A method and system for assisting in the writing of biomedical literature abstracts based on semantic analysis

By employing a semantic analysis-based approach and utilizing a pre-trained language model and Transformer architecture to process biomedical literature, this method addresses the issues of document format differences and user needs, generating high-quality abstracts and improving writing efficiency and user experience.

CN120030998BActive Publication Date: 2026-11-17CHANGSHA MEDICAL UNIV
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
CN202411763861.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-12-03
Publication Date
2026-11-17
Estimated Expiration
2044-12-03

AI Technical Summary

Technical Problem

Existing technologies struggle to effectively parse and process the format differences in biomedical literature, are unable to generate abstracts in various styles according to user needs, and have low system performance, resulting in insufficient efficiency and quality in writing literature abstracts.

Method used

We employ a semantic analysis-based approach, utilizing a pre-trained biomedical language model and Transformer architecture, combined with natural language processing techniques and a rule engine, to perform format recognition, deep semantic understanding, summary generation, and user interaction, and optimize the model through self-learning.

Benefits of technology

It improves the efficiency and quality of document abstract writing, can handle various document formats, and generates concise, coherent abstracts that meet user needs, thus enhancing the system's practicality and flexibility.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120030998B_ABST
    Figure CN120030998B_ABST
Patent Text Reader

Abstract

This invention proposes a method and system for assisting in the writing of biomedical literature abstracts based on semantic analysis, belonging to the field of computer vision and natural language processing technology. It includes the following steps: format recognition and content extraction of input biomedical literature; deep semantic understanding of the parsed literature content; processing the literature content based on the Transformer architecture to generate abstracts; providing a user-friendly interface that allows users to input literature, view, and edit generated abstracts; and self-learning, responsible for collecting and analyzing user feedback information to optimize the model. Furthermore, this invention also proposes a corresponding system, including a literature parsing module, a semantic analysis module, an abstract generation module, a user interaction module, and a self-learning module. This invention can deeply understand and analyze literature content to obtain high-quality abstracts, accelerate the abstract generation process, significantly reduce the time and effort required for manual abstract writing, reduce resource consumption, and shorten the generation time.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of computer vision and natural language processing technology, and in particular relates to a method and system for assisting in the writing of biomedical literature abstracts based on semantic analysis. Background Technology

[0002] In the biomedical field, researchers need to read and write a large number of English literature abstracts. This process not only consumes a significant amount of time and energy but also requires researchers to possess a high level of professional knowledge and language expression skills. The quality of the literature abstracts directly affects the dissemination of the literature and the presentation of research results. Therefore, how to write literature abstracts efficiently and accurately has become an important issue for researchers.

[0003] Existing automatic abstract generation technologies primarily rely on rules and statistical methods, making it difficult to handle the technical terminology and complex sentence structures in biomedical literature. These technologies often fail to fully understand the deeper semantics of the literature, resulting in abstracts that lack accuracy and coherence, and fail to effectively convey the core content. Furthermore, existing technologies struggle to adapt to different styles and formats when generating abstracts, leading to insufficient versatility and practicality.

[0004] With the development of natural language processing technology, especially the application of deep learning and pre-trained language models (such as BioBERT), new solutions have been provided for the generation of biomedical literature summaries. By utilizing these advanced language models, it is possible to gain a deeper understanding of the literature content, extract key information and core viewpoints, and thus generate more accurate and coherent literature summaries.

[0005] However, while these technologies have improved the quality of abstract generation to some extent, many challenges remain in practical applications. For example, the content and format of different documents vary significantly, posing challenges for effective parsing and processing; how to generate abstracts in various styles based on user needs; and how to continuously optimize and improve system performance through user feedback. Therefore, there is an urgent need for a semantic analysis-based method and system for assisting in the writing of biomedical literature abstracts to address these issues and improve the efficiency and quality of abstract writing. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for assisting in the writing of biomedical literature abstracts based on semantic analysis, thereby improving the efficiency and quality of literature abstract writing; and to solve the technical problems of existing technologies where different literatures have large differences in content and format, making it impossible to effectively parse and process literatures, generate abstracts in various styles according to user needs, and resulting in low performance of abstract generation systems.

[0007] To solve the above-mentioned technical problems, the specific technical solution of the present invention is as follows:

[0008] A semantic analysis-based method for assisting in the writing of biomedical literature abstracts includes the following steps:

[0009] Step S1: Perform format recognition and content extraction on the input biomedical literature;

[0010] Step S2: Perform deep semantic understanding on the parsed document content;

[0011] Step S3: Process the document content based on the Transformer architecture to generate a concise and coherent abstract;

[0012] Step S4: Provide a user-friendly interface that allows users to input references, view and edit generated abstracts;

[0013] Step S5: Self-learning, responsible for collecting and analyzing user feedback information to optimize the model.

[0014] Further, step S1 includes the following steps:

[0015] Step S11: Identify the file format by the file header information or file extension, and then use the corresponding parsing tool to parse and process the text content;

[0016] Step S12: Using natural language processing technology and a rule engine, extract the title, author, keywords, introduction, methods, results, and discussion sections from the extracted text content;

[0017] Step S13: Perform classification and recognition processing using a deep learning model;

[0018] Step S14: Standardize the terminology of the extracted information to ensure consistent terminology and format.

