context-aware strategy to obtain the final query set.

[0173] In this embodiment, the dynamic retrieval mechanism corresponding to step S31 can automatically identify information gaps based on self-feedback content and generate targeted retrieval queries, realizing the transformation from passive retrieval to active information acquisition.

[0174] The process of extracting information needs from feedback involves natural language understanding and semantic analysis; therefore, the information need extraction function InfoNeed in this embodiment... i for:

[0175]

[0176] In the formula, This represents the i-th specific feedback. Information demand conversion function, type i Keywords indicate the type of information requirement. i This refers to key information extracted from specific feedback, constraints i The constraints represent the information demand, argmax represents the set of parameters that take the maximum value, and K represents the set of selected keywords. This represents the i-th specific feedback. All entities, This indicates that keyword k is in the i-th specific feedback. The TF-IDF weights in the equation.

[0177] The function for the direct extension strategy is:

[0178] ;

[0179] In the formula, The direct extended query generated for the i-th information request is extract_nouns(q), a function that extracts phrases from the user question q.

[0180] The function for the semantic rewriting strategy is:

[0181] ;

[0182] In the formula, Let R represent the semantic rewrite query generated by the i-th information request, and let R represent the query rewriter based on the pre-trained language model.

[0183] The function for the context-aware strategy is:

[0184] ;

[0185] In the formula, This represents the context-aware query generated for the i-th information request. The context_from(y) function represents the fusion function. t ) represents the iterative input y t The contextual information mentioned in the text.

[0186] The final query set is generated through a weighted fusion of direct expansion, semantic rewriting, and context-aware strategies, and its function is as follows:

[0187] ;

[0188] In the formula, This represents the final query set for the i-th information request. For the union operation, k t It represents the total number of specific feedbacks in the t-th iteration.

[0189] S32. For the final query set, supplementary information is obtained by filtering through multiple rounds of parallel retrieval, document deduplication, and multi-dimensional relevance filtering.

[0190] The enhanced retrieval stage, corresponding to step S32, not only aims to acquire new information but also ensures the quality and relevance of the retrieval results. This stage employs techniques such as multi-round retrieval, deduplication filtering, and relevance ranking to guarantee the high quality of the retrieved information. Therefore, supplementary information is obtained in this embodiment. The function is:

[0191]

[0192] In the formula, This represents the collection of documents after deduplication. This indicates that document p belongs to the document collection. , The correlation threshold, This is a deduplication function. Represents the original document collection. IsDuplicate(p) is the repetition threshold. i ,p j ) indicates document p i and p j A function to determine if a content is duplicated. This indicates the indicator function, sim(p) i ,p j ) indicates document p i and p j semantic similarity, query j This indicates the j-th query. Let D represent the retrieval function, D represent the knowledge base, and k represent the retrieval function. dynamic This indicates the number of most relevant documents to return, Score(p,q,F) t () represents the overall score. and Let represent the corresponding weights, and sim(p,q) represent the semantic similarity between document p and iterative input q. This indicates that document p and all feedback F t The maximum similarity is given by Quality(p), which represents the quality score of document p, w1, w2, and w3 represent the corresponding weights, A(p) represents the authority score, C(p) represents the completeness score, and R(p) represents the readability score.

[0193] Step S4: Based on the priority of the specific feedback, the supplementary information and iterative input are progressively fused to obtain the optimized content.

[0194] In this embodiment, information fusion is the key step in the entire iterative process, determining whether the new information can effectively improve the response quality. The progressive fusion algorithm designed in this invention can improve specific aspects while maintaining the original high-quality content. Specifically, step S4 includes:

[0195] Prioritize the feedback based on its importance, urgency, and feasibility, and then integrate the supplementary information and iterative inputs in stages, including structural adjustment, content supplementation, and reference correction, to obtain optimized content.

[0196] The feedback prioritization process comprehensively considers importance, urgency, and feasibility. Importance is assessed based on the problem's impact on overall quality, urgency reflects the problem's severity, and feasibility assesses the difficulty of improvement. Therefore, the priority ranking function in this embodiment... for:

[0197]

[0198] In the formula, α, β, and γ represent the corresponding weights, This represents the importance evaluation function. Indicates the urgency assessment function, This represents the feasibility assessment function. Indicates the current quality loss. Indicates the maximum possible mass loss. E represents the quality threshold for evaluation dimension i. i (y t ) represents the i-th evaluation dimension relative to the iterative input y. t Quality rating Indicates the number of relevant search documents. Indicates the minimum number of documents required.

