Retrieval method and device for interpretable academic papers
By using multi-dimensional academic element decomposition and hybrid retrieval technology, combined with knowledge extraction and semantic ranking, the problem of insufficient semantic understanding and interpretability in academic paper retrieval is solved, achieving accurate, efficient, and interpretable academic paper retrieval.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUAZHONG NORMAL UNIV
- Filing Date
- 2026-01-14
- Publication Date
- 2026-05-08
AI Technical Summary
Existing academic paper retrieval methods suffer from insufficient semantic understanding, lack of interpretability of retrieval results, and weak fine-grained matching capabilities, failing to effectively meet the efficient and transparent retrieval needs of researchers.
By decomposing academic elements in multiple dimensions and using a hybrid retrieval mechanism, the system extracts knowledge from papers and queries in multiple dimensions using a knowledge extraction model. Combined with keyword and vector retrieval, it generates interpretable text and provides structured paper representation and semantic ranking.
It significantly improves the accuracy, relevance, and interpretability of academic paper retrieval, providing academic researchers with an efficient and transparent literature retrieval tool.
Smart Images
Figure CN121996777A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of information retrieval and natural language processing, specifically to an interpretable academic paper retrieval method and apparatus based on multi-dimensional academic element extraction. Background Technology
[0002] The number of academic papers is growing exponentially, and how to quickly and accurately retrieve relevant research results from this massive amount of literature is a significant challenge for researchers. Existing academic paper retrieval methods mainly fall into two categories: keyword-based retrieval and semantic embedding-based retrieval. However, these methods have the following limitations in practical applications: Keyword-based retrieval methods (such as BM25) rely on precise lexical matching, ranking relevance by calculating the degree of match between query terms and words in documents. The advantages of these methods are fast retrieval speed and high interpretability, but they suffer from significant lexical gap problems: they cannot match when queries and papers use different but semantically related terms; for example, the query "deep learning" cannot match papers using "neural network." They are insensitive to synonyms, abbreviations, and variations of academic terms; they focus only on literal matching and fail to understand the semantic intent of the query, resulting in low retrieval recall.
[0003] Semantic embedding-based retrieval methods map queries and papers to a high-dimensional vector space and utilize vector similarity for retrieval. While these methods can alleviate the lexical gap problem to some extent and capture semantic similarity, they also have drawbacks: They encode the entire paper or query into a single vector using holistic embedding, resulting in coarse-grained matching that fails to distinguish the relevance of papers across different academic dimensions. For example, two papers may have similar research methods but completely different research subjects, leading to high overall similarity but low actual relevance. Furthermore, the retrieval process is black-box; users cannot understand why a particular paper was retrieved, lacking interpretability. Finally, they ignore the structural features of academic papers, failing to fully utilize the semantic information in dimensions such as research subject, research method, and research field.
[0004] More importantly, existing methods generally lack consideration for the specific characteristics of academic paper retrieval scenarios. Academic paper retrieval differs fundamentally from general document retrieval: academic queries often focus on specific dimensions; for example, users may only be interested in papers that "use a certain research method" or "target a specific research object," rather than papers that are generally similar. Academic papers possess highly structured semantic features; elements such as research object, research method, research field, dataset, and key technologies constitute the core semantic framework of the paper, and this structured information is not fully utilized. Researchers need to understand the matching criteria of search results in order to quickly determine the relevance of papers, but existing methods lack the ability to interpret search results.
[0005] Furthermore, both keyword-based and embedding-based methods have their advantages and disadvantages. The former is accurate but has low recall, while the latter has high recall but may introduce irrelevant results. How to organically combine the advantages of both methods is a problem that urgently needs to be solved. In recent years, knowledge extraction technology, especially information extraction methods based on pre-trained language models, has made significant progress, enabling the accurate extraction of structured information such as entities and relationships from unstructured text, providing a technical foundation for fine-grained semantic analysis of academic papers. However, how to apply knowledge extraction technology to academic paper retrieval, construct a multi-dimensional retrieval framework, and provide interpretable retrieval results still lacks a systematic technical solution. Summary of the Invention
[0006] The purpose of this invention is to provide an interpretable academic paper retrieval method and apparatus, which aims to solve the problems of insufficient semantic understanding, lack of interpretability of retrieval results, and weak fine-grained matching ability of existing academic paper retrieval methods. By introducing multi-dimensional academic element decomposition and hybrid retrieval mechanism, the accuracy, relevance and interpretability of academic paper retrieval can be significantly improved, providing academic researchers with an efficient and transparent literature retrieval tool.