[0019] Further, step S2 includes the following steps:

[0020] Step S21: Use the SpaCy tool to perform word segmentation, part-of-speech tagging, and named entity recognition preprocessing on the document content;

[0021] Step S22: Perform deep semantic analysis on the document content using a pre-trained biomedical language model;

[0022] Step S23: Construct a topic model, perform topic classification and cluster analysis on the literature, and extract the main topics and related content.

[0023] Further, step S3 includes the following steps:

[0024] Step S31: Select a suitable summary template according to user needs;

[0025] Step S32: Use the encoder of the Transformer model to perform semantic encoding, generate hidden state representations, and generate a summary through the decoder.

[0026] Step S33: Abstract proofreading. Perform grammatical and logical proofreading on the generated abstract to ensure that the abstract content is accurate and coherent.

[0027] Further, step S32 includes the following steps:

[0028] Step S321: Transformer encoder, using the embeddings provided by the pre-trained language model, inputs into the Transformer encoder for semantic encoding.

[0029] The Transformer encoder includes a multi-head self-attention mechanism and a feedforward neural network:

[0030] H = TransformerEncoder(E)

[0031] Here, E is the embedded representation of the document content generated by a pre-trained language model, TransformerEncoder represents the Transformer encoder, and H is the hidden state output by the Transformer encoder. The computation of the Transformer encoder mainly involves the calculation formula for multi-head self-attention, as shown below:

[0032]

[0033] MultiHead(Q,K,V)=Concat(head1,...,head i , ...,head h W o

[0034] Where Attention(Q, K, V) represents the core operation of the attention mechanism, Q represents the query matrix, K represents the key matrix, V represents the value matrix, and d k The key vector dimension is represented by `Q`, `softmax` by `normalized exponential function`, `T` by `transpose`, `MultiHead(Q, K, V)` by `multihead attention`, `Concat` by `concatenation`, and `head` by `head`. i Let W represent the i-th attention matrix, h represent the number of attention heads, and W represent the number of attention heads. O This represents the learnable parameter matrix.

[0035]

[0036] Among them, W i Q It is a learnable parameter matrix.

[0037] The encoder layer of the present invention can be represented as:

[0038] EncLayer(H)=LayerNorm(H+MultiHead(H,H,H))

[0039] EncLayer(H)=LayerNorm(H+FFN(H))

[0040] Where FFN(H) = max(0, HW1+b1)W2+b2 is a feedforward neural network. max represents the ReLU activation function, W1 and W2 represent the weight matrices, b1 and b2 represent the biases, LayerNorm represents the normalization operation of the normalization layer, and EncLayer(H) represents the encoding operation of the encoding layer.

[0041] Step S322: The Transformer decoder takes the encoder's output and passes it to the Transformer decoder to generate a summary. The Transformer decoder also includes multi-head self-attention mechanisms, encoder-decoder attention mechanisms, and feedforward neural networks.

[0042] S = TransformerDecoder(H,T)

[0043] Where H is the encoder's hidden state, T is the digest template, S is the generated digest, and TransformerDecoder represents the Transformer decoder.

[0044] Similar to the encoder layer, the decoder layer of this invention can be represented as:

[0045] DecLayer(H,E)=LayerNorm(H+MultiHead(H,H,H))

[0046] DecLayer(H,E)=LayerNorm(H+Attention(H,E,E))

[0047] DecLayer(H,E)=LayerNorm(H+FFN(H))

[0048] Here, H comes from the decoder, E comes from the encoder, and DecLayer(H, E) represents the decoding operation of the decoding layer.

[0049] Further, step S4 includes the following steps:

[0050] Step S41: Literature input. Users can input literature files by dragging and dropping or uploading. The system supports multiple file formats.

[0051] Step S42: Abstract editing. Users can view, edit, and modify the generated abstracts. The system provides real-time feedback and suggestions.

[0052] Step S43: Summary preview. Users can preview different styles of summary effects and select the most suitable summary version.

[0053] The present invention also provides a biomedical literature abstract writing assistance system based on semantic analysis, for implementing the method described in any one of claims 1-6, including a literature parsing module, a semantic analysis module, an abstract generation module, a user interaction module, and a self-learning module.

[0054] Furthermore, the document parsing module is used to perform format recognition and content extraction on the input biomedical documents;

[0055] The semantic analysis module is used to perform deep semantic understanding on the parsed document content;

[0056] The abstract generation module processes document content based on the Transformer architecture to generate concise and coherent abstracts;

[0057] The user interaction module provides a user-friendly interface that allows users to input references, view and edit generated abstracts;

[0058] The self-learning module is responsible for collecting and analyzing user feedback information to optimize the model.

[0059] Compared to existing technologies, this invention offers the following advantages: The semantic analysis-based biomedical literature abstract writing method and system provided by this invention, through the use of a pre-trained biomedical language model and deep learning technology, enables in-depth understanding and analysis of the literature content during abstract generation, yielding high-quality abstract data. This abstract data is then processed in the abstract generation module. Because this abstract data closely approximates the final required abstract content, the abstract generation process is accelerated, significantly reducing the time and effort required for manual abstract writing, minimizing resource consumption, and shortening the generation time.