[0199] The fusion process employs a phased strategy, with each phase handling a specific type of feedback. This includes prioritizing feedback related to organizational structure, followed by feedback related to information coverage, then feedback related to citation accuracy, and finally modeling each fusion step as a conditional text generation process. Therefore, the optimization content of this embodiment is... t+1 The fusion generation function is:

[0200]

[0201] In the formula, Indicates a reference to the fusion function, y content F represents the intermediate response after content supplementation. citation This indicates feedback from the reference class. This indicates supplementary information including a supplementary document collection. This represents the content fusion function, y struct F represents the intermediate response after structural supplementation. coverage This indicates content-based feedback. F represents the structure fusion function. struct This indicates a structured feedback. Let y represent the probability distribution function. new This represents the new text generated during the current fusion phase, y current f represents the pre-processed text generated during the current fusion phase. i P represents the feedback at the current integration stage. relvant P(y) represents a subset of supplementary documents related to the current fusion phase. new |y current ,f i ,P relevant ) indicates the preceding text y in the current fusion phase. current Feedback f i and supplementary document subset P relevant Generate new text y under common constraints new The probability, w j Represents the new text y new The j-th word in; w <j P(w) represents all words preceding the j-th word in the new text. j |w <j y current ,f i,P relevant ) indicates all the preceding words w <j And the pre-text y of the current fusion phase current Feedback f i and supplementary document subset P relevant Generate j words w under the common constraints j The probability of.

[0202] Step S5: Verify and correct the references in the optimized content to obtain the iterative output.

[0203] In this embodiment, the citation verification mechanism ensures that every academic claim in the generated content is supported by reliable literature, which is crucial for applications in professional fields. Specifically, step S5 includes:

[0204] The citations in the optimized content are verified and corrected through automatic identification of citation requirements, multi-level citation matching algorithms, and quantitative evaluation of citation accuracy, resulting in iterative output.

[0205] In this embodiment, a hybrid approach based on rules and machine learning is used to identify the statements that need to be referenced. Therefore, the function for automatically identifying reference requirements in this embodiment is:

[0206]

[0207] In the formula, CitationCandidates(y t+1 ) indicates a quotation statement, S(y t+1 ) indicates optimized content y t+1 The set of sentences in the set, where RequiresCitation(s) represents a function to determine whether sentence s needs to be cited, F s (s) indicates a factual rating. S represents the factual threshold. s (s) represents the specificity score. Indicates a specificity threshold. Indicates an indicator function, Indicates parameters In the factual classification model, E(s) represents the set of entities in sentence s, N(s) represents the set of numerical information in sentence s, and W(s) represents the total vocabulary set in sentence s. This indicates the number of elements in the set.

[0208] The reference matching process is divided into two stages: coarse-grained matching and fine-grained verification. Therefore, the function of the multi-level reference matching algorithm in this embodiment is:

[0209]

[0210] In the formula, Dcandidate (s) represents the coarse-grained matching value, P t+1 This represents the set of documents retrieved in round t+1. This indicates that document p belongs to the retrieved document set P. t+1 sim(s,p) represents the semantic similarity between sentence s and document p. The value represents the coarse matching threshold; BestMatch(s,p) represents the fine matching value; argmax represents selecting the paragraph with the highest score; BERT-Score(s,π) represents the BERT semantic similarity score between sentence s and paragraph π; w represents the words in sentence s; w' represents the words in paragraph π; h w and h w' This represents the BERT embedding vectors of words w and w', where max represents the maximum value. This represents the cosine similarity.

[0211] This embodiment designs a multi-dimensional evaluation index for citation accuracy. Specifically, the function for quantitatively evaluating citation accuracy in this embodiment is:

[0212]

[0213] In the formula, CQ(c,s,p) is the comprehensive citation quality score of citation c, sentence s, and document p; w1, w2, and w3 represent the corresponding weights; CC represents the content consistency score; PA represents the positional accuracy score; and CR represents the context relevance score. The expressions indicate relational judgments: `content(c)` refers to the claimed content in `c`, and `extract(p)` extracts the original content from document `p`. This indicates the indicator function, where position(c) represents the annotation position referencing c, and ValidPositions(p) is the set of valid positions for document p. The semantic similarity is represented by context(s), which represents the content of sentence s in the optimized content, and surrounding(p,c) which represents the content before and after the reference position in document p.