[0007] To achieve the above objectives, in a first aspect, the present invention provides an interpretable academic paper retrieval method, comprising: S1. Define n academic element dimensions, perform semantic parsing on the title and abstract of the paper to be retrieved, extract each of the n academic element dimensions to form an academic element set; combine the title, abstract and academic element set to output a structured paper representation; n is a positive integer; S2. Vectorize the text content of each academic element to obtain a vector representation, and establish a mapping relationship between academic elements and papers based on the text content and vector representation. S3. Receive user queries, perform semantic parsing on the queries, extract according to n academic element dimensions, and output the query element set. If a certain academic element dimension is not involved in the query, the text content of that academic element dimension is empty. S4. Traverse the set of query elements, merge the search results of the same academic element dimension, aggregate the search results of all non-empty academic element dimensions, and output the set of candidate papers. S5. Reorder the candidate paper set and output the sorted paper result set; S6. Extract the retrieval score and query element set of each paper in the paper result set under each academic element dimension, generate interpretable text based on the retrieval score and query elements, and obtain interpretable academic paper retrieval results.
[0008] According to the interpretable academic paper retrieval method provided by the present invention, S1 specifically includes: Obtain the title and abstract of the paper to be searched, and use them as input text: Input = {Title, Abstract}. Define n academic element dimensions, which include research object, research method, research field, dataset / experimental object, key technology, research question, application scenario, and evaluation index. The first knowledge extraction model is used to perform semantic parsing on the input text, extracting elements from each of the defined n academic element dimensions. The first knowledge extraction model includes a large-scale language model, a sequence labeling model, an information extraction model, and a named entity recognition model. The extraction results form an academic element set Elements = {E1, E2, ..., E...} n}, where E i ={Type, Content} (i = 1, 2, ..., n) represents the type and text content of the i-th academic element dimension; The title, abstract, and elements are combined to output a structured paper representation, Paper = {Title, Abstract, Elements}.
[0009] According to the interpretable academic paper retrieval method provided by the present invention, S2 specifically includes: Store each academic element E in the academic element set Elements i Text(E) of the text content of (i = 1, 2, ..., n) i = Content(E) i ); The text content of each academic element is vectorized and encoded using the embedding model, resulting in a vector representation Vec(E). i = Embedding(Text(E) i (i = 1, 2, ..., n); A keyword index, Index_keyword, is constructed based on text content, and a vector index, Index_vector, is constructed based on vector representation. Establish a mapping relationship between academic elements and papers, recording each academic element (E). i The paper and academic element type (E) to which (i = 1, 2, ..., n) belong. i It supports reverse locating the complete paper from the retrieved academic elements.
[0010] According to the interpretable academic paper retrieval method provided by the present invention, S3 specifically includes: Receive user input queries, which include natural language questions, keyword combinations, and research question descriptions; The second knowledge extraction model is used to perform semantic parsing on the query Query, extracting elements according to the n academic element dimensions defined in S1. The second knowledge extraction model may be the same as or different from the first knowledge extraction model. The extraction results form a set of query elements QueryElements = {QE1, QE2, ..., QE...} n}, where QE i = {Type, Content} (i = 1,2, ..., n) represents the query element type and text content of the i-th academic element dimension. If the query does not involve a certain academic element dimension, the text content of that academic element dimension is empty. According to the interpretable academic paper retrieval method provided by the present invention, S4 specifically includes: Iterate through the set of query elements: QueryElements = {QE1, QE2, ..., QE n For each query element whose text content is not empty, QE i (i = 1, 2, ..., n), perform keyword search and vector search respectively; QE for the same query element i The keyword search results and vector search results (i = 1, 2, ..., n) are fused to obtain the search results for this academic element dimension: Results(QE) i ); Aggregate all search results for academic elements (excluding empty text content), deduplicate and merge them using paper identifiers, resulting in a candidate paper set: CandidatePapers = Aggregate({Results(QE... i )|Content(QE i )≠ The function `Score_dim(Paper, i=1,2,...,n)` records the retrieval score of each candidate paper across various academic elements. j QE i ).