[0060] (1) The document parsing module in this invention can efficiently identify and process various document formats (such as PDF, DOCX, TXT) and extract structured information from them. This process utilizes natural language processing technology and a rule engine to ensure the accuracy of key information extracted from the documents (such as titles, authors, keywords, etc.). By using a biomedical terminology dictionary for terminology standardization, the consistency and accuracy of terminology are guaranteed, thereby improving the quality of subsequent semantic analysis and abstract generation.

[0061] (2) The semantic analysis module in this invention utilizes deep learning and pre-trained biomedical language models (such as BioBERT) to perform in-depth semantic understanding of the document content. Through preprocessing operations such as word segmentation, part-of-speech tagging, and named entity recognition, sentence vector representations are generated, and core viewpoints and key information are identified. A topic model (such as LDA) is constructed to perform topic classification and cluster analysis, ensuring the accuracy of the extracted main topics and related content. This module significantly improves the depth and accuracy of understanding the document content, providing a solid foundation for generating high-quality abstracts.

[0062] (3) The abstract generation module in this invention is based on the Transformer architecture and can select appropriate abstract templates (such as concise, detailed, and professional) according to user needs. It then generates concise, coherent, and accurate abstracts through an encoder and decoder. The generated abstracts undergo grammatical and logical proofreading to ensure content coherence and accuracy, effectively conveying the core viewpoints of the document. This module significantly improves the quality and readability of abstracts.

[0063] (4) The user interaction module in this invention provides a user-friendly interface, allowing users to input literature files via drag-and-drop or file upload, supporting multiple formats. Users can preview and edit the generated abstracts in real time and select different abstract styles. The system provides real-time feedback and suggestions to help users generate the most satisfactory abstracts. This module enhances the user experience and improves the system's usability and flexibility.

[0064] (5) The self-learning module continuously optimizes the model by collecting and analyzing user feedback, improving the accuracy of abstract generation and user satisfaction. The system records user editing operations and feedback, and uses this data to optimize the model through supervised learning methods. Based on the latest technological developments and user feedback, the system regularly upgrades its model and methods to ensure its advanced and practical application in processing biomedical literature abstracts. This module enables the system to self-improve, maintaining long-term effectiveness and advancement.

[0065] In summary, this invention, through the synergistic effect of its five modules, significantly improves the efficiency and quality of document abstract writing, solves many problems in the prior art, and has broad application prospects and important practical significance. Attached Figure Description

[0066] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0067] Figure 1 This is a schematic diagram of the semantic analysis-based biomedical literature abstract writing method of the present invention.

[0068] Figure 2 This is a schematic diagram of the document parsing process of the present invention.

[0069] Figure 3 This is a schematic diagram of the semantic analysis process of the present invention.

[0070] Figure 4 This is a schematic diagram of the abstract generation process of the present invention.

[0071] Figure 5 This is a schematic diagram of the user interaction process of the present invention.

[0072] Figure 6 This is a schematic diagram of the self-learning process of the present invention.

[0073] Figure 7 This is a schematic diagram of the semantic analysis-based biomedical literature abstract writing system of the present invention.

[0074] Figure 8 This is a schematic diagram of a UI interface in the embodiment. Detailed Implementation

[0075] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0076] To improve the efficiency and quality of literature abstract writing, this invention proposes a semantic analysis-based method for assisting in the writing of biomedical literature abstracts, such as... Figure 1 As shown, it includes the following steps:

[0077] Step S1: Perform format recognition and content extraction on the input biomedical literature.

[0078] First, the document format (e.g., PDF, DOCX, TXT) is identified by the document header information or file extension, and the text content is extracted using appropriate parsing tools. Second, natural language processing techniques are used to extract structured information from the document, such as the title, authors, keywords, introduction, methods, results, and discussion. Finally, the extracted information undergoes terminology standardization, using biomedical terminology dictionaries (e.g., UMLS, MeSH) for mapping and standardization to ensure consistency in terminology and expression. Figure 2 As shown.

[0079] Step S11: Identify the file format by the file header information or file extension, and then use the corresponding parsing tool to parse and process the text content.

[0080] File extensions are suffixes in filenames used to identify file types, such as .pdf, .docx, and .txt. By checking the file extension, the system can quickly determine the file format and select the appropriate parser or tool to process it. For example: .pdf indicates a PDF document format, which can be extracted using PDF parsing tools. .docx indicates a Microsoft Word document format, which can be parsed using appropriate processing tools. .txt indicates a plain text file, which can be read directly from the text content.

[0081] Sometimes file extensions may be inaccurate or modified. In such cases, the file format can be more reliably identified through the file header information (also known as the file signature or magic number). The file header information consists of specific identifiers in the first few bytes of a file stored, used to indicate the actual type of the file. Different file types typically have unique header identifiers. For example, PDF files usually begin with %PDF. ZIP files typically use the PK identifier (.docx and .xlsx are also based on the ZIP format). JPEG image files begin with FFD8.

[0082] The text content is extracted using the corresponding parsing tool in the following manner:

[0083]

[0084] Where F is the input file, format(F) indicates the return file format, parsePDF indicates parsing a PDF file, parseDOCX(F) indicates parsing a DOCX file, parseTXT indicates parsing a text file, and f format (F) represents the text content extracted after the input document is parsed.

[0085] Step S12: Using natural language processing technology and a rule engine, extract the title, author, keywords, introduction, methods, results, and discussion from the extracted text content.