[0214] Step S6: Determine whether the iteration termination condition is met based on the iteration output. If so, use the iteration output as the final response; otherwise, input the iteration output into step S2 as the iteration input for iterative optimization until the final response is obtained.

[0215] In this embodiment, step S6 includes:

[0216] S61. Determine whether the iteration termination condition is met based on the iteration output. If so, use the iteration output as the final response; otherwise, input the iteration output into step S2 as the iteration input for iterative optimization until the final response is obtained.

[0217] The iterative control mechanism ensures that the system stops optimizing at an appropriate time, avoiding quality degradation or waste of computational resources caused by excessive iteration. The function ShouldStop(t) for determining the iteration termination condition is:

[0218]

[0219] In the formula, QM represents that all dimensions meet the standard. This indicates that the improvement is not significant and the improvement trend is converging, where t is the current iteration number, T max E is the preset maximum number of iterations. i (y t ) represents the score of the i-th evaluation dimension, min i E i (y t The lowest score across all evaluation dimensions. Let represent the minimum threshold for the i-th evaluation dimension, SI represent the magnitude of improvement, and IR represent the magnitude of improvement in the current iteration. Indicates the improvement amplitude threshold. w represents the trend convergence threshold. i The corresponding weights are represented by CT, the trend improvement effect is represented by T(t), the average improvement rate in the current iteration is represented by Δt, and the iteration sliding window is represented by E. overall (y k ) represents the overall quality score for the kth round.

[0220] Step S62: During the iteration process, a quality degradation detection algorithm is used for detection. If quality degradation is detected, a graded rollback strategy is adopted.

[0221] To prevent quality degradation during the iteration process, a real-time quality monitoring and intelligent rollback mechanism were designed. Therefore, the function of the quality degradation detection algorithm in this embodiment is:

[0222] ;

[0223] In the formula, QD(t) represents the quality degradation detection result, and E overall (y t Eoverall(y) represents the overall quality score for round t. t-1 ) represents the overall quality score in round t-1. This represents the degradation threshold.

[0224] When quality degradation is detected, the system adopts a tiered rollback strategy, and its rollback decision function is:

[0225]

[0226] In the formula, RS(t) represents the result of the rollback strategy, LR represents local rollback, and GR represents global rollback. This represents the best historical quality score up to round t-1, and "otherwise" indicates other scores. rollback Indicates the content after rollback, y t-1 Let y represent the content of round t-1. best This indicates the round in which the best historical quality score was achieved. This indicates the iteration number corresponding to the best historical iteration output.

[0227] Example 2

[0228] Based on the self-feedback iterative optimization-based scientific literature question-answering method, and building upon the first embodiment described above, researchers often need to obtain high-quality literature analysis results within a limited timeframe in actual scientific research workflows. While traditional multi-round iterative optimization can improve answer quality, it is often accompanied by a significant increase in computational cost and a prolonged response time, which greatly limits the practical application value of the system. Therefore, step S6 of this embodiment further includes:

[0229] Step S63: During the iteration process, the time complexity of the entire iteration process is calculated, and parallelization optimization is performed based on the time complexity to parallelize the feedback generation, retrieval operation, and verification process.

[0230] The time complexity function for the entire iteration process is:

[0231] ;

[0232] In the formula, Let |Q| represent the time complexity function, |D| represent the knowledge base size, |P| represent the number of documents retrieved, |F| represent the average document complexity, and |F| represent the number of feedbacks. M represents the complexity of the fusion operation.

[0233] The function for parallelizing the feedback generation is:

[0234] ;

[0235] In the formula, Indicates scores across different dimensions. Represents a parallel computation function. These represent different evaluation functions.

[0236] The function for parallelizing the retrieval operation is as follows:

[0237] ;

[0238] In the formula, P new This indicates the addition of a new collection of documents for retrieval. Q represents the parallel retrieval function. additional This represents a dynamically generated supplementary query set, and D represents the knowledge base.