[0011] According to the interpretable academic paper retrieval method provided by the present invention, the keyword retrieval is as follows: Results_keyword(QE i =KeywordSearch(Content(QE) i ),Index_keyword), Keyword retrieval scores are calculated based on the BM25 algorithm or the TF-IDF algorithm. Vector retrieval is: Results_vector(QE i =VectorSearch(Vec(Content(QE)) i ),Index_vector), Semantic retrieval scores are calculated based on vector similarity, where Vec(Content(QE)) i )) = Embedding(Content(QE i )). According to the interpretable academic paper retrieval method provided by the present invention, S5 specifically includes: The semantic ranking model Ranker is used to calculate the semantic similarity score between the query Query and each paper in the candidate paper set CandidatePapers. The expression is: SemanticScore(Paper) j = Ranker(Query, Paper) j (j = 1, 2, ..., |CandidatePapers|), where Paper j Semantic matching can be performed using the paper's title, abstract, or full text; Candidate papers are sorted from highest to lowest according to their semantic similarity score (SemanticScore). The top k papers with the highest scores are selected, and the sorted paper set RankedPapers = {Paper1, Paper2, ...,Paper...} is output. k}, where Paper1 is the highest-scoring paper, Paper k This represents the paper with the k-th highest score, where k is the preset number of returned results.
[0012] According to the interpretable academic paper retrieval method provided by the present invention, S6 specifically includes: Extract each paper from the paper results set. j (j = 1, 2, ..., k) Search scores Score_dim(Paper) across various academic element dimensions j QE i (i = 1, 2, ..., n), identify academic element dimensions with non-zero search scores to form a matching dimension set MatchedDims = {i | Score_dim(Paper j QE iIf )>0}, extract the set of query elements: QueryElements = {QE1, QE2, ..., QE} n The query dimensions and their text content that are not empty in Chinese; Based on the retrieval score Score_dim(Paper) j QE i The query elements (QueryElements) are used to generate an interpretable text explanation (Paper). j (j = 1, 2, ..., k); The sorted set of papers and their corresponding interpretable texts are combined to output the interpretable academic paper retrieval results: FinalResults={(Paper1,Explanation(Paper1)),(Paper2,Explanation(Paper2)),...,(Paper k ,Explanation(Paper k ))}.
[0013] According to the present invention, an interpretable academic paper retrieval method is provided, wherein the interpretable text is Explanation (Paper). j The generation methods for ) (j = 1, 2, ..., k) include: Generate using a large-scale language model (LLM): Construct a prompt word for the paper. j The academic element set (Elements), the query element set (QueryElements), and the retrieval score (Score_dim) for each academic element dimension. j QE i As input, a large language model (LLM) is invoked to generate natural, fluent, and interpretable text (Explanation Paper). j = LLM(Prompt); Generated based on a predefined template: Select the appropriate template (Template(Type(QE)) according to the query feature type in the MatchedDims matching dimension set. i The text content of the query element (Content(QE)) will be displayed. i ), Paper j The text content of academic elements (E) i '), Retrieval score Score_dim(Paper j QE iThe text is filled into placeholders in the template to generate structured, interpretable text.
[0014] Secondly, the present invention provides an interpretable academic paper retrieval device, comprising: The extraction unit defines multiple academic element dimensions, performs semantic parsing on the title and abstract of the paper to be retrieved, extracts from each of the n academic element dimensions to form an academic element set, and combines the title, abstract, and academic element set to output a structured paper representation; n is a positive integer. A unit is established to vectorize the text content of each academic element to obtain a vector representation, and a mapping relationship between academic elements and papers is established based on the text content and vector representation. The query unit is used to receive user queries, perform semantic parsing on the queries, extract elements according to n academic element dimensions, and output a set of query elements. If a certain academic element dimension is not involved in the query, the text content of that academic element dimension will be empty. The retrieval unit is used to traverse the set of query elements, merge the retrieval results of the same academic element dimension, aggregate the retrieval results of all non-empty academic element dimensions, and output a set of candidate papers. The rearrangement unit is used to rearrange the candidate paper set and output the sorted paper result set. The output unit is used to extract the retrieval score and query element set of each paper in the paper result set under each academic element dimension, generate interpretable text based on the retrieval score and query elements, and obtain interpretable academic paper retrieval results.
