Academic opinion extraction method and system applied to academic literature

Abstract

The present invention provides an academic opinion extraction method and system applied to academic documents, aiming to solve the problem in the prior art that it is difficult to automatically and accurately extract opinion information from academic documents. First, the original data set of the academic documents to be analyzed is obtained and preprocessed for structural normalization; then, a pre-trained opinion feature extraction model is called to extract semantic features from the preprocessed documents; then, opinion detection results are generated for each document through opinion detection processing, including text position marks of the academic opinions to be extracted and an opinion content summary; finally, an academic opinion extraction result including an opinion content summary and corresponding text position marks is generated and output to a target analysis terminal to support academic content analysis, thereby realizing the automatic and accurate extraction of academic opinions, which helps to improve the efficiency and quality of academic research.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method and system for extracting academic opinions applied to academic literature. Background Art

[0002] In the field of academic research, scholarly literature, as a crucial vehicle for knowledge dissemination and innovation, continues to grow in volume and complexity. When conducting academic research, researchers often need to extract key information from vast amounts of literature, particularly academic perspectives, to gain a deeper understanding of the dynamics of their research field and grasp research trends. However, traditional methods for reading and analyzing academic literature rely primarily on manual reading and note-taking. This approach is not only inefficient but also susceptible to subjective factors, resulting in incomplete and inaccurate information extraction.

[0003] With the rapid development of natural language processing and machine learning technologies, although some automated literature analysis tools have emerged, most of these tools focus on surface-level analysis of documents, such as keyword extraction and text classification, but fail to deeply explore the academic viewpoints in the documents and their specific positions in the text. Summary of the Invention

[0004] In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides a method for extracting academic opinions from academic documents, the method comprising:

[0005] Obtaining an original document data set of the academic document to be analyzed, wherein the original document data set includes a plurality of document text units having a chapter structure;

[0006] Preprocessing the original document data set to obtain a preprocessed document data set with a normalized structure, wherein the preprocessed document data set includes document text segments divided into paragraphs and corresponding chapter category tags;

[0007] Calling a pre-trained viewpoint feature extraction model to perform semantic feature extraction on the pre-processed document data set to obtain document semantic features of each document text unit, wherein the document semantic features include semantic relevance between text segments and core argument direction features;

[0008] Performing opinion detection processing on the semantic features of the document using the opinion feature extraction model to generate opinion detection results for each document text unit, wherein the opinion detection results include text position marks of the academic opinions to be extracted and an opinion content summary;

[0009] An academic opinion extraction result including an opinion content summary and corresponding text position marks is generated based on the opinion detection result, and the academic opinion extraction result is output to a target analysis terminal to support academic content analysis operations.

[0010] On the other hand, an embodiment of the present invention also provides an academic opinion extraction system applied to academic literature, including a processor and a machine-readable storage medium, the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0011] Based on the above aspects, the embodiments of the present invention realize the automatic and accurate extraction of opinion information in academic documents by comprehensively applying technical means such as document preprocessing, semantic feature extraction and opinion detection. First, the original document data is preprocessed for structural normalization to provide a unified and standardized text basis for subsequent opinion extraction; then, the pre-trained opinion feature extraction model is called to perform semantic feature extraction on the preprocessed document data to effectively capture the semantic correlation and core argument direction characteristics between text fragments; then, through opinion detection processing, text position marks and opinion content summaries containing the academic opinions to be extracted are accurately generated; finally, academic opinion extraction results containing opinion content summaries and corresponding text position marks are generated to provide scientific researchers with intuitive and comprehensive academic opinion information. Overall, the technical solution of this application significantly improves the automation level and accuracy of academic opinion extraction, helps scientific researchers quickly grasp the core opinions in the document, improves the efficiency and quality of academic research, and promotes the dissemination and innovation of academic knowledge. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 It is a schematic diagram of the execution flow of the academic opinion extraction method applied to academic literature provided by an embodiment of the present invention.

[0013] Figure 2 Schematic diagram of exemplary hardware and software components of an academic opinion extraction system applied to academic literature provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0014] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 This is a flow chart of an academic opinion extraction method applied to academic documents provided by an embodiment of the present invention. The academic opinion extraction method applied to academic documents is introduced in detail below.

[0015] Step S110: obtaining an original document data set of the academic document to be analyzed, wherein the original document data set includes a plurality of document text units having a chapter structure.

[0016] In academic research and analysis, to comprehensively and accurately extract scholarly insights, the first step is to obtain a dataset of the original literature from the academic documents to be analyzed. This literature comes from a wide range of sources, including specialized academic databases and websites of well-known academic journals. For example, in the field of medical research, relevant academic literature can be obtained from sources such as PubMed and the Chinese Biomedical Literature Database. Each article has a specific chapter structure, with common sections such as title, abstract, introduction, methods, results, discussion, and conclusion. The title summarizes the core of the research, the abstract distills the key points, the introduction explains the background and purpose, the methods explain the research approach, the results present the research findings, the discussion analyzes the results, and the conclusion summarizes the research value. The original literature dataset is composed of numerous document text units with this chapter structure.

[0017] When collecting academic literature data for training and analysis, it is imperative to strictly adhere to laws and regulations and ensure that the data collection is legally authorized. For example, when obtaining literature data from professional academic databases, it is important to establish a legal partnership with the database operator. Key information such as the scope, purpose, and duration of data use should be clearly defined through a formal data use agreement. For example, when collaborating with a well-known academic journal database, the agreement will stipulate that the acquired data may only be used for research to extract scholarly insights and may not be used for commercial gain or other unauthorized purposes. Furthermore, the payment of the appropriate fees required by the agreement is required to obtain legal rights to use the data. For literature not included in the database or requiring additional permissions, contact the authors or copyright holders directly to request authorization to use their literature data. During communication, clearly explain the purpose, method, and expected outcomes of the data use to gain the authors' understanding and support. Communication can occur via email, formal letters, or other means, and records of relevant communications should be retained as evidence of authorization. Even for publicly available academic literature data, it is important to carefully review the terms of use and copyright notice. Some public data may have specific usage restrictions, such as requiring proper citation and attribution when used. When collecting and using this data, we strictly abide by its regulations to ensure that the use of data complies with legal requirements.

[0018] Step S120: pre-processing the original document data set to obtain a pre-processed document data set with normalized structure, wherein the pre-processed document data set includes document text segments divided into paragraphs and corresponding chapter category tags.

[0019] Because the original document dataset contains documents from diverse sources and with widely varying formats and structures, it is difficult to analyze them later. Therefore, preprocessing is necessary to standardize the document structure. This preprocessed document dataset contains document text segments divided into paragraphs and corresponding chapter category tags, which helps clarify the chapters to which the text segments belong within the document.

[0020] Step S121: performing format unification processing on each document text unit in the original document data set to eliminate the typesetting differences between different documents and obtain initial document texts with consistent format.