[0239] The function for parallelizing the verification process is as follows:

[0240] ;

[0241] In the formula, VR represents the newly added set of search documents. Represents a parallel verification function, C t Let P represent the set of references in round t. t This represents the corresponding collection of documents.

[0242] Computational complexity analysis identified which parts of the system are most time-consuming and grow fastest with the input size. By parallelizing feedback generation, retrieval operations, and the verification process, the order of complexity is reduced, ultimately optimizing system efficiency. This ensures that the system improves question-answering quality without sacrificing practicality due to excessive computational cost.

[0243] Example 3

[0244] Please refer to Figure 2 A scientific literature question-and-answer system 1 based on self-feedback iterative optimization includes a memory 3, a processor 2, and a computer program stored on the memory 3 and run on the processor 2. When the processor 2 executes the computer program, it implements the steps in Embodiment 1 or 2 above.

[0245] In summary, the present invention has the following advantages:

[0246] (1) By introducing a multi-dimensional self-feedback evaluation mechanism, the present invention enables the scientific literature question-and-answer system to autonomously identify the deficiencies of the generated content in terms of organizational structure, information coverage, academic depth, etc., and automatically generate targeted improvement strategies, thereby significantly improving the quality stability and professional level of academic question-and-answer.

[0247] (2) The dynamic retrieval mechanism designed in this invention can automatically identify information gaps based on self-feedback content and generate targeted supplementary retrieval queries, realizing the transformation from passive response to active information acquisition. This mechanism is particularly suitable for scientific question-and-answer scenarios that require interdisciplinary knowledge integration, and can significantly improve the comprehensiveness and precision of literature retrieval.

[0248] (3) This invention achieves fully automated processing from citation requirement identification, literature matching, content verification to format standardization by constructing a multi-level citation verification and attribution mechanism. This mechanism ensures that every academic statement generated by the scientific literature question-and-answer system has reliable literature support, significantly improving the credibility and practicality of the system in scenarios such as academic writing assistance, research report generation, and professional consultation.

[0249] (4) The progressive information fusion algorithm designed in this invention ensures the logical consistency and academic rigor of the content in each information integration process through phased processing, priority sorting, and consistency checks. This algorithm is particularly suitable for academic application scenarios that require the integration of multi-source information, such as review literature writing, research status analysis, and technology development trend prediction.

[0250] (5) This invention achieves an optimal balance between computational efficiency and generated quality by designing intelligent iteration termination conditions, efficient parallel retrieval strategies, optimized information fusion algorithms, and quality monitoring mechanisms. The system can adaptively adjust optimization strategies according to different application scenarios, ensuring both the professional level of academic answers and meeting the timeliness requirements of practical work.

[0251] Therefore, this invention addresses key technical issues in scientific literature question-answering systems by employing an iterative optimization method based on a large language model with a self-feedback mechanism. This method not only significantly improves the quality of scientific literature information retrieval, comprehensive analysis, and knowledge question-answering responses but also meets the timeliness requirements in practical work.

[0252] Since the systems / devices described in the above embodiments of the present invention are systems / devices used to implement the methods of the above embodiments of the present invention, those skilled in the art can understand the specific structure and modifications of the systems / devices based on the methods described in the above embodiments of the present invention, and therefore will not be repeated here. All systems / devices used in the methods of the above embodiments of the present invention fall within the scope of protection of the present invention.

[0253] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0254] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (devices), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions.

[0255] It should be noted that any reference numerals placed between parentheses in the claims should not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claims. The word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. The invention can be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In claims that enumerate several means, several of these means may be embodied by the same hardware. The use of the terms first, second, third, etc., is merely for convenience of expression and does not indicate any order. These terms can be understood as part of the component names.

[0256] Furthermore, it should be noted that in the description of this specification, the terms "one embodiment," "some embodiments," "embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0257] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the claims should be interpreted to include both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0258] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, then this invention should also include these modifications and variations.