[0086] Preliminary extraction can be performed using regular expression-based methods, followed by classification and recognition using deep learning models.

[0087] Specifically, the initial extraction steps based on regular expressions are as follows:

[0088] Step S121: Title extraction. Titles are usually located at the top of the document and have a specific format (such as large font, bold). Regular expressions can be used to match the title format:

[0089] Title=match(TitlePatterm,Document)

[0090] Where Title represents the extracted title, match represents the matching function, TitlePattern is the regular expression used to match the title, and Document represents the text content, that is, the text content f extracted from the input document after parsing in step S11. format (F).

[0091] Step S122: Author extraction. The author list usually follows the title and may include a comma-separated list of names. This can be done using regular expressions for matching and extraction.

[0092] Authors=match(AuthorPattern,Doccmemt)

[0093] Here, Authors represents the extracted authors, and AuthorPattern is a regular expression used for the authors.

[0094] Step S123: Keyword extraction. Keywords are usually identified by "keywords" or "Keywords" before the abstract or introduction. Match the keyword section using regular expressions:

[0095] Keywords=match(KeywordsPattern,Document)

[0096] Here, Keywords represents the extracted keywords, and KeywordsPattern is a regular expression used to match the keyword part.

[0097] Step S124: Introduction Extraction. The introduction section is usually identified by "Introduction" or "Introduction" and ends before the "Methods" section. The introduction can be extracted through token matching and text segmentation:

[0098] Introduction=extract(ntroductionmPattern,Document)

[0099] Here, Introduction represents the extracted introduction, extract represents the extraction function, and IntroductionPattern is a marker used to match the introduction section.

[0100] Step S125: Method extraction. The method section is usually identified by "Methods" or "Methods" and ends before the "Results" section. Methods can be extracted through token matching and text segmentation.

[0101] Methods=extract(MethodsPattern,Documemt)

[0102] In this context, Methods represents the extracted methods, and MethodsPattern is a marker used to match the method part.

[0103] Step S126: Results Extraction. The results section is typically labeled "Results" and ends before the "Discussion" section. Results can be extracted using token matching and text segmentation.

[0104] Results=extract(ResultsPattern,Document)

[0105] Here, Results represents the extracted results, and ResultsPattern is a marker used to match the result portion.

[0106] Step S127: Discussion extraction. The discussion section is usually identified by "Discussion" or "Discussion" and ends at the end of the document. Discussion can be extracted using token matching and text segmentation.

[0107] Discussion=etract(DiscussionPattern,Document)

[0108] Here, Discussion represents the extracted discussion, and DiscussionPattern is a marker used to match the discussion section.

[0109] Step S13: Perform classification and recognition processing using a deep learning model, as follows:

[0110] Step S131: Text paragraph classification. Based on the initial extraction, a pre-trained language model is used to classify and identify the document paragraphs.

[0111] Deep learning models can employ pre-trained language models such as BERT and BioBERT. These models can understand the semantics of paragraphs and identify their structured components.

[0112] The pre-trained language model used in this embodiment is the BERT model. Assume the input document content is D, and each paragraph in the document is P. i Each paragraph P is generated using a BERT pre-trained language model. i Convert to embedding vector E i :

[0113] E i =BERT(P i )

[0114] Here, BERT stands for BERT pre-trained language model.

[0115] Step S132: Classification model. A classification model (such as Bidirectional Long Short-Term Memory Network (BiLSTM), Convolutional Neural Network (CNN), etc.) is used to classify the embedded vector E. i The classification process involves determining the section to which a paragraph belongs (e.g., title, author, keywords, introduction, methods, results, and discussion). The output of the classification model is a label L indicating the section to which the paragraph belongs. i :

[0116] L i =Classify(E i )

[0117] Classify is a classification model.

[0118] Step S133: Classification model training. The classification model is trained through supervised learning. The training dataset includes document paragraphs and their corresponding labels. The training objective is to minimize the classification loss function. The classification loss function is expressed as follows:

[0119]

[0120] Where N is the number of training samples, C is the number of label categories, and y ij It's a real label. It is a predicted probability.

[0121] Step S134: Identification Results. Based on the output of the classification model, the paragraphs in the document are classified and labeled to form structured information.

[0122] StructuredDgta={(P i L i |i = 1, 2, ..., n}

[0123] Where StructuredData represents structured information, P i It's a paragraph, L i is the tag of the paragraph, and n represents the total number of tag categories.

[0124] Step S14: Standardize the extracted information to ensure consistent terminology and format. Use biomedical terminology dictionaries (such as UMLS, MeSH, etc.) to map and standardize the terms:

[0125] stamdardize(Term)=map(Term,Dictionary)

[0126] Here, `standardize` is the term standardization function, `map` is the mapping operation, `Term` is the extracted term, and `Dictionary` is the term dictionary.

[0127] Step S2: Perform deep semantic understanding on the parsed document content.

[0128] First, the document content undergoes preprocessing operations such as word segmentation, part-of-speech tagging, and named entity recognition. Next, a pre-trained biomedical language model (such as BioBERT) is used to generate vector representations of each sentence in the document and identify core viewpoints and key information. Finally, a topic model (such as Latent Dirichlet Allocation, LDA) is constructed to perform topic classification and cluster analysis on the document content, extracting main topics and related content, such as... Figure 3 As shown.