[0021] Different documents come in various file formats, including plain text, rich text, and PDF. Different methods should be used to eliminate these differences in layout to obtain a consistent format for the initial document.

[0022] Step S1211: Identify the file format type of the document text unit in the original document data set, where the file format type includes at least one of plain text format, rich text format, and PDF format.

[0023] First, we need to identify the file format type of the document text units in the original document dataset. This can be determined by the file extension: for example, files ending in .txt are plain text, .docx are likely rich text, and .pdf are PDF. We can also perform more accurate identification by reading the file header information, ensuring the correct handling of documents in different formats.

[0024] Step S1212: Mark-clearing processing is performed on the document text unit in rich text format, removing the formatting marks of bold, italics, and underline, and retaining the text content.

[0025] For rich text document units, the formatting tags such as bold, italic, and underline are mainly used for formatting when displaying the document, and have no actual semantic value for extracting academic opinions. Parsing tools can be used to parse the markup language of rich text files to identify and remove these formatting tags. For example, for rich text files based on HTML format, regular expressions can be used to match and remove 、 、 etc. tags to obtain pure text content.

[0026] Step S1213: Perform text extraction processing on the document text unit in PDF format, identify the text content in the scanned PDF version through optical character recognition technology, and generate an editable text string.

[0027] If the document text unit is in PDF format, and it is a scanned PDF, the text in it exists in the form of an image and cannot be processed directly. In this case, you need to use optical character recognition (OCR) technology. Select an appropriate OCR engine, such as TesseractOCR, and adjust the relevant parameters based on the document language type, resolution, etc. to recognize the text in the image and generate an editable text string.

[0028] Step S1214: performing line break normalization processing on the extracted text string, merging consecutive line breaks into a single line break, and eliminating text breaks caused by forced line breaks within a paragraph.

[0029] The extracted text string may contain too many consecutive line breaks, causing fragmented text within the paragraph and hindering semantic understanding. Programmatically traversing the text string and combining consecutive line breaks into a single line break can improve the paragraph structure of the text. For example, in Python, you can use the string replace method to replace multiple consecutive line breaks with a single line break.

[0030] Step S1215: Perform unified encoding format processing on all document text units, and convert texts in different encoding formats into a unified Unicode encoding format.

[0031] Different documents may use different encoding formats, such as UTF-8 and GBK. To avoid encoding errors in subsequent processing, it is necessary to unify the encoding format of all document text units into Unicode. This can be achieved using the encoding conversion functions in programming languages. For example, in Python, use the encode() and decode() methods for encoding conversion.

[0032] Step S1216: Perform content alignment processing on the processed document text units of each format to generate an initial document text with consistent format.

[0033] The document text units of different formats that have undergone the above processing are aligned in content, including unified font, font size, line spacing and other format settings, so that all documents have a unified style in appearance, which is convenient for subsequent processing and analysis, and finally generates an initial document text with consistent format.

[0034] Step S122: Perform chapter identification processing on the initial document text, extract the text content of the document title, abstract, introduction, methods, results, discussion, and conclusion parts, and determine the chapter boundaries of each part.

[0035] In order to accurately extract the text content of each important chapter in the document, it is necessary to perform chapter recognition processing on the initial document text and determine the chapter boundaries of each part.

[0036] Step S1221: Identify the first line of text content of the initial document text, extract the text containing the title keywords as the document title, and record the starting position and ending position of the title.

[0037] Using preset title keyword rules, the first line of the original document text is identified. Title keywords are typically words that summarize the core content of the document. Regular expressions are used to match the first line of text for these keywords. If a match is found, the text is identified as the document title, and its starting and ending character positions in the original document text are recorded.

[0038] Step S1222: Search for a paragraph containing the summary keyword in the text after the title, determine the starting position of the summary part, and extract the continuous text to the position where the next chapter keyword appears as the end position of the summary part.

[0039] After determining the title, search the text following the title for a paragraph containing the abstract keyword (e.g., "Abstract") to determine the starting point of the abstract. Continue extracting text until you encounter the next section keyword (e.g., "Introduction"), marking that point as the end of the abstract.

[0040] Step S1223: Search the text after the abstract for paragraphs containing the introduction keyword, method keyword, result keyword, discussion keyword, and conclusion keyword, and determine the starting position of each chapter.

[0041] In the text after the abstract, use regular expressions to search for paragraphs containing introduction keywords (such as "introduction", "research background"), method keywords (such as "method", "experimental design"), result keywords (such as "results", "experimental results"), discussion keywords (such as "discussion", "analysis"), and conclusion keywords (such as "conclusion", "research conclusion"), so as to determine the starting position of each chapter.

[0042] Step S1224: perform boundary verification processing on the text content after each chapter keyword to check whether the subsequent text contains feature words related to the chapter theme, and the feature words include at least one of research background, experimental design, and data analysis.

[0043] To ensure the accuracy of chapter divisions, verify the boundaries of the text following each chapter's keywords. Check whether the subsequent text contains key words related to the chapter's theme. For example, in the introduction, check for key words such as "research status" and "problem raised." In the methods section, check for key words such as "experimental method" and "sample collection." By determining whether these key words are present in the subsequent text, we can verify the rationality of the chapter boundaries.

[0044] Step S1225: Record the starting character position and ending character position of each chapter, and generate a chapter boundary marker set.

[0045] After completing the chapter start position determination and boundary verification, the start character position and end character position of each chapter are recorded, thereby organizing the start character position and end character position of each chapter into a chapter boundary mark set.

[0046] Step S1226: extract the text content of each chapter according to the chapter boundary marker set, and generate a chapter content set including the document title, abstract, introduction, methods, results, discussion, and conclusion.

[0047] According to the chapter boundary marker set, the text content of each chapter is extracted from the initial document text, and these contents are combined together to generate a chapter content set including the document title, abstract, introduction, methods, results, discussion, and conclusion.

[0048] Step S123: performing paragraph segmentation processing on the initial document text based on the chapter boundaries, dividing the continuous text in each chapter into multiple paragraph units, and obtaining document text segments divided by paragraphs.

[0049] After determining the boundaries and content of each chapter, the continuous text within each chapter is segmented into paragraphs based on features such as line breaks and punctuation. A new paragraph is identified when a line break is encountered and the next line of text is semantically independent of the previous line, or when punctuation marks such as periods, exclamation points, and question marks are used to indicate the end of a sentence. Through paragraph segmentation, the chapter text is divided into multiple paragraph units, resulting in document fragments divided by paragraphs. These fragments serve as the basic text units for further analysis.

[0050] Step S124: punctuation standardization is performed on each paragraph unit, the full-width and half-width symbol formats are unified, and missing punctuation is corrected to generate a paragraph text with standardized punctuation.

[0051] Due to different document sources and input methods, punctuation within paragraphs may be inconsistent or missing. To address this issue, full-width and half-width punctuation are converted to half-width punctuation through character encoding conversion. For missing punctuation, it is assessed and supplemented based on the text's semantics and grammatical rules. For example, in a longer sentence, periods are added at appropriate locations based on the sentence's logical structure and intended meaning. Punctuation standardization generates standardized paragraph text, making the text's semantic expression clearer and more accurate.