Claims

1. A method for answering scientific literature questions based on self-feedback iterative optimization, characterized in that, The method comprises the following steps: Step S1, receiving a user question, and generating an initial response by using a search enhancement generation strategy; Step S2, taking the initial response as an iterative input, performing multi-dimensional evaluation on the iterative input, and obtaining structured feedback with multiple specific feedbacks; Step S3, extracting information requirements from the structured feedback, and performing supplementary search according to the information requirements to obtain supplementary information; Step S4, performing progressive information fusion on the supplementary information and the iterative input according to the priority ranking of the specific feedbacks to obtain optimized content; Step S5, verifying and correcting references in the optimized content to obtain an iterative output; Step S6, judging whether the iterative output meets an iterative termination condition, if yes, taking the iterative output as a final response, or if no, inputting the iterative output into step S2 as an iterative input for iterative optimization until a final response is obtained; The step S2 comprises: taking the initial response as an iterative input, performing content organization quality evaluation, information coverage quantitative evaluation and reference accuracy evaluation on the iterative input, and obtaining structured feedback with multiple specific feedbacks; wherein the content organization quality evaluation function E org (y t q) is: ； ； ； where y t represents the iterative input, q represents the user question, a, b, g represent the corresponding weights, Cohe(y t ) represents the coherence scoring function, Stru(y t ) represents the structure quality scoring function, Rele(y t ) represents the relevance scoring function, n represents the total number of sentences, sim(s i ,s i+1 ) represents the semantic similarity between the ith sentence and the (i+1)th sentence in the iterative input, N valid represents the number of valid paragraphs, N total represents the total number of paragraphs, represents the average paragraph integrity score; wherein the information coverage quantification evaluation function E cov (y t q) is: ； ； ； In the formula, Topics(q) represents the total topic of user question q, and topic1, topic2, and topic3 ​​are also included. m This represents the 1st, 2nd, and mth topics after decomposition. i This represents the i-th topic. Indicates whether the i-th topic was input in the t-th iteration. cover, Indicates an indicator function, This indicates the depth of discussion surrounding a topic in the responses. Indicates the topic Number of related sentences Indicate topic The expected length of the discussion; wherein the reference accuracy evaluation function is: ； ； ； where C t represents a set of citations, P t represents a set of search documents in the tth iteration, c i represents the ith citation, Exis(c i ) represents an existence verification function, Cons(c i , source(c i ) represents the ith citation c i 's consistency score with the content of the source document source(c i ), c i .source represents the source document of citation c i , c i .position represents the position of citation c i in the source document, valid_positions represents the range of valid citation positions in the source document, Cons(c i , p j ) represents the ith citation c i 's consistency verification function with the jth search document p j , BLEU represents Bilingual Evaluation Underdevelopment, extract(p j , c i .position) represents the extracted source content in p j at position c i .position, and c i .content represents the cited source content in citation c i . The structured feedback F t The generating function of the i-th specific feedback in the middle is: ; wherein represents the feedback generating function, E i (y t ) represents the quality score of the ith evaluation dimension for the iteration input y t represents the quality threshold of the evaluation dimension i;​ wherein each feedback comprises the following three components: ； ； ； where issue i represents the problem identified in the i-th specific feedback, represents the problem identification function, suggestion i represents the improvement direction for the problem issue i , represents the improvement identification function, info_need i represents the specific information required to implement the improvement direction suggestion i , extr_requ represents the information requirement function.

2. The self-feedback iteration optimization based scientific literature question answering method of claim 1, wherein, The step S4 comprises: performing priority ranking according to importance, urgency and feasibility of the specific feedbacks, and performing stage-by-stage fusion on the supplementary information and the iterative input according to a structure adjustment stage, a content supplement stage and a reference correction stage to obtain optimized content; wherein the priority ordering function is: ； ； ； ； wherein a, b, g represent the corresponding weights respectively, represents the importance evaluation function, represents the urgency evaluation function, represents the feasibility evaluation function, represents the current quality loss, represents the maximum possible quality loss, represents the quality threshold of the evaluation dimension i, E i (y t ) represents the quality score of the ith evaluation dimension on the iterative input y t , represents the number of relevant search documents, represents the minimum number of documents required; wherein the optimization content y t+1 The fusion generating function is: ； ； ； ； wherein, represents a reference fusion function, y content represents an intermediate response after content supplement, F citation represents a reference class feedback, represents supplement information including a supplement document set, represents a content fusion function, y struct represents an intermediate response after structure supplement, F coverage represents a content class feedback, represents a structure fusion function, F struct represents a structure class feedback, represents a probability distribution function, y new represents a new text generated in a current fusion stage, y current represents a previous text generated in a current fusion stage, f i represents a feedback in a current fusion stage, P relvant represents a supplement document subset related to a current fusion stage, P(y new |y current ,f i ,P relevant ) represents a probability of generating a new text y current under the joint constraint of a previous text y i , a feedback f relevant , and a supplement document subset P new in a current fusion stage, w j represents the jth word in the new text y new ; w <j represents all words before the jth word in the new text, P(w j |w <j , y current ,f i ,P relevant ) represents a probability of generating the jth word w <j under the joint constraint of all previous words w current and a previous text y i , a feedback f relevant , and a supplement document subset P j in a current fusion stage.