[0015] This invention has at least the following technical effects: This invention provides an interpretable academic paper retrieval method and apparatus based on multi-dimensional academic element extraction, aiming to address the problems of insufficient semantic understanding, lack of interpretability, and weak fine-grained matching ability in existing academic paper retrieval methods. The method first uses a knowledge extraction model to extract multiple academic element dimensions from paper titles and abstracts, including research object, research method, and research field. Then, it constructs a dual-modal index of text content and vectors for each academic element. User queries are extracted according to the same academic element dimensions. Keyword retrieval and vector retrieval are performed based on each query element, fusing results from the same dimension and aggregating results from across dimensions. A semantic ranking model is used to reorder candidate papers. Finally, interpretability descriptions are generated based on the matching information of each academic element dimension. By introducing multi-dimensional academic element decomposition and a hybrid retrieval mechanism, this invention significantly improves the accuracy, relevance, and interpretability of academic paper retrieval, providing academic researchers with an efficient and transparent literature retrieval tool. Attached Figure Description
[0016] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.
[0017] In the attached diagram: Figure 1 This is a flowchart illustrating the academic paper retrieval method of the present invention. Detailed Implementation
[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.
[0019] The following detailed description of some embodiments of the present invention will be provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.
[0020] This invention presents a method for academic paper retrieval utilizing knowledge extraction technology and a hybrid retrieval strategy. Academic paper retrieval is a crucial means for researchers to obtain relevant literature and understand research frontiers; however, traditional retrieval methods have limitations in understanding complex academic queries, providing accurate matching results, and explaining the basis for retrieval. Academic papers possess highly structured semantic features, containing information across multiple dimensions such as research objects, research methods, research fields, and datasets. With the development of natural language processing technology, particularly the advancements in knowledge extraction models and semantic understanding technology, new technical pathways have been provided for fine-grained retrieval of academic papers. This invention focuses on combining multi-dimensional academic element extraction with hybrid retrieval techniques to construct an interpretable academic paper retrieval method, thereby improving retrieval accuracy, result relevance, and user experience.
[0021] Please see Figure 1 This invention provides an interpretable academic paper retrieval method based on multi-dimensional academic element extraction. Through knowledge extraction technology, the method decomposes papers into multiple academic element dimensions, constructs a bimodal index, implements multi-dimensional hybrid retrieval, re-ranks papers using a semantic ranking model, and generates interpretability descriptions, thereby achieving accurate, efficient, and interpretable academic paper retrieval. The method includes the following steps: S1: Define n academic element dimensions, perform semantic parsing on the title and abstract of the paper to be retrieved, extract each of the n academic element dimensions to form an academic element set; combine the title, abstract and academic element set to output a structured paper representation; n is a positive integer.
[0022] Furthermore, the implementation process in step S1 is as follows: S1.1: Paper Input and Dimension Definition. Obtain the title and abstract of the paper to be retrieved as input text: Input = {Title, Abstract}. Define n academic element dimensions, including but not limited to ResearchObject, ResearchMethod, ResearchField, Dataset / Experimental Object, KeyTechnology, ResearchProblem, ApplicationScenario, and EvaluationMetric.
[0023] S1.2: Academic Element Extraction. The input text is semantically parsed using a first-knowledge extraction model. Extraction is performed on each of the defined n academic element dimensions. This first-knowledge extraction model includes, but is not limited to, large-scale language models, sequence labeling models, information extraction models, and named entity recognition models. The extraction results form an academic element set Elements = {E1, E2, ..., E...} n}, where E i = {Type, Content} (i = 1, 2, ..., n) represents the type Type and the specific text content Content of the i-th academic element dimension. For example, the text content of the research method dimension may be "contrastive learning, loss function optimization", and the text content of the research object dimension may be "image classification task".
[0024] S1.3: Structured Representation Output. The title, abstract, and extracted elements of the paper are combined to output a structured paper representation Paper = {Title, Abstract, Elements}.
[0025] S2: Bimodal Index Construction. The text content of each academic element is vectorized to obtain a vector representation. Based on the text content and vector representation, a mapping relationship between academic elements and papers is established.
[0026] Furthermore, the implementation process in step S2 is as follows: S2.1: Text content storage. For each academic element E in the academic element set Elements. i (i = 1, 2, ..., n), storing its text content Text(E) i = Content(E) i For example, a text index library TextDB can be built: Text(E1), Text(E2), ..., Text(E...). n )}.
[0027] S2.2: Vectorization Encoding. The text content of each academic element is vectorized using the embedding model to obtain the vector representation Vec(E). i = Embedding(Text(E) i (i = 1, 2, ..., n), of course, the embedding model can also include, but is not limited to, pre-trained language models such as BERT, Sentence-BERT, BGE, E5, etc.