[0052] Step S125: performing category tagging processing on the chapter contents obtained by the chapter identification processing, and adding a corresponding chapter category tag to each paragraph unit, wherein the chapter category tag includes at least one of the title category, abstract category, method category, and conclusion category.

[0053] After completing paragraph segmentation and punctuation standardization, add corresponding chapter category tags to each paragraph unit based on the chapter contents obtained from the previous chapter identification process. Paragraph units in the title chapter are marked as title category, paragraph units in the abstract chapter are marked as abstract category, paragraph units in the method chapter are marked as method category, and paragraph units in the conclusion chapter are marked as conclusion category. For paragraph units in chapters such as introduction, results, and discussion, corresponding category tags can also be added according to specific research needs and analysis purposes, such as introduction-background category, introduction-purpose category, etc. By adding chapter category tags, the chapter affiliation and semantic function of each paragraph unit in the document can be clarified.

[0054] Step S126: integrating the document text segments divided into paragraphs and the corresponding chapter category tags to generate a pre-processed document data set with normalized structure.

[0055] After completing paragraph segmentation, punctuation standardization, and chapter category tagging, the document text segments divided by paragraphs and the corresponding chapter category tags are integrated. This can be done using a dictionary data structure, with paragraph text as the key and chapter category tags as the value, stored in a dictionary. Alternatively, a database table can be used, with paragraph text and chapter category tags as separate fields, with each record corresponding to a paragraph unit. Through data integration, all paragraph text segments and chapter category tags are combined to generate a structurally standardized preprocessed document data set, facilitating subsequent semantic feature extraction and opinion detection operations.

[0056] Step S130: calling the pre-trained viewpoint feature extraction model to perform semantic feature extraction processing on the pre-processed document data set to obtain the document semantic features of each document text unit, wherein the document semantic features include the semantic relevance between text segments and the core argument direction features.

[0057] After obtaining a structurally normalized preprocessed document dataset, we apply a pretrained opinion feature extraction model to extract semantic features. This pretrained opinion feature extraction model, trained on a large dataset of academic documents, effectively captures the semantic information within the text. By processing the preprocessed document dataset with this model, we obtain the semantic features of each document unit, including the semantic relevance between text segments and the core argument direction.

[0058] Step S131: inputting the document text segments in the pre-processed document data set into the text encoding layer of the viewpoint feature extraction model, performing lemma processing on each paragraph unit and generating a lemma embedding vector.

[0059] The document text fragments in the preprocessed document dataset are input into the text encoding layer of the opinion feature extraction model, where each paragraph unit is tokenized and a token embedding vector is generated.

[0060] Step S1311: performing word segmentation processing on each paragraph unit, dividing the continuous text into word units with semantic meanings, wherein the word units include Chinese words and English words.

[0061] For Chinese paragraph units, use word segmentation tools such as Jieba to segment the text into appropriate words according to Chinese language rules; for English paragraph units, use the split() method in Python to segment the text into words based on spaces, and segment continuous text into word units with semantic meaning.

[0062] Step S1312: performing lowercase conversion processing on the word units after word segmentation to unify the uppercase and lowercase formats of English words.

[0063] In order to eliminate the influence of the case of English words on subsequent processing, the English words in the word unit after word segmentation are converted to lowercase, and all English words are converted to lowercase format to make the representation of English words more unified.

[0064] Step S1313: removing meaningless symbols in word units, wherein the meaningless symbols include at least one of special punctuation marks, mathematical symbols, and hyperlinks, and retaining word units with semantic information.

[0065] Tokens may contain meaningless symbols such as special punctuation, mathematical symbols, and hyperlinks. These symbols are not helpful for semantic understanding and need to be removed. You can use regular expressions to match these meaningless symbols and remove them, retaining only tokens with semantic information.

[0066] Step S1314: input the processed word unit into the vocabulary mapping module of the text encoding layer, and map each word unit to a corresponding word index according to a predefined vocabulary.

[0067] The predefined vocabulary contains a large number of common tokens. The processed token units are input into the vocabulary mapping module of the text encoding layer. The vocabulary is used to find the token index corresponding to each token unit. If the token unit exists in the vocabulary, it is mapped to the corresponding index; if it does not exist, a special index (such as the unknown word index) can be used to represent it.

[0068] Step S1315: Call the embedding matrix of the text encoding layer to perform vector conversion processing on the word unit index to generate a word unit embedding vector for each word unit, and the dimension of the word unit embedding vector is consistent with the number of columns of the embedding matrix.

[0069] The embedding matrix of the text encoding layer is a pre-trained matrix with the number of rows equal to the size of the vocabulary and the number of columns representing the dimensionality of the word-unit embedding vector. This embedding matrix is called, and the word-unit index is used as the matrix row index. The corresponding row vector is extracted from the embedding matrix, which is the word-unit embedding vector for each word unit. The dimensionality of the word-unit embedding vector is the same as the number of columns in the embedding matrix.

[0070] Step S1316: Perform position encoding processing on the word unit embedding vector of each paragraph unit, add a position encoding vector representing its position in the paragraph to each word unit embedding vector, and generate a word unit embedding vector set containing position information.

[0071] To enable the model to perceive the position of a word in a paragraph, we perform position encoding on the word embedding vector of each paragraph unit. Based on the word's position in the paragraph, we generate a corresponding position encoding vector. This position encoding vector is then added to the word embedding vector to produce a set of word embedding vectors containing the position information.

[0072] Step S132: performing contextual modeling processing on the word embedding vector through the self-attention mechanism module of the text encoding layer to generate a local context feature vector containing the semantic relationship between words in the paragraph.

[0073] After obtaining a set of word-memory embedding vectors containing positional information, the self-attention mechanism module in the text encoding layer performs contextual modeling on these embedding vectors. The self-attention mechanism calculates the correlation between word-memory elements and generates a local context feature vector that contains the semantic relationships between words within a paragraph by performing operations such as weighted summation on the word-memory embedding vectors. This local context feature vector reflects the semantic relationships between words within a paragraph.

[0074] Step S133: Input the local context feature vector into the cross-paragraph association layer of the viewpoint feature extraction model, analyze the semantic coherence between different paragraph units, and calculate the semantic overlap and topic continuity parameters of adjacent paragraphs.

[0075] The local context feature vector is input into the cross-paragraph association layer of the opinion feature extraction model. This layer analyzes the semantic coherence between different paragraph units. By comparing the local context feature vectors of adjacent paragraphs, the semantic overlap and topic continuity parameters are calculated. Semantic overlap can be measured by calculating the similarity between vectors (such as cosine similarity). The topic continuity parameter is determined based on the occurrence and distribution of topic keywords within a paragraph to determine the degree of semantic connection between adjacent paragraphs.