3. The self-feedback iteration optimization based scientific literature question answering method of claim 1, wherein, The step S3 comprises: extracting information requirements from the structured feedback, and performing supplementary search on the structured feedback by using a direct expansion strategy, a semantic rewriting strategy and a context awareness strategy to obtain a final query set; where the information need extraction function InfoNeed i is: ； ； wherein, represents the information requirement conversion function of the ith specific feedback , type i represents the type of information requirement, keywords i represents the key information extracted from the specific feedback, constraints i represents the constraint condition of the information requirement, argmax represents the parameter set taking the maximum value, K represents the selected key word set, represents all entities of the ith specific feedback , represents the TF-IDF weight of the key word k in the ith specific feedback . wherein a function of the direct expansion strategy is: ； wherein extract_nouns(q) is a function that extracts phrases from the user question q, for the i-th information need generated by the direct expansion query wherein a function of the semantic rewriting strategy is: ； wherein Ri denotes the semantic rewritten query generated by the ith information need, and R denotes the query rewriter based on the pre-trained language model. wherein a function of the context awareness strategy is: ； wherein, represents a context-aware query generated by the ith information need, represents a fusion function, context_from(y t ) represents the context information raised in the iteration input y t . wherein the final query set is generated by weighted fusion of the direct expansion strategy, the semantic rewriting strategy and the context awareness strategy, and a function thereof is: ； wherein denotes the final query set for the i-th information need, is a union operation, k t is the total number of specific feedback in the t-th iteration.

4. The self-feedback iteration optimization based scientific literature question answering method of claim 3, wherein, The step S3 further comprises: for the final query set, performing screening by using multi-round parallel search, document deduplication and multi-dimensional relevance filtering to obtain supplementary information; where the function of the supplementary information is: ； ； ； ； ； ； wherein, represents the document set after deduplication, represents that the document p belongs to the document set , is a relevance threshold, is a deduplication function, represents the original document set, is a duplication threshold, IsDuplicate(p i ,p j ) represents a judgment function whether the documents p i and p j are duplicated, represents an indication function, sim(p i ,p j ) represents semantic similarity of the documents p i and p j , query j represents the jth query, represents a retrieval function, D represents a knowledge base, k dynamic represents the number of the most relevant documents set to return, Score(p, q, F t ) represents a comprehensive score, and represent corresponding weights, sim(p, q) represents semantic similarity of the document p and the iterative input q, represents the maximum similarity of the document p to all feedback F t , Quality(p) represents a quality score of the document p, w1, w2 and w3 represent corresponding weights, A(p) represents an authority score, C(p) represents a completeness score, and R(p) represents a readability score.