[0129] Step S21: Use the SpaCy tool to perform word segmentation, part-of-speech tagging, and named entity recognition preprocessing on the document content. The processing steps are as follows:

[0130] Step S211: Tokenization involves breaking down the input document content D into a sequence of words T. In SpaCy, tokenization is performed based on predefined language models and rules:

[0131] T = SpaCy.tokenize(D)

[0132] Where SpaCy.tokenize represents the tokenization operation of SpaCy, and the result T after tokenization is:

[0133] T = [t1, t2, ... t] i , ...,tn ]

[0134] Among them, t i Let represent the i-th word in the document, and n represent the total number of words.

[0135] Step S212: Part-of-speech tagging involves assigning appropriate part-of-speech labels to each word. Part-of-speech labels can be nouns, verbs, adjectives, etc. Part-of-speech tagging is performed using a pre-trained model from SpaCy:

[0136] POS(t i =SpaCy.pos t ag(t i )

[0137] Where SpaCy.postag represents the part-of-speech tagging operation of SpaCy, POS(t i ) represents the part-of-speech tagging result of the i-th word.

[0138] Step S213: Named entity recognition (NER) identifies entities with specific meanings in text and assigns them corresponding labels. NER is performed using SpaCy's NER model.

[0139] NER(t i =SpaCy.ner(t i )

[0140] Where SpaCy.ner represents the named entity recognition operation of SpaCy, NER(t i () represents the named entity recognition result of the i-th word. The results of named entity recognition of the literature are as follows:

[0141] T NER =[(t1.NER(t1)),...,(t i .NER(t i )),...,(t n .NER(t n ))]

[0142] Step S214: Combine the preprocessing results, integrating the results of word segmentation, part-of-speech tagging, and named entity recognition to form complete information for each word:

[0143] PreprocessedData

[0144] =[(t1.POS(t1),NER(t1)),...,(t i .POS(t i ), NER(t i )),...,(t n .POS(tn ),NER(t n ))]

[0145] The combined preprocessing results can be represented as a structured dataset containing each word, its part-of-speech tag, and named entity tag:

[0146] PreprocessedData={(t i .POS(t i ),NER(t i ))|i=1,2,...,n}

[0147] This dataset provides detailed semantic information for subsequent semantic analysis and summary generation. By utilizing SpaCy for word segmentation, part-of-speech tagging, and named entity recognition, key information in the documents can be effectively extracted, improving the accuracy and quality of summary generation.

[0148] Step S22: Perform deep semantic analysis on the document content using a pre-trained biomedical language model (such as BioBERT). The model calculates the vector representation of each sentence in the document and identifies core viewpoints and key information. The deep semantic analysis of the document content involves the following processing:

[0149] Step S221: Semantic embedding, using a pre-trained language model (such as BERT) to convert each sentence in the literature into a semantic embedding vector, which captures the semantic information of the sentence.

[0150] Suppose that the content of the document is processed into a sentence sequence S = {s1, ..., s2}. i , ..., s m}, where m represents the number of sentences obtained after processing, and each sentence s i Convert to sentence embedding vector E i :

[0151] E i =BioBERT(s) i )

[0152] Among them, E i It is sentence s i The semantic embedding vector, BioBERT represents the BioBERT pre-trained biomedical language model.

[0153] Step S222: Sentence Relationship Recognition. This involves analyzing the relationships between semantic embedding vectors to identify logical and semantic relationships between sentences. This can be achieved by calculating the similarity or distance between sentence embedding vectors. (Sentence s) i and s j The similarity between sim(E) i Ej Cosine similarity can be used to calculate it:

[0154]

[0155] Among them, E i E j Let ||·|| represent the semantic embedding vector, and let ||·|| represent the norm. Based on similarity, a sentence relationship graph can be constructed to represent the correlation between sentences.

[0156] Step S223: Keyword and Topic Extraction. A deep learning model is used to identify keywords and topics in the document. This can be achieved by analyzing the clustering properties of semantic embedding vectors. Assume all sentence embedding vectors in the document are {E1,...,E...} i ,...,E m The data is clustered into several topics, where m represents the number of sentences, and each topic represents a core piece of content. Topic extraction can be performed using clustering algorithms (such as K-means).

[0157] C1, ...,C i , ...,C k =K-meams(E1, ..., E i , ..., E m (k)

[0158] Among them, C i Let i represent the i-th topic, K-means represent the K-means clustering algorithm, and k is the number of topics.

[0159] Step S224: Dependency Parsing. Dependency parsing identifies the internal grammatical structure and dependency relations of a sentence, further enhancing semantic understanding. SpaCy provides efficient dependency parsing tools. Assume sentence s... i The dependency tree is represented as T i Each node v and edge e represents the words of the sentence and the grammatical relationships between them:

[0160] T i =SpaCy.dependency parse(s i )

[0161] Here, SpaCy.dependency parse represents the dependency parsing operation of SpaCy, which can identify the subject-verb-object structure, modification relations, etc. of a sentence.