[0076] Step S134: constructing a semantic association model between paragraphs based on the semantic overlap and topic continuity parameters, and generating semantic association descriptors between text segments, wherein the semantic association descriptors are used to represent the strength of the logical cohesion of paragraph units.

[0077] Based on the calculated semantic overlap and topic continuity parameters, a semantic association model between paragraphs is constructed. This model generates semantic association descriptors between text segments by weightedly combining semantic overlap and topic continuity parameters. The semantic association descriptor is a multidimensional vector that represents the strength of the logical cohesion of a paragraph unit.

[0078] Step S135: Jointly analyze and process the local context feature vector and the semantic correlation descriptor between paragraphs, extract the core discussion topics repeatedly emphasized in the document text unit, and generate core discussion direction features. The core discussion direction features include a topic keyword set and a topic occurrence frequency parameter.

[0079] The local context feature vector and the inter-paragraph semantic relevance descriptor are jointly analyzed and processed. By performing topic mining on the local context feature vector and combining it with the inter-paragraph semantic relevance descriptor to determine the continuity of topics across different paragraphs, the core thematic themes repeatedly emphasized within the document text unit are extracted. The core thematic direction features include a thematic keyword set and a thematic frequency parameter. The thematic keyword set represents the key expressions of the core thematic themes, while the thematic frequency parameter reflects the frequency of each topic in the document. This information can be used to identify the core thematic direction of the document.

[0080] Step S136: performing feature fusion processing on the semantic relevance descriptors and core argument direction features between the text segments to generate document semantic features of each document text unit.

[0081] The semantic association descriptors between text segments and the core argument direction features are fused. These two features can be combined using a splicing method to generate a document semantic feature for each document unit. The document semantic feature is a comprehensive feature vector that contains the semantic association information between text segments and the core argument direction information.

[0082] Step S140: performing opinion detection processing on the semantic features of the document through the opinion feature extraction model to generate opinion detection results for each document text unit, wherein the opinion detection results include text position marks of the academic opinions to be extracted and an opinion content summary.

[0083] After obtaining the semantic features of each document's text unit, we then use the opinion feature extraction model to perform opinion detection, generating an opinion detection result. This result includes the text location markers of the academic opinion to be extracted and a summary of the opinion's content. Through in-depth analysis of the document's semantic features, we can accurately locate the academic opinion within the document and extract its core content.

[0084] Step S141: inputting the document semantic features into the viewpoint positioning layer of the viewpoint feature extraction model, analyzing the distribution density of the subject keyword set in the core argument direction feature in the paragraph unit, and identifying candidate paragraphs where the subject keywords are clustered.

[0085] The semantic features of the document are input into the opinion positioning layer of the opinion feature extraction model. The main task of this layer is to analyze the distribution density of the theme keyword set in the core argument direction feature in the paragraph unit. The theme keyword set reflects the core argument theme of the document, and its distribution density in the paragraph can reflect the degree of relevance between the paragraph and the core argument. The distribution density is determined by calculating the frequency of the theme keywords in each paragraph. When the distribution density of theme keywords in a paragraph is high, it means that the paragraph may contain important academic viewpoints and it will be identified as a candidate paragraph. For example, in a document on climate change, if theme keywords such as "greenhouse gas emissions" and "global temperature rise" appear frequently in a paragraph, then the paragraph may be a candidate paragraph.

[0086] Step S142: Perform sentiment analysis on the local context feature vector of the candidate paragraphs, extract sentiment parameters representing affirmative, negative or innovative arguments, and select target paragraphs whose sentiment parameters exceed a preset threshold.

[0087] Sentiment analysis is performed on the local context feature vectors of the candidate paragraphs to extract sentiment parameters that indicate affirmative, negative, or innovative arguments. Sentiment parameters reflect the emotional attitude of the arguments in the paragraph and are crucial for determining the nature of academic viewpoints.

[0088] Step S1421: inputting the local context feature vector of the candidate paragraph into the sentiment analysis submodule of the opinion feature extraction model, wherein the sentiment analysis submodule includes a positive classifier, a negative classifier, and an innovative classifier.

[0089] The local context feature vector of the candidate paragraph is input into the sentiment analysis submodule of the opinion feature extraction model. This submodule contains a positive classifier, a negative classifier, and an innovative classifier. These three classifiers are used to determine whether the paragraph's sentiment is positive, negative, or innovative, respectively. Each classifier is trained with a large amount of data and can accurately classify the local context feature vector input.

[0090] Step S1422: The local context feature vector is classified by the affirmative classifier to generate an affirmative probability value indicating that the paragraph content holds a positive attitude towards a certain proposition.

[0091] The affirmativeness classifier classifies the local context feature vector and, based on its internal classification rules and training model, generates a probability value indicating whether the paragraph content affirms a particular claim. This probability value reflects the likelihood that the paragraph expresses a positive view. For example, in a paragraph discussing the effectiveness of a treatment, the affirmativeness classifier analyzes the local context feature vector to determine the probability that the paragraph expresses a positive view of the treatment.

[0092] Step S1423: Classify the local context feature vector using the negativity classifier to generate a negativity probability value indicating that the paragraph content has a negative attitude towards a certain proposition.

[0093] The negativity classifier similarly processes the local context feature vectors, generating a negativity probability value indicating that the paragraph's content disapproves of a particular proposition. This probability reflects the likelihood that a paragraph expresses a disapproving viewpoint. For example, in a document evaluating the effectiveness of a policy, the negativity classifier analyzes the local context feature vectors of the relevant paragraphs to determine the probability that they disapprove of the policy.

[0094] Step S1424: The local context feature vector is classified by the innovation classifier to generate an innovation probability value indicating that the paragraph content proposes a new theory or new method.

[0095] The Innovation Classifier processes the local context feature vector to generate a probability value indicating whether the paragraph's content proposes a new theory or method. This probability value is used to measure the paragraph's degree of innovation. In some cutting-edge research literature, the Innovation Classifier can help identify paragraphs that propose novel ideas or methods.

[0096] Step S1425: Merge the positive probability value, negative probability value and innovative probability value into a sentiment tendency parameter set.

[0097] The positive probability value, the negative probability value and the innovative probability value are combined into a sentiment tendency parameter set, which comprehensively reflects the sentiment tendency characteristics of the candidate paragraph.

[0098] Step S1426: Filter candidate paragraphs in the sentiment tendency parameter set whose probability value exceeds a preset threshold as target paragraphs, where the preset threshold is used to distinguish between arguments with clear sentiment tendencies and neutral descriptions.

[0099] A preset threshold is set to distinguish between arguments with clear sentiment and neutral descriptions. Candidate paragraphs with at least one probability value exceeding the preset threshold in the sentiment parameter set are selected as target paragraphs. For example, if the probability value for affirmation exceeds the preset threshold, it indicates that the paragraph has a strong positive attitude towards a certain proposition; if the probability value for innovation exceeds the preset threshold, it indicates that the paragraph may propose a new theory or method. This screening method can focus on paragraphs that express clear opinions.