5. The self-feedback iteration optimization based scientific literature question answering method of claim 1, wherein, The step S5 comprises: verifying and correcting references in the optimized content by using reference requirement automatic identification, a multi-level reference matching algorithm and reference accuracy quantitative evaluation to obtain an iterative output; wherein a function of the reference requirement automatic identification is: ； ； ； ； In the formula, CitationCandidates(y t+1 ) represents a citation sentence, S(y t+1 ) represents a sentence set in the optimization content y t+1 , RequiresCitation(s) represents a judgment function whether the sentence s needs a citation, F s (s) represents a fact score, represents a fact threshold value, S s (s) represents a specificity score, represents a specificity threshold value, represents an indication function, represents a fact classification model of parameters , E(s) represents an entity set in the sentence s, N(s) represents a numerical information set in the sentence s, W(s) represents a total vocabulary set in the sentence s, represents the number of elements in the set; wherein a function of the multi-level reference matching algorithm is: ； ； ； where D candidate (s) denotes a coarse-grained matching value, P t+1 denotes the set of retrieved documents for the t+1 iteration, denotes that the passage p belongs to the set of retrieved documents P t+1 , sim(s, p) denotes the semantic similarity between the sentence s and the passage p, denotes a coarse matching threshold; BestMatch(s, p) denotes a fine-grained matching value, argmax denotes the selection of the passage with the highest score, BERT-Score(s, π) denotes the BERT semantic similarity score between the sentence s and the passage π, w denotes a word in the sentence s, w' denotes a word in the passage π, h w and h w' denote the BERT embedding vectors of the word w and the word w', max denotes the maximum value, denotes the cosine similarity; wherein a function of the reference accuracy quantitative evaluation is: ； ； ； ； where CQ(c,s,p) is a comprehensive reference quality score referencing c, sentence s and document p, w1, w2 and w3 represent corresponding weights, CC represents a content consistency value, PA represents a position accuracy value, CR represents a context relevance value, represents a relationship judgment, content(c) represents a claimed content in reference c, extract(p) represents an extracted original content in document p, represents an indicator function, position(c) represents an annotated position of reference c, ValidPositions(p) is a set of valid positions of document p, represents a semantic similarity, context(s) represents a content of sentence s in the optimized content, surrounding(p,c) represents a content before and after the reference position in document p.

6. The self-feedback iteration optimization based scientific literature question answering method of claim 1, wherein, a judgment function ShouldStop(t) of the iterative termination condition of the step S6 is: ； ； ； ； ； ； where QM represents all dimensions of compliance, represents that the improvement range is not significant and the improvement trend converges, t is the current iteration round number, T max is the preset maximum iteration number, E i (y t ) represents the score of the i-th evaluation dimension, min i E i (y t ) is the lowest score in all evaluation dimensions, represents the lowest threshold of the i-th evaluation dimension, SI represents the improvement range effect, IR represents the improvement range of the current iteration round, represents the improvement range threshold, represents the trend convergence threshold, w i represents the corresponding weight, CT represents the improvement trend effect, T(t) represents the average improvement rate of the current iteration round, Δt represents the iteration sliding window, E overall (y k ) represents the overall quality score of the k-th round.

7. The self-feedback iteration optimization based scientific literature question answering method of claim 6, wherein, The step S6 further comprises: in the iterative process, detecting by using a quality degradation detection algorithm, and if quality degradation is found, adopting a hierarchical rollback strategy; wherein a function of the quality degradation detection algorithm is: ； where QD(t) represents the quality degradation detection result, E overall (y t ) represents the overall quality score of the tth round, Eoverall(y t-1 ) represents the overall quality score of the (t-1)th round, represents the degradation threshold value; wherein a rollback decision function of the hierarchical rollback strategy is: ； ； ； where RS(t) represents the result of the rollback strategy, LR represents a local rollback, and GR represents a global rollback, represents the historical best quality score up to the t-1th round, otherwise represents other, and y rollback represents the content after rollback, and y t-1 represents the content of the t-1th round, and y best represents the content of the round in which the historical best quality score is located, represents the round corresponding to the historical best iteration output.

8. The self-feedback iteration optimization based scientific literature question answering method of claim 6, wherein, The step S6 further comprises: in the iterative process, calculating a time complexity of the entire iterative process, and performing parallel optimization of feedback generation parallelization, search operation parallelization and verification process parallelization according to the time complexity; wherein a function of the time complexity is: ； wherein denotes the time complexity function, T denotes the number of iterations, |Q| denotes the number of queries, |D| denotes the size of the knowledge base, |P| denotes the number of retrieved documents, L denotes the average complexity of the documents, |F| denotes the number of feedbacks, and M denotes the complexity of the fusion operation. Wherein, the function of feedback generation parallelization is: ； wherein denotes different dimension scores, denotes parallel computing functions, denotes different evaluation functions; Wherein, the function of retrieval operation parallelization is: ； where P new represents a newly added search document set, represents a parallel search function, Q additional represents a dynamically generated supplementary query set, and D represents a knowledge base; Wherein, the function of verification process parallelization is: ； wherein VR represents a newly added search document set, represents a parallel verification function, C t represents a reference set of the tth round, P t represents a corresponding document set.

9. A scientific literature question answering system based on self-feedback iterative optimization, comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that, The processor implements the computer program to realize the scientific and technological literature question and answer method based on self-feedback iterative optimization in any one of claims 1 to 8.

Images