[0028] S2.3: Bimodal Index Construction. A keyword index, Index_keyword, is constructed based on text content, supporting keyword retrieval based on algorithms such as BM25 and TF-IDF. A vector index, Index_vector, is constructed based on vector representation, supporting vector retrieval based on metrics such as cosine similarity and Euclidean distance. This vector index employs FAISS and other methods.
[0029] S2.4: Establishing Mapping Relationships. Establish mapping relationships between academic elements and papers, recording each academic element (E). i The Paper and Academic Element Type (E) to which (i = 1, 2, ..., n) belong i It supports reverse locating the complete paper from the retrieved academic elements.
[0030] S3: Query Academic Element Extraction. Receive user queries, perform semantic parsing, extract elements according to the same n academic element dimensions as S1, and output a set of query elements. If a certain academic element dimension is not mentioned in the query, the text content of that academic element dimension will be empty.
[0031] Furthermore, the implementation process in step S3 is as follows: S3.1: Query Input Reception. Receives user-input queries, which can be in the form of natural language questions, keyword combinations, research question descriptions, etc.
[0032] S3.2: Query Element Extraction. The query is semantically parsed using a second knowledge extraction model. Elements are extracted according to the n academic element dimensions defined in step S1. This second knowledge extraction model may be the same as or different from the first knowledge extraction model in S1.2. The extraction results form a query element set QueryElements = {QE1, QE2, ..., QE...}. n}, where QE i = {Type, Content} (i = 1, 2, ..., n) represents the query element type and specific text content of the i-th academic element dimension. For example, the content of the research method dimension may be "contrastive learning", and the content of the research object dimension may be "image classification". If a certain academic element dimension is not involved in the query, the text content Content of that academic element dimension is empty.
[0033] S4: Traverse the set of query elements, merge the search results of the same academic element dimension, aggregate the search results of all non-empty academic element dimensions, and output a set of candidate papers.
[0034] Furthermore, the implementation process in step S4 is as follows: S4.1: Single-dimensional hybrid search. Traverse the query element set QueryElements = {QE1, QE2, ..., QE...} n For each query element whose text content is not empty, QE i (i = 1, 2, ..., n), perform keyword search and vector search respectively: The keyword search term is: Results_keyword(QE) i = KeywordSearch(Content(QE) i (),Index_keyword), calculate keyword retrieval scores based on algorithms such as BM25 and TF-IDF; Vector retrieval is: Results_vector(QE) i ) = VectorSearch(Vec(Content(QE i Vec(Content(QE)), Index_vector), calculates semantic retrieval scores based on vector similarity, where Vec(Content(QE)) i =Embedding(Content(QE)) i )).
[0035] S4.2: Merging results within the same dimension. This involves combining QE results for the same query element. iThe keyword search results and vector search results (i = 1, 2, ..., n) are fused together. The fusion method includes, but is not limited to, weighted fusion. Results(QE) i = λ ×Results_keyword(QE i ) + (1-λ) × Results_vector(QE i The search results for this academic element dimension are obtained by combining reciprocal ranking with Regression Rank (RRF), linear combination, and normalization fusion. i ), where λ is the weighting coefficient, and its value ranges from [0,1].
[0036] S4.3: Cross-dimensional result aggregation. Aggregate search results from academic element dimensions where all content is not empty, deduplicate and merge them using paper identifiers, resulting in a candidate paper set CandidatePapers = Aggregate({Results(QE i ) |Content(QE i ) ≠ , i = 1, 2, ..., n}), each candidate paper records its retrieval score (also known as matching score) across various academic element dimensions. Score_dim(Paper j QE i It supports subsequent sorting and interpretable generation.
[0037] It's important to note that Aggregate performs the function of merging and deduplicating results. For example, suppose a user queries "image classification using contrastive learning," and we extract two dimensions: Dimension 1, QE1 (Research Method) = "Contrastive Learning"; Dimension 2, QE2 (Research Subject) = "Image Classification." In Dimension 1, we find Paper A (0.85 points) and Paper B (0.72 points); in Dimension 2, we find Paper A (0.91 points) and Paper C (0.68 points). The Aggregate operation merges these results: Paper A appears in both dimensions, so it's recorded as {Research Method: 0.85, Research Subject: 0.91}; Paper B appears only in Dimension 1, so it's recorded as {Research Method: 0.72, Research Subject: 0}; Paper C appears only in Dimension 2, so it's recorded as {Research Method: 0, Research Subject: 0.68}. The final candidate paper set contains three papers: A, B, and C. Each paper has its score in each academic dimension, with a score of 0 for academic dimensions where no matching paper was found.