[0100] Step S143: In the target paragraph, locate key nodes of argument logic based on semantic association descriptors between text segments. The key nodes include text locations where claims are made, evidence is provided, and conclusions are drawn.

[0101] In the target paragraph, the key nodes of the argument logic are located based on the semantic relevance descriptors between text segments. The semantic relevance descriptor reflects the strength of the argument logic of the paragraph unit. By analyzing its value and changing trend, the text locations of key argument links such as making claims, providing evidence, and drawing conclusions can be found. For example, when the semantic relevance descriptor shows that the connection between two text segments is close and there is a clear causal relationship, it may indicate the location of a claim and corresponding evidence; when a text segment is closely related to the previous content and has a summary nature, it may be the location of the conclusion.

[0102] Step S144: performing summary generation processing on the text content of the key nodes, extracting short sentences containing core propositions or conclusions as a summary of the viewpoint content.

[0103] The text content at key nodes is summarized to extract short sentences containing the core proposition or conclusion as a summary of the viewpoint. The summary generation process considers the semantic information and importance of the text, removing redundant information and retaining the most core content. For example, in the text at a key node proposing a theory, the summary generation process will extract the core statement of the theory; in the text at a key node reaching a conclusion, the summary sentence can be extracted to form a summary of the viewpoint.

[0104] Step S145: Record the starting character position and the ending character position of the key node in the document text unit, and generate a corresponding text position mark.

[0105] Record the starting and ending character positions of key nodes within a document's text unit to generate corresponding text position markers. Text position markers accurately identify the specific location of a point within the document, making it easier to review and reference it later. By recording the starting and ending character positions, you can precisely locate the text scope of key nodes, facilitating analysis and organization of the points.

[0106] Step S146: performing an association matching process on the opinion content summary and the corresponding text position mark to generate an opinion detection result for each document text unit.

[0107] The opinion summary is then matched with the corresponding text location markers, ensuring that each opinion summary corresponds to a precise text location marker. This matching process generates opinion detection results for each document text unit. The opinion detection results present the text location and core content of the academic opinion to be extracted in a structured manner.

[0108] Step S150: Generate an academic opinion extraction result including an opinion content summary and corresponding text position marks based on the opinion detection result, and output the academic opinion extraction result to a target analysis terminal to support academic content analysis operations.

[0109] After obtaining the opinion detection results of each document text unit, it is necessary to generate academic opinion extraction results based on these results, which include an opinion content summary and corresponding text position marks, and output them to the target analysis terminal to support various academic content analysis operations.

[0110] Step S151: performing redundancy elimination processing on the opinion content summaries in the opinion detection results, merging opinion contents expressing the same or similar propositions, and generating a compressed opinion content set.

[0111] The opinion content summaries in the opinion detection results are processed for redundancy elimination. Since there may be multiple opinion content summaries expressing the same or similar propositions in the literature, these redundant information will increase the complexity of subsequent analysis, so they need to be merged.

[0112] For example, step S1511: calculating the semantic similarity between any two opinion contents in the opinion content summary, wherein the semantic similarity is determined by comparing the keyword overlap rate of the opinion contents and the matching degree of the core argument direction features.

[0113] The semantic similarity between any two opinion pieces in the opinion content summary is calculated by comparing the keyword overlap rate of the opinion pieces and the matching degree of the core argument direction features. The keyword overlap rate can be determined by counting the ratio of the number of identical keywords in the two opinion pieces to the total number of keywords. The matching degree of the core argument direction features is determined by comprehensively judging the theme keyword set and theme frequency parameters in the core argument direction features. For example, if two opinion pieces share most of the same keywords and their core argument direction features also highly match, then their semantic similarity is high.

[0114] Step S1512: constructing a similarity matrix of opinion contents, where the elements in the similarity matrix represent the semantic similarity between two opinion contents.

[0115] Based on the calculated semantic similarity, a similarity matrix for the opinion content is constructed. The similarity matrix is a two-dimensional matrix, where each element represents the degree of semantic similarity between two opinion contents. The rows and columns of the matrix correspond to different opinion contents. The larger the value of an element, the more semantically similar the two corresponding opinion contents are.

[0116] Step S1513: performing a hierarchical clustering analysis based on the similarity matrix, and dividing the viewpoint contents whose semantic similarity exceeds a preset similarity threshold into the same cluster group.

[0117] Hierarchical clustering analysis is performed based on a similarity matrix. A threshold is set, and opinions with semantic similarity exceeding this threshold are grouped together. Hierarchical clustering analysis gradually merges similar opinions into clusters based on the values of the elements in the similarity matrix, forming a hierarchical clustering structure. This approach allows opinions expressing the same or similar views to be clustered together.

[0118] Step S1514: performing content fusion processing on the opinion contents in each cluster group, and extracting the common core propositions of each opinion content as the merged opinion content.

[0119] The idea content within each cluster is fused, analyzing the semantic information of each cluster's ideas and extracting their shared core propositions as the merged idea content. This process comprehensively considers the presentation and emphasis of each idea, removing any differences and retaining the shared core. For example, within a cluster about a specific technology application, the shared statements regarding the advantages of that technology are extracted from each idea to form the merged idea content.

[0120] Step S1515: Record the number of original opinion content corresponding to each merged opinion content, and generate an opinion content record containing merge information.

[0121] The number of original viewpoints corresponding to each merged viewpoint is recorded and saved together with the merged viewpoint to generate a viewpoint record containing merge information. This record can reflect information about the merge process, such as how many original viewpoints were merged from a merged viewpoint.

[0122] Step S1516: sorting the merged opinion content and the corresponding merge information to generate a compressed opinion content set, wherein the semantic similarity of different opinion contents in the compressed opinion content set is lower than the preset similarity threshold.

[0123] The merged opinion content and the corresponding merged information are organized to form a compressed opinion content set. The semantic similarity of different opinion contents in this set is lower than the preset similarity threshold, which means that the opinion contents in the set are highly independent and diverse, removing redundant information and making the opinion content more refined.

[0124] Step S152: performing topic grouping processing on the opinion content set according to the topic relevance of each opinion content in the opinion content set, and generating a plurality of topic-related opinion subsets.

[0125] The compressed opinion content is grouped based on thematic relevance of each opinion piece. Thematic keywords and core argumentation of each opinion piece are analyzed, and thematically related opinion pieces are grouped into the same opinion subset. For example, in a paper on energy research, opinion pieces related to solar energy utilization are grouped into one subset, while those related to wind energy utilization are grouped into another. This creates multiple thematically related opinion subsets, facilitating subsequent in-depth analysis of opinion pieces on different topics.

[0126] Step S153: extracting the text position marks corresponding to the content summary of each opinion subset, counting the distribution of each text position mark in the document text unit, and generating an opinion distribution density map.