[0162] Step S225: Integrate the semantic analysis results, combining the results of word segmentation, part-of-speech tagging, named entity recognition, semantic embedding, sentence relation identification, keyword and topic extraction, and dependency parsing to form the deep semantic analysis results of the document. The integrated semantic analysis results are expressed as follows:

[0163] SemamticAnalysis={(s i E i ,sim(E i E j ), C i T i |i = 1, 2, ..., m}

[0164] Among them, s i It's a sentence, E i It is a sentence embedding vector, sim(E i E j ) represents sentence similarity, C i It's the theme, T i It is a dependency tree.

[0165] Deep semantic analysis comprehensively understands the semantic information of documents through multi-level and multi-angle analysis, providing strong support for subsequent abstract generation.

[0166] Step S23: Construct a topic model (such as Latent Dirichlet Allocation, LDA) to perform topic classification and cluster analysis on the literature, and extract the main topics and related content:

[0167]

[0168] Where θ is the document topic distribution, α is the prior of the topic distribution, and z i It's the topic, w i Let β be the word, β be the word-topic distribution, P(θ|α) represent the probability of θ occurring given α, and P(z) be the probability of θ occurring. i |θ) represents z under condition θ i The probability of occurrence, P(w) i |z i ,β) represents in z i w under β conditions i The probability of occurrence.

[0169] Step S3: Process the document content based on the Transformer architecture to generate a concise and coherent abstract.

[0170] First, a suitable summary template (e.g., concise, detailed, or professional) is selected based on user needs. Then, the encoder of the Transformer model performs semantic encoding to generate hidden state representations, and the decoder generates the summary. The generated summary undergoes grammatical and logical proofreading to ensure accuracy. Figure 4 As shown.

[0171] Step S31: Select a suitable abstract template based on user needs. Abstract templates include concise, detailed, and professional types.

[0172]

[0173] Where u represents the template type selected by the user, Template(u) represents the summary template selected according to the user's needs, concise represents concise, detailed represents detailed, professional represents professional, Concise represents concise summary template, Detailed represents detailed summary template, and Professional represents professional summary template.

[0174] Step S32: Use the encoder of the Transformer model to perform semantic encoding, generate hidden state representations, and generate a summary through the decoder.

[0175] The Transformer model for generating the summary includes an encoder and a decoder. The encoder and decoder of this invention will be described below:

[0176] Step S321: Transformer encoder, using the embeddings provided by the pre-trained language model, inputs into the Transformer encoder for semantic encoding.

[0177] The Transformer encoder includes a multi-head self-attention mechanism and a feedforward neural network:

[0178] H = TransformerEncoder(E)

[0179] Here, E is the embedded representation of the document content generated by a pre-trained language model, TransformerEncoder represents the Transformer encoder, and H is the hidden state output by the Transformer encoder. The computation of the Transformer encoder mainly involves the calculation formula for multi-head self-attention, as shown below:

[0180]

[0181] MultiHead(Q,K,V)=Concat(head1,...,head i , ...,head h W o

[0182] Where Attention(Q, K, V) represents the core operation of the attention mechanism, Q represents the query matrix, K represents the key matrix, V represents the value matrix, and d k The key vector dimension is represented by `Q`, `softmax` by `normalized exponential function`, `T` by `transpose`, `MultiHead(Q, K, V)` by `multihead attention`, `Concat` by `concatenation`, and `head` by `head`. i Let W represent the i-th attention matrix, h represent the number of attention heads, and W represent the number of attention heads. O This represents the learnable parameter matrix.

[0183]

[0184] Among them, W i Q , It is a learnable parameter matrix.

[0185] The encoder layer of the present invention can be represented as:

[0186] EncLayer(H)=LayerNorm(H+MultiHead(H,H,H))

[0187] EncLayer(H)=LayerNorm(H+FFN(H))

[0188] Where FFN(H) = max(0, HW1+b1)W2+b2 is a feedforward neural network. max represents the ReLU activation function, W1 and W2 represent the weight matrices, b1 and b2 represent the biases, LayerNorm represents the normalization operation of the normalization layer, and EncLayer(H) represents the encoding operation of the encoding layer.

[0189] Step S322: The Transformer decoder takes the encoder's output and passes it to the Transformer decoder to generate a summary. The Transformer decoder also includes multi-head self-attention mechanisms, encoder-decoder attention mechanisms, and feedforward neural networks.

[0190] S = TransformerDecoder(H,Te)

[0191] Where H is the encoder's hidden state, Te is the digest template, S is the generated digest, and TransformerDecoder represents the Transformer decoder.

[0192] Similar to the encoder layer, the decoder layer of this invention can be represented as:

[0193] DecLayer(H,E)=LayerNorm(H+MultiHead(H,H,H))

[0194] DecLayer(H,E)=LayerNorm(H+Attention(H,E,E))

[0195] DecLayer(H,E)=LayerNorm(H+FFN(H))

[0196] Here, H comes from the decoder, E comes from the encoder, and DecLayer(H, E) represents the decoding operation of the decoding layer.

[0197] Step S33: Abstract proofreading. Perform grammatical and logical proofreading on the generated abstract (such as Grammarly or LanguageTool) to ensure that the abstract content is accurate and coherent.

[0198] This invention uses syntax checking tools and logical consistency checking algorithms:

[0199] Proofread(S)=GrammarCheck(S)+LogicCheck(S)

[0200] Here, Proofread represents the text proofreading operation, S is the generated summary, GrammarCheck is the grammar check, and LogicCheck is the logic check.