[0038] S5: Reordering based on a semantic ranking model. Reorders the candidate paper set and outputs the sorted paper result set.
[0039] Furthermore, the implementation process in step S5 is as follows: S5.1: Semantic Similarity Calculation. The semantic ranking model Ranker is used to calculate the semantic similarity score between the query and each paper in the candidate paper set CandidatePapers. The semantic ranking model includes, but is not limited to, ranking models based on cross-encoders (such as Cross-Encoder), semantic matching models based on dual-tower structures (such as BGE-reranker), and re-ranking models based on Transformers. The expression for the semantic similarity score is: SemanticScore(Paper... j = Ranker(Query, Paper) j (j = 1, 2, ..., |CandidatePapers|), where Paper j Semantic matching can be performed using the paper's title, abstract, or full text.
[0040] S5.2: Paper Reordering Output. Sort candidate papers from highest to lowest according to their semantic similarity score (SemanticScore), select the top k papers with the highest scores, and output the sorted paper set RankedPapers = {Paper1, Paper2, ..., Paper...}. k}, where Paper1 is the highest-scoring paper, Paper k This represents the paper with the k-th highest score, where k is the preset number of returned results.
[0041] S6: Explainable Content Generation. Extract the search scores and query element sets for each paper in the paper results set under each academic element dimension. Generate explainable text based on the search scores and query elements to obtain explainable academic paper search results.
[0042] Furthermore, the implementation process in step S6 is as follows: S6.1: Matching Information Extraction. For each sorted paper... j (j = 1, 2, ..., k), extract its retrieval score Score_dim(Paper) across each academic element dimension. j QE i (i = 1, 2, ..., n), identify academic element dimensions with non-zero search scores to form a matching dimension set MatchedDims = {i | Score_dim(Paper j QE i If )>0}, extract the set of query elements: QueryElements = {QE1, QE2, ..., QE} nThe query dimensions and their text content that are not empty within the range}.
[0043] S6.2: Interpretable text generation. Based on retrieval score Score_dim(Paper) j QE i The query elements (QueryElements) are used to generate an interpretable text explanation (Paper). j (j = 1, 2, ..., k), and the generation methods include, but are not limited to: Generate using a large language model: Construct a prompt word for the paper. j The academic element set (Elements), the query element set (QueryElements), and the retrieval score (Score_dim) for each academic element dimension. j QE i As input, a large language model (LLM) is invoked to generate natural, fluent, and interpretable text (Explanation Paper). j = LLM(Prompt); Generated based on a predefined template: Select the appropriate template (Template(Type(QE)) according to the query feature type in the MatchedDims matching dimension set. i The text content of the query element (Content(QE)) will be displayed. i ), Paper j The text content of academic elements (E) i '), Retrieval score Score_dim(Paper j QE i Information such as ) is filled into the placeholders in the template to generate structured interpretable text. For example, the template could be "This paper is in {Type(QE) i )} Dimension and query matching: the paper's {Type(QE i )} is {Content(E i ')}, and the query {Content(QE i The search result is related to {Score_dim(Paper)}, and the retrieval score is {Score_dim(Paper)}. j QE i )}".
[0044] S6.3: Final Result Output. The sorted set of papers and their corresponding interpretable texts are combined to output the interpretable academic paper retrieval results as follows: FinalResults={(Paper1,Explanation(Paper1)),(Paper2,Explanation(Paper2)),...,(Paper k ,Explanation(Paper k Each result item contains complete information about the paper (title, abstract, authors, publication information, etc.) and an interpretability description of the paper being retrieved. Users can simultaneously view the retrieved papers and the academic elements that match the query, the specific matching content, and the search score, thereby quickly determining the relevance of the paper.
[0045] Based on the same inventive concept, another embodiment of the present invention provides an interpretable academic paper retrieval device based on multi-dimensional academic element extraction. This device is used to implement the method of the aforementioned embodiment, and includes: The extraction unit defines multiple academic element dimensions, performs semantic parsing on the title and abstract of the paper to be retrieved, extracts from each of the n academic element dimensions to form an academic element set, and combines the title, abstract, and academic element set to output a structured paper representation; n is a positive integer. A unit is established to vectorize the text content of each academic element to obtain a vector representation, and a mapping relationship between academic elements and papers is established based on the text content and vector representation. The query unit is used to receive user queries, perform semantic parsing on the queries, extract elements according to n academic element dimensions, and output a set of query elements. If a certain academic element dimension is not involved in the query, the text content of that academic element dimension will be empty. The retrieval unit is used to traverse the set of query elements, merge the retrieval results of the same academic element dimension, aggregate the retrieval results of all non-empty academic element dimensions, and output a set of candidate papers. The rearrangement unit is used to rearrange the candidate paper set and output the sorted paper result set. The output unit is used to extract the retrieval score and query element set of each paper in the paper result set under each academic element dimension, generate interpretable text based on the retrieval score and query elements, and obtain interpretable academic paper retrieval results.