[0127] Extract the text location markers corresponding to the summary of each viewpoint subset and count the distribution of these text location markers within the document's text units. By analyzing the distribution of text location markers, we can understand the density of different viewpoints within the document. For example, we can count the number of text location markers that contain a particular viewpoint subset in each chapter and generate a viewpoint distribution density map based on the statistical results. This viewpoint distribution density map graphically displays the distribution of different viewpoints within the document, helping us quickly understand the distribution patterns of viewpoints within the document.

[0128] Step S154: Add a topic label to each viewpoint subset, wherein the topic label is generated based on a topic keyword set in the core discussion direction feature.

[0129] A topic label is assigned to each viewpoint subset. The topic label is generated based on the set of topic keywords in the core argument direction feature. Keywords related to the topic of the viewpoint subset are extracted from the core argument direction feature and combined into a topic label. For example, for a viewpoint subset on biodiversity conservation, keywords such as "biodiversity" and "conservation measures" are extracted from the core argument direction feature to form the topic label "biodiversity conservation." The topic label can succinctly summarize the thematic content of the viewpoint subset, making it easier to identify and classify the viewpoint subset.

[0130] Step S155: Integrate the subject labels, opinion content subsets and corresponding opinion distribution density maps to generate academic opinion extraction results containing a structured hierarchy.

[0131] Data integration of topic tags, opinion content subsets, and corresponding opinion distribution density maps is performed to form a structured, hierarchical academic opinion extraction result. This integration can be performed in a tree or list structure, with topic tags as top-level nodes, opinion content subsets as middle-level nodes, and opinion distribution density maps as related information for lower-level nodes. This structured integration approach creates a clear, hierarchical structure for academic opinion extraction, making it easier to understand and analyze.

[0132] Step S156: The academic opinion extraction result is transmitted to the target analysis terminal through the data interface, and the target analysis terminal is used to perform academic opinion comparison, research trend analysis or literature quality assessment operations.

[0133] The results of academic opinion extraction are transmitted to the target analysis terminal through the data interface. The target analysis terminal can be a specialized data analysis software or hardware device that can use the received academic opinion extraction results to perform various academic content analysis operations, such as academic opinion comparison, research trend analysis, or literature quality assessment. For example, in the academic opinion comparison operation, the target analysis terminal can compare the content of opinions on the same topic in different documents to identify differences and similarities in opinions; in the research trend analysis, the research trends in a certain field can be analyzed based on the opinion distribution density map and changes in topic labels; in the literature quality assessment, the quality of the literature can be evaluated based on the innovation and rationality of the opinions.

[0134] In the above embodiment, the opinion feature extraction model mainly consists of a text encoding layer, a cross-paragraph association layer, an opinion location layer, and a sentiment analysis submodule, etc. These modules and layers work together to achieve effective extraction of opinions from academic documents.

[0135] The text encoding layer is the foundational processing layer of the opinion feature extraction model. It is primarily responsible for tokenizing the input document fragments and generating token embeddings. It comprises a vocabulary mapping module and an embedding matrix. The vocabulary mapping module maps token units to corresponding token indices based on a predefined vocabulary, while the embedding matrix converts token indices into token embeddings. Term embeddings are also positionally encoded to preserve the position of tokens within the paragraph.

[0136] The cross-paragraph association layer receives the local context feature vector output by the text encoding layer. Its main function is to analyze the semantic coherence between different paragraph units. By calculating the semantic overlap and topic continuity parameters of adjacent paragraphs, it constructs an inter-paragraph semantic association model and generates semantic association descriptors between text segments.

[0137] The opinion positioning layer receives the semantic features of the document, analyzes the distribution density of the subject keyword set in the core argument direction features in the paragraph unit, identifies the candidate paragraphs where the subject keywords are clustered, and provides positioning information for subsequent opinion detection.

[0138] The sentiment analysis submodule includes a positive classifier, a negative classifier, and an innovative classifier, which are used to perform sentiment tendency analysis on the local context feature vector of the candidate paragraph and generate sentiment tendency parameters representing positive, negative or innovative arguments.

[0139] There are clear connections between these modules and layers. The output of the text encoding layer serves as the input to the cross-paragraph association layer. The output of the cross-paragraph association layer is combined with the local context feature vector of the text encoding layer to generate document semantic features. The document semantic features are input to the opinion positioning layer, and the local context feature vectors of the candidate paragraphs selected by the opinion positioning layer are input to the sentiment analysis submodule.

[0140] The quality and diversity of training data are crucial to the performance of the model. Collect a large amount of academic literature as training data. These literature should cover different fields, research directions, and writing styles to ensure that the model has wide applicability.

[0141] Manually annotate the collected academic literature, including the text location of each academic point of view and a summary of the point of view. The annotation process must adhere to unified standards and specifications to ensure accuracy and consistency. For example, when annotating a medical research paper, the text location of each research conclusion is clearly indicated, and the core content of the conclusion is extracted as a summary of the point of view.

[0142] The labeled data is divided into training, validation, and test sets. The training set is used to learn model parameters, the validation set is used to adjust model hyperparameters during training, and the test set is used to evaluate the model's final performance. The division ratio can be adjusted based on actual conditions, but generally follows certain principles to ensure that each dataset can be fully utilized.

[0143] Before training the opinion feature extraction model, a series of training parameters need to be set. These parameters will affect the training process and final performance of the opinion feature extraction model.

[0144] The learning rate controls the step size of each parameter update in the opinion feature extraction model. If the learning rate is too large, the opinion feature extraction model may skip the optimal solution; if the learning rate is too small, the opinion feature extraction model will converge very slowly. A dynamic learning rate strategy can be used: using a larger learning rate at the beginning of training and gradually reducing it as training progresses to balance the convergence speed and accuracy of the opinion feature extraction model.

[0145] Batch size refers to the number of data samples fed into the opinion feature extraction model during each training session. Larger batch sizes can improve training stability and efficiency, but may consume more memory. Smaller batch sizes can increase the randomness of the opinion feature extraction model, helping it escape local optima. Choosing an appropriate batch size depends on your computing resources and the complexity of the opinion feature extraction model.

[0146] The number of training epochs indicates how many times the model traverses the entire training dataset. If the number of training epochs is too small, the model may not fully learn the features in the data; if the number of training epochs is too large, the model may overfit. You can determine the appropriate number of training epochs by observing the performance on the validation set. Stop training when the performance on the validation set no longer improves.

[0147] The training process of the opinion feature extraction model is an iterative optimization process. By continuously adjusting the model parameters, the output results of the model are made as close as possible to the actual results of the annotation.

[0148] For example, a document text fragment from the training set is input into the opinion feature extraction model and processed sequentially according to the model's modules and hierarchical structure. The text encoding layer tokenizes and embeds the text, the cross-paragraph association layer analyzes the semantic associations between paragraphs, the opinion location layer locates candidate paragraphs, and the sentiment analysis submodule performs sentiment analysis, ultimately outputting the predicted opinion detection results. The predicted results of the opinion feature extraction model are compared with the annotated true results, and the loss value is calculated. The loss function can be a cross-entropy loss function, which measures the degree of difference between the model's predicted results and the true results. The smaller the loss value, the more accurate the model's prediction. Based on the calculated loss value, the backpropagation algorithm is used to calculate the gradient of each model parameter. The gradient represents the rate of change of the loss function with respect to the parameter, and the gradient can be used to determine the direction of parameter update.