[0201] Abstract proofreading is a crucial step in generating high-quality abstracts. By checking grammar, logical consistency, and information accuracy, it ensures that the generated abstract is accurate, coherent, and meets the reader's expectations.

[0202] Step S4: Provide a user-friendly interface that allows users to input references, view, and edit generated abstracts. Users can input reference files via drag-and-drop or file upload; the system supports multiple formats. Users can preview and edit generated abstracts in real time and choose different abstract styles. The system provides real-time feedback and suggestions to help users generate the most satisfactory abstracts, enhancing the user experience, such as... Figure 5 As shown.

[0203] Step S41: Literature input. Users can input literature files by dragging and dropping, uploading, etc. The system supports multiple file formats.

[0204] Step S42: Abstract editing. Users can view, edit, and modify the generated abstracts. The system provides real-time feedback and suggestions.

[0205] Step S43: Summary preview. Users can preview different styles of summary effects and select the most suitable summary version.

[0206] Step S5: Self-learning, responsible for collecting and analyzing user feedback to optimize the model. The system records user editing operations and feedback, and uses this data to optimize the model through supervised learning methods, improving the accuracy of abstract generation and user satisfaction. The system regularly upgrades its model and methods based on the latest technological developments and user feedback to ensure its advanced and practical application in processing biomedical literature abstracts. Figure 6 As shown.

[0207] Step S51: Collect user feedback, record user editing operations and feedback information, including modified content, satisfaction ratings, etc.

[0208] Step S52: Model optimization. Based on collected user feedback, supervised learning methods are used to optimize the model to improve the accuracy of summary generation and user satisfaction.

[0209] Step S53: Method upgrade. Based on the optimized model and the development of new technologies, the method is upgraded regularly to ensure its advanced nature and practicality.

[0210] Furthermore, this invention proposes a biomedical literature abstract writing assistance system based on semantic analysis, including a literature parsing module, a semantic analysis module, an abstract generation module, a user interaction module, and a self-learning module, such as... Figure 7 As shown.

[0211] The document parsing module is used to identify the format and extract the content of input biomedical documents.

[0212] Specifically, the document format (e.g., PDF, DOCX, TXT) is identified by the document header information or file extension, and the text content is extracted using appropriate parsing tools. Secondly, natural language processing techniques are used to extract structured information from the document, such as title, authors, keywords, introduction, methods, results, and discussion. Finally, the extracted information undergoes terminology standardization, using biomedical terminology dictionaries (e.g., UMLS, MeSH) for mapping and standardization to ensure consistency in terminology and expression.

[0213] The semantic analysis module is used to perform deep semantic understanding on the parsed document content.

[0214] Specifically, the semantic analysis module performs preprocessing operations on the document content, including word segmentation, part-of-speech tagging, and named entity recognition. Next, a pre-trained biomedical language model (such as BioBERT) is used to generate vector representations of each sentence in the document and identify core viewpoints and key information. Finally, a topic model (such as Latent Dirichlet Allocation, LDA) is constructed to perform topic classification and cluster analysis on the document content, extracting main topics and related content.

[0215] The abstract generation module processes document content based on the Transformer architecture, generating concise and coherent abstracts.

[0216] Specifically, the summary generation module selects a suitable summary template (such as concise, detailed, or professional) based on user needs. Then, it uses the encoder of the Transformer model to perform semantic encoding, generating hidden state representations, and then uses the decoder to generate the summary. The generated summary undergoes grammatical and logical proofreading to ensure accuracy.

[0217] The user interaction module provides a user-friendly interface that allows users to input references, view and edit generated abstracts.

[0218] like Figure 8 The diagram illustrates a UI interface of this invention. Users can input literature files through drag-and-drop functionality or file upload via the user interaction module. The system supports multiple formats. Users can preview and edit the generated abstracts in real time and select different abstract styles. The system provides real-time feedback and suggestions to help users generate the most satisfactory abstracts, enhancing the user experience.

[0219] The self-learning module is responsible for collecting and analyzing user feedback information to optimize the model.

[0220] The system records user editing actions and feedback, and uses this data to optimize the model through supervised learning methods, improving the accuracy of abstract generation and user satisfaction. The system regularly upgrades its models and methods based on the latest technological developments and user feedback to ensure its advanced and practical application in generating abstracts for biomedical literature.

[0221] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A semantic analysis-based biomedical literature abstract-assisted writing method, characterized in that, Includes the following steps: Step S1: Perform format recognition and content extraction on the input biomedical literature; Step S2: Perform deep semantic understanding on the parsed document content, including the following steps: Step S21: Use the SpaCy tool to perform word segmentation, part-of-speech tagging, and named entity recognition preprocessing on the document content; Step S22: Perform deep semantic analysis on the document content using a pre-trained biomedical language model; Step S23: Construct a topic model, perform topic classification and cluster analysis on the literature, and extract the main topics and related content; Step S3: Process the document content based on the Transformer architecture to generate a concise and coherent abstract; Step S4: Provide a user-friendly interface that allows users to input references, view and edit generated abstracts; Step S5: Self-learning, responsible for collecting and analyzing user feedback information to optimize the model; Step S1 includes the following steps: Step S11: Identify the file format by the file header information or file extension, and then use the corresponding parsing tool to parse and process the text content; Step S12: Using natural language processing technology and a rule engine, extract the title, author, keywords, introduction, methods, results, and discussion sections from the extracted text content; Step S13: Perform classification and recognition processing using a deep learning model; Step S14: Standardize the terminology of the extracted information to ensure consistent terminology and format.