[0046] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the embodiments disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. It should be understood that the invention is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.
Claims
1. An interpretable academic paper retrieval method, characterized in that, include: S1. Define n academic element dimensions, perform semantic parsing on the title and abstract of the paper to be retrieved, extract each of the n academic element dimensions to form an academic element set; combine the title, abstract and academic element set to output a structured paper representation; n is a positive integer; S2. Vectorize the text content of each academic element to obtain a vector representation, and establish a mapping relationship between academic elements and papers based on the text content and vector representation. S3. Receive user queries, perform semantic parsing on the queries, extract according to n academic element dimensions, and output the query element set. If a certain academic element dimension is not involved in the query, the text content of that academic element dimension is empty. S4. Traverse the set of query elements, merge the search results of the same academic element dimension, aggregate the search results of all non-empty academic element dimensions, and output the set of candidate papers. S5. Reorder the candidate paper set and output the sorted paper result set; S6. Extract the retrieval score and query element set of each paper in the paper result set under each academic element dimension, generate interpretable text based on the retrieval score and query elements, and obtain interpretable academic paper retrieval results.
2. The interpretable academic paper retrieval method according to claim 1, characterized in that, S1 specifically includes: Obtain the title and abstract of the paper to be retrieved, and use them as input text Input = {Title,Abstract}. Define n academic element dimensions, which include research object, research method, research field, dataset / experimental object, key technology, research question, application scenario, and evaluation index. The first knowledge extraction model is used to perform semantic parsing on the input text, extracting elements from each of the defined n academic element dimensions. This first knowledge extraction model includes a large-scale language model, a sequence labeling model, an information extraction model, and a named entity recognition model. The extraction results form an academic element set Elements = {E1, E2, ..., E...} n }, where E i ={Type, Content} (i = 1, 2, ..., n) represents the type and text content of the i-th academic element dimension; The title, abstract, and elements are combined to output a structured paper representation, Paper = {Title, Abstract, Elements}.
3. The interpretable academic paper retrieval method according to claim 2, characterized in that, S2 specifically includes: Store each academic element E in the academic element set Elements i Text(E) of the text content of (i = 1, 2, ..., n) i = Content(E) i ); The text content of each academic element is vectorized and encoded using the embedding model, resulting in a vector representation Vec(E). i = Embedding(Text(E) i (i = 1, 2, ..., n); A keyword index, Index_keyword, is constructed based on text content, and a vector index, Index_vector, is constructed based on vector representation. Establish a mapping relationship between academic elements and papers, recording each academic element (E). i The paper and academic element type (E) to which (i = 1, 2, ..., n) belong. i It supports reverse locating the complete paper from the retrieved academic elements.
4. The interpretable academic paper retrieval method according to claim 3, characterized in that, S3 specifically includes: Receive user input of a query, which includes natural language questions, keyword combinations, and research question descriptions; The second knowledge extraction model is used to perform semantic parsing on the query Query, and extraction is performed according to the n academic element dimensions defined in S1. The second knowledge extraction model may be the same as or different from the first knowledge extraction model. The extraction result forms a query element set QueryElements = {QE1, QE2, ..., QE...} n }, where QE i = {Type, Content} (i = 1, 2, ..., n) represents the query element type and text content of the i-th academic element dimension. If a certain academic element dimension is not involved in the query, the text content of that academic element dimension is empty.
5. The interpretable academic paper retrieval method according to claim 4, characterized in that, S4 specifically includes: Iterate through the set of query elements: QueryElements = {QE1, QE2, ..., QE n For each query element whose text content is not empty, QE i (i = 1, 2, ..., n), perform keyword search and vector search respectively; QE for the same query element i The keyword search results and vector search results (i = 1, 2, ..., n) are fused to obtain the search results for this academic element dimension: Results(QE) i ); Aggregate all search results for academic elements (excluding empty text content), deduplicate and merge them using paper identifiers, resulting in a candidate paper set: CandidatePapers = Aggregate({Results(QE... i )|Content(QE i )≠ The function `Score_dim(Paper, i=1,2,...,n)` records the retrieval score of each candidate paper across various academic elements. j QE i ).