[0149] Based on the calculated gradient and the preset learning rate, the model parameters are updated. The updated parameters will make the model output a prediction closer to the actual result during the next forward propagation.

[0150] The process of forward propagation, loss calculation, backpropagation, and parameter update is repeated until the preset number of training rounds is reached or the performance of the validation set no longer improves.

[0151] After model training is complete, the model's performance needs to be evaluated using a test set. For example, appropriate evaluation metrics can be selected to measure model performance, such as accuracy, recall, and F1 score. Accuracy indicates the proportion of samples correctly predicted by the model out of the total number of samples, recall indicates the proportion of positive samples correctly predicted by the model out of the actual positive samples, and F1 score is the harmonic mean of accuracy and recall. By comprehensively considering these metrics, the model's performance can be comprehensively evaluated. If the performance of the opinion feature extraction model is unsatisfactory, the model needs to be optimized. This can be achieved by adjusting training parameters, increasing training data, and improving the model structure. For example, if the model is found to be overfitting, you can try increasing the regularization term or reducing the model complexity. If the model's accuracy is low, you can consider increasing the diversity of the training data or adjusting parameters such as the learning rate.

[0152] Figure 2 A schematic diagram illustrates exemplary hardware and software components of an academic opinion extraction system 100 for academic literature that can implement the concepts of the present application, as provided in some embodiments of the present application. For example, a processor 120 can be used in the academic opinion extraction system 100 for academic literature and perform the functions described in the present application.

[0153] The academic opinion extraction system 100 for academic documents can be a general-purpose server or a special-purpose server, both of which can be used to implement the academic opinion extraction method for academic documents in this application. Although only one server is shown in this application, for convenience, the functions described in this application can be implemented in a distributed manner on multiple similar platforms to balance the processing load.

[0154] For example, the academic opinion extraction system 100 applied to academic documents may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and storage media 140 in different forms, such as a disk, ROM, or RAM, or any combination thereof. Exemplarily, the academic opinion extraction system 100 applied to academic documents may also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The method of the present application can be implemented according to these program instructions. The academic opinion extraction system 100 applied to academic documents also includes an I / O interface 150 between the computer and other input and output devices.

[0155] For ease of explanation, only one processor is described in the academic opinion extraction system 100 applied to academic documents. However, it should be noted that the academic opinion extraction system 100 applied to academic documents in the present application may also include multiple processors, so the steps performed by one processor described in the present application may also be performed jointly or individually by multiple processors. For example, if the processor of the academic opinion extraction system 100 applied to academic documents executes step A and step B, it should be understood that step A and step B may also be executed jointly by two different processors or individually in one processor. For example, the first processor executes step A, the second processor executes step B, or the first processor and the second processor execute steps A and B together.

[0156] In addition, an embodiment of the present invention also provides a readable storage medium, in which computer-executable instructions are preset. When a processor executes the computer-executable instructions, the academic opinion extraction method applied to academic literature as described above is implemented.

[0157] It should be noted that in order to simplify the description of the present invention and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, multiple features are sometimes combined into one embodiment, figure or description thereof.

Claims

1. A method for extracting academic opinions from academic literature, characterized in that: The method comprises: Obtaining an original document data set of the academic document to be analyzed, wherein the original document data set includes a plurality of document text units having a chapter structure; Preprocessing the original document data set to obtain a preprocessed document data set with a normalized structure, wherein the preprocessed document data set includes document text segments divided into paragraphs and corresponding chapter category tags; Calling a pre-trained viewpoint feature extraction model to perform semantic feature extraction on the pre-processed document data set to obtain document semantic features of each document text unit, wherein the document semantic features include semantic relevance between text segments and core argument direction features; Performing opinion detection processing on the semantic features of the document using the opinion feature extraction model to generate opinion detection results for each document text unit, wherein the opinion detection results include text position marks of the academic opinions to be extracted and an opinion content summary; Generating an academic opinion extraction result including an opinion content summary and corresponding text position marks based on the opinion detection result, and outputting the academic opinion extraction result to a target analysis terminal to support an academic content analysis operation; The opinion detection process is performed on the semantic features of the document by the opinion feature extraction model to generate an opinion detection result for each document text unit. The opinion detection result includes a text position mark of the academic opinion to be extracted and an opinion content summary, including: Inputting the document semantic features into the opinion positioning layer of the opinion feature extraction model, analyzing the distribution density of the theme keyword set in the core argument direction feature in the paragraph unit, and identifying candidate paragraphs where the theme keywords are concentrated; Performing sentiment analysis on the local context feature vectors of the candidate paragraphs, extracting sentiment parameters representing affirmative, negative, or innovative arguments, and screening target paragraphs whose sentiment parameters exceed a preset threshold; In the target paragraph, locating key nodes of argument logic based on semantic relevance descriptors between text segments, wherein the key nodes include text locations where claims are made, evidence is provided, and conclusions are drawn; Performing summary generation processing on the text content of the key nodes, extracting short sentences containing core propositions or conclusions as a summary of the viewpoint content; Recording the starting character position and the ending character position of the key node in the document text unit, and generating a corresponding text position mark; The opinion content summary is associated and matched with the corresponding text position mark to generate the opinion detection result of each document text unit.

2. The method for extracting academic opinions from academic documents according to claim 1, wherein: The preprocessing is performed on the original document data set to obtain a preprocessed document data set with normalized structure, wherein the preprocessed document data set contains document text segments divided into paragraphs and corresponding chapter category tags, including: Performing format unification processing on each document text unit in the original document data set to eliminate the typesetting differences between different documents and obtain initial document texts with consistent format; Performing chapter recognition processing on the initial document text, extracting the text content of the document title, abstract, introduction, methods, results, discussion, and conclusion parts, and determining the chapter boundaries of each part; Performing paragraph segmentation processing on the initial document text based on the chapter boundaries, dividing the continuous text in each chapter into multiple paragraph units, and obtaining document text segments divided by paragraphs; Punctuation is standardized for each paragraph unit, unifying the full-width and half-width symbol formats and correcting missing punctuation to generate paragraph text with standardized punctuation. Performing category tagging on each chapter content obtained by the chapter identification process, adding a corresponding chapter category tag to each paragraph unit, wherein the chapter category tag includes at least one of a title category, an abstract category, a method category, and a conclusion category; The document text segments divided into paragraphs and the corresponding chapter category tags are integrated to generate a pre-processed document data set with normalized structure.