2. The method for assisting in writing biomedical literature abstracts based on semantic analysis according to claim 1, characterized in that, Step S3 includes the following steps: Step S31: Select a suitable summary template according to user needs; Step S32: Use the encoder of the Transformer model to perform semantic encoding, generate hidden state representations, and generate a summary through the decoder; Step S33: Abstract proofreading. Perform grammatical and logical proofreading on the generated abstract to ensure that the abstract content is accurate and coherent.

3. The method for biomedical literature abstracting aided writing based on semantic analysis according to claim 2, characterized in that, Step S32 includes the following steps: Step S321: Transformer encoder, using the embeddings provided by the pre-trained language model, inputs into the Transformer encoder for semantic encoding; The Transformer encoder includes a multi-head self-attention mechanism and a feedforward neural network: H = TransformerEncoder(E) Here, E is the embedded representation of the document content generated by a pre-trained language model, TransformerEncoder represents the Transformer encoder, and H is the hidden state output by the Transformer encoder. The computation of the Transformer encoder involves the calculation formula of multi-head self-attention, as shown below: MultiHead(Q, K, V) = Concat( ) wherein Attention(Q, K, V) represents the core operation of the attention mechanism, Q represents a query matrix, K represents a key matrix, and V represents a value matrix, denotes the dimension of the key vector, softmax denotes a normalized exponential function, T denotes transposition, MultiHead(Q, K, V) denotes multi-head attention, and Concat denotes concatenation, denotes the i-th attention matrix, and h denotes the number of attention heads, denotes a learnable parameter matrix; wherein, 、 、 are learnable parameter matrices; The encoder layer is represented as: EncLayer(H)=LayerNorm(H+MultiHead(H, H, H)) EncLayer(H)=LayerNorm(H+FFN(H)) FFN(H) = max(0, H ) is a feed-forward neural network; max denotes a ReLU activation function, , denotes a weight matrix, , denotes a bias, and LayerNorm denotes a normalization operation of a normalization layer, and EncLayer(H) denotes an encoding operation of an encoding layer. Step S322: Transformer decoder, which passes the encoder's output to the Transformer decoder to generate a summary; the Transformer decoder also includes a multi-head self-attention mechanism, an encoder-decoder attention mechanism, and a feedforward neural network: S = TransformerDecoder(H, T) Where H is the hidden state of the encoder, T is the digest template, S is the generated digest, and TransformerDecoder represents the Transformer decoder; similar to the encoder layer, the decoder layer is represented as: DecLayer(H,E)=LayerNorm(H+MultiHead(H,H,H)) DecLayer(H,E)=LayerNorm(H+Attention(H,E,E)) DecLayer(H,E)=LayerNorm(H+FFN(H)) Here, H comes from the decoder, E comes from the encoder, and DecLayer(H, E) represents the decoding operation of the decoding layer.

4. The method for assisting in writing biomedical literature abstracts based on semantic analysis according to claim 1, characterized in that, Step S4 includes the following steps: Step S41: Literature input. Users can input literature files by dragging and dropping or uploading. The system supports multiple file formats. Step S42: Abstract editing. Users can view, edit, and modify the generated abstracts. The system provides real-time feedback and suggestions. Step S43: Abstract preview. Users can preview abstracts of different styles and select the most suitable abstract version.

5. A semantic analysis-based system for assisting in the writing of biomedical literature abstracts, characterized in that, The method for implementing any one of claims 1-4 includes a document parsing module, a semantic analysis module, a summary generation module, a user interaction module, and a self-learning module; The literature parsing module performs the following methods when identifying the format and extracting the content of input biomedical literature: It identifies the file format by the file header information or file extension, then uses the corresponding parsing tool to extract the text content; it uses natural language processing technology and a rule engine to extract the title, author, keywords, introduction, methods, results, and discussion sections from the extracted text content; and it uses a deep learning model for classification and recognition. The extracted information is processed to standardize terminology to ensure consistent terminology and format. The semantic analysis module is used to perform deep semantic understanding on the parsed document content; The abstract generation module processes document content based on the Transformer architecture to generate concise and coherent abstracts; The user interaction module provides a user-friendly interface that allows users to input references, view and edit generated abstracts; The self-learning module is responsible for collecting and analyzing user feedback information to optimize the model.

6. The semantic analysis-based biomedical literature abstract writing system according to claim 5, characterized in that, The semantic analysis module performs preprocessing operations such as word segmentation, part-of-speech tagging, and named entity recognition on the document content. Then, it uses a pre-trained biomedical language model to generate vector representations of each sentence in the document and identify core viewpoints and key information. Finally, it constructs a topic model to perform topic classification and cluster analysis on the document content and extract the main topics and related content.

7. The semantic analysis-based biomedical literature abstract writing system according to claim 5, characterized in that, The summary generation module selects a suitable summary template based on user needs, then uses the encoder of the Transformer model to perform semantic encoding, generates hidden state representations, and generates a summary through the decoder. The generated summary undergoes grammatical and logical proofreading.

Citation Information

Patent Citations

  • News interpretation method and system based on natural language processing

    CN119025670A