6. The interpretable academic paper retrieval method according to claim 5, characterized in that, The keyword retrieval is as follows: Results_keyword(QE i )=KeywordSearch(Content(QE i ),Index_keyword), Keyword retrieval scores are calculated based on the BM25 algorithm or the TF-IDF algorithm. The vector retrieval is as follows: Results_vector(QE i )=VectorSearch(Vec(Content(QE i )),Index_vector), Semantic retrieval scores are calculated based on vector similarity, where Vec(Content(QE)) i )) = Embedding(Content(QE i )).
7. The interpretable academic paper retrieval method according to claim 5, characterized in that, S5 specifically includes: The semantic ranking model Ranker is used to calculate the semantic similarity score between the query Query and each paper in the candidate paper set CandidatePapers. The expression is: SemanticScore(Paper) j = Ranker(Query, Paper) j (j = 1, 2, ..., |CandidatePapers|), where Paper j Semantic matching can be performed using the paper's title, abstract, or full text; Candidate papers are sorted from highest to lowest according to their semantic similarity score (SemanticScore). The top k papers with the highest scores are selected, and the sorted paper set RankedPapers = {Paper1, Paper2, ...,Paper...} is output. k }, where Paper1 is the highest-scoring paper, Paper k This represents the paper with the k-th highest score, where k is the preset number of returned results.
8. The interpretable academic paper retrieval method according to claim 7, characterized in that, S6 specifically includes: Extract each paper from the paper results set. j (j = 1, 2, ..., k) Search scores Score_dim(Paper) across various academic element dimensions j QE i (i = 1, 2, ..., n), identify academic element dimensions with non-zero search scores to form a matching dimension set MatchedDims = {i | Score_dim(Paper j QE i If ) > 0}, extract the set of query elements: QueryElements = {QE1, QE2, ..., QE} n The query dimensions and their text content that are not empty in Chinese; Based on the retrieval score Score_dim(Paper) j QE i The query elements (QueryElements) are used to generate an interpretable text explanation (Paper). j (j = 1, 2, ..., k); The sorted set of papers and their corresponding interpretable texts are combined to output the interpretable academic paper retrieval results: FinalResults={(Paper1,Explanation(Paper1)),(Paper2,Explanation(Paper2)),...,(Paper k ,Explanation(Paper k ))}。 9. The interpretable academic paper retrieval method according to claim 8, characterized in that, The interpretable text Explanation (Paper) j The generation methods for ) (j = 1, 2, ..., k) include: Generate using a large-scale language model (LLM): Construct a prompt word for the paper. j The academic element set (Elements), the query element set (QueryElements), and the retrieval score (Score_dim) for each academic element dimension. j QE i As input, a large language model (LLM) is invoked to generate natural, fluent, and interpretable text (Explanation Paper). j = LLM(Prompt); Generated based on a predefined template: Select the appropriate template (Template(Type(QE)) according to the query feature type in the MatchedDims matching dimension set. i The text content of the query element (Content(QE)) will be displayed. i ), Paper j The text content of academic elements (E) i '), Retrieval score Score_dim(Paper j QE i The text is filled into placeholders in the template to generate structured, interpretable text.
10. An interpretable academic paper retrieval device, characterized in that, include: The extraction unit defines multiple academic element dimensions, performs semantic parsing on the title and abstract of the paper to be retrieved, extracts from each of the n academic element dimensions to form an academic element set, and combines the title, abstract, and academic element set to output a structured paper representation; n is a positive integer. A unit is established to vectorize the text content of each academic element to obtain a vector representation, and a mapping relationship between academic elements and papers is established based on the text content and vector representation. The query unit is used to receive user queries, perform semantic parsing on the queries, extract elements according to n academic element dimensions, and output a set of query elements. If a certain academic element dimension is not involved in the query, the text content of that academic element dimension will be empty. The retrieval unit is used to traverse the set of query elements, merge the retrieval results of the same academic element dimension, aggregate the retrieval results of all non-empty academic element dimensions, and output a set of candidate papers. The rearrangement unit is used to rearrange the candidate paper set and output the sorted paper result set. The output unit is used to extract the retrieval score and query element set of each paper in the paper result set under each academic element dimension, generate interpretable text based on the retrieval score and query elements, and obtain interpretable academic paper retrieval results.