3. The method for extracting academic opinions from academic documents according to claim 2, wherein: The step of performing format unification processing on each document text unit in the original document data set to eliminate typesetting differences between different documents and obtain initial document texts with consistent formats includes: Identifying a file format type of a document text unit in the original document data set, wherein the file format type includes at least one of a plain text format, a rich text format, and a PDF format; Perform markup removal on the text units of documents in rich text format, remove the formatting marks such as bold, italic, and underline, and retain the text content; Perform text extraction on the document text units in PDF format, identify the text content in the scanned PDF using optical character recognition technology, and generate editable text strings; Normalize the line breaks of the extracted text string, merge consecutive line breaks into a single line break, and eliminate text breaks caused by forced line breaks within a paragraph; Perform unified encoding processing on all document text units, and convert texts in different encoding formats into a unified Unicode encoding format; The processed document text units of each format are aligned in content to generate an initial document text with consistent format.

4. The method for extracting academic opinions from academic documents according to claim 2, wherein: The chapter identification process is performed on the initial document text to extract the text content of the document title, abstract, introduction, methods, results, discussion, and conclusion, and determine the chapter boundaries of each part, including: Identify the first line of text content of the initial document text, extract the text containing the title keywords as the document title, and record the starting position and ending position of the title; Search for the paragraph containing the abstract keyword in the text after the title, determine the starting position of the abstract part, and extract the continuous text to the position where the next chapter keyword appears as the end position of the abstract part; In the text following the abstract, search for paragraphs containing the introduction keywords, method keywords, result keywords, discussion keywords, and conclusion keywords, and determine the starting position of each chapter. Perform boundary verification on the text content after each chapter keyword to check whether the subsequent text contains feature words related to the chapter topic, including at least one of research background, experimental design, and data analysis; Record the starting and ending character positions of each chapter and generate a chapter boundary marker set; The text content of each chapter is extracted according to the chapter boundary marker set to generate a chapter content set including the document title, abstract, introduction, methods, results, discussion, and conclusion.

5. The method for extracting academic opinions from academic documents according to claim 1, wherein: The pre-trained viewpoint feature extraction model is called to perform semantic feature extraction on the pre-processed document data set to obtain the document semantic features of each document text unit. The document semantic features include the semantic relevance between text segments and the core argument direction features, including: Inputting the document text fragments in the preprocessed document data set into the text encoding layer of the opinion feature extraction model, performing lemma processing on each paragraph unit and generating a lemma embedding vector; Performing contextual modeling on the word embedding vector through the self-attention mechanism module of the text encoding layer to generate a local context feature vector containing the semantic relationship between words in the paragraph; Inputting the local context feature vector into the cross-paragraph association layer of the opinion feature extraction model, analyzing the semantic coherence between different paragraph units, and calculating the semantic overlap and topic continuity parameters of adjacent paragraphs; Building a semantic association model between paragraphs based on the semantic overlap and topic continuity parameters, and generating semantic association descriptors between text segments, wherein the semantic association descriptors are used to represent the strength of the logical cohesion of paragraph units; Performing a joint analysis on the local context feature vector and the inter-paragraph semantic relevance descriptor to extract the core discussion topics repeatedly emphasized in the document text unit and generate a core discussion direction feature, wherein the core discussion direction feature includes a topic keyword set and a topic occurrence frequency parameter; The semantic relevance descriptors and core argument direction features between the text fragments are subjected to feature fusion processing to generate the document semantic features of each document text unit.

6. The method for extracting academic opinions from academic documents according to claim 5, characterized in that: The step of inputting the document text segments in the pre-processed document data set into the text encoding layer of the opinion feature extraction model, performing lemma processing on each paragraph unit and generating a lemma embedding vector comprises: Perform word segmentation on each paragraph unit to divide the continuous text into word units with semantic meaning, wherein the word units include Chinese words and English words; Convert the word units after word segmentation to lowercase to unify the uppercase and lowercase formats of English words; Removing meaningless symbols from word units, wherein the meaningless symbols include at least one of special punctuation marks, mathematical symbols, and hyperlinks, and retaining word units with semantic information; Inputting the processed word unit into the vocabulary mapping module of the text encoding layer, mapping each word unit to a corresponding word index according to a predefined vocabulary; Calling the embedding matrix of the text encoding layer to perform vector conversion processing on the word unit index to generate a word unit embedding vector for each word unit, where the dimension of the word unit embedding vector is consistent with the number of columns of the embedding matrix; Perform position encoding on the word unit embedding vector of each paragraph unit, add a position encoding vector representing its position in the paragraph to each word unit embedding vector, and generate a word unit embedding vector set containing position information.

7. The method for extracting academic opinions from academic documents according to claim 1, wherein: The sentiment tendency analysis process is performed on the local context feature vector of the candidate paragraphs, the sentiment tendency parameters representing affirmative, negative or innovative arguments are extracted, and the target paragraphs whose sentiment tendency parameters exceed a preset threshold are screened, including: Inputting the local context feature vector of the candidate paragraph into the sentiment analysis submodule of the opinion feature extraction model, wherein the sentiment analysis submodule includes a positive classifier, a negative classifier, and an innovative classifier; Classifying the local context feature vector using the affirmative classifier to generate a positive probability value indicating that the paragraph content holds a positive attitude towards a certain proposition; Classifying the local context feature vector using the negativity classifier to generate a negativity probability value indicating that the paragraph content has a negative attitude towards a certain proposition; Classifying the local context feature vector using the innovative classifier to generate an innovative probability value indicating whether the paragraph content proposes a new theory or method; Combining the positive probability value, the negative probability value and the innovative probability value into a sentiment tendency parameter set; The candidate paragraphs whose probability value in the sentiment tendency parameter set exceeds a preset threshold are selected as target paragraphs, and the preset threshold is used to distinguish between discussions with clear sentiment tendencies and neutral descriptions.

8. The method for extracting academic opinions from academic documents according to claim 1, wherein: The generating of an academic opinion extraction result including an opinion content summary and corresponding text position marks based on the opinion detection result, and outputting the academic opinion extraction result to a target analysis terminal to support an academic content analysis operation, includes: performing redundancy elimination processing on the opinion content summaries in the opinion detection results, merging opinion contents expressing the same or similar propositions, and generating a compressed opinion content set; performing thematic grouping processing on the opinion content set according to the thematic relevance of each opinion content in the opinion content set to generate a plurality of thematically related opinion subsets; Extract the text position markers corresponding to the content summary of each opinion subset, count the distribution of each text position marker in the document text unit, and generate an opinion distribution density map; Add a topic label to each subset of opinions, wherein the topic label is generated based on a set of topic keywords in the core discussion direction feature; Integrate the topic tags, opinion content subsets, and corresponding opinion distribution density maps to generate academic opinion extraction results with a structured hierarchy; The academic opinion extraction results are transmitted to a target analysis terminal through a data interface, and the target analysis terminal is used to perform academic opinion comparison, research trend analysis or literature quality assessment operations.

9. An academic opinion extraction system applied to academic literature, characterized in that: It includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the academic opinion extraction method applied to academic literature as described in any one of claims 1 to 8 above.

Images