A method for generating scientific hypothesis graphs based on hypothesis relationship identification

Abstract

The present invention discloses a method for generating a scientific hypothesis map based on hypothesis relationship recognition, comprising the following steps: (1) collecting academic title information and full text in a target field; (2) identifying and screening hypothesis relationship declarative sentences; (3) identifying multiple types of structured hypothesis relationships; (4) distinguishing multiple types of hypothesis relationships; (5) processing hypothesis relationship standardization; (6) automatically generating a scientific hypothesis map. The present invention comprehensively improves the recognition accuracy and automation level of hypothesis relationships in academic texts.

Description

Technical Field

[0001] The present invention relates to the technical field of natural language processing and information extraction, and in particular to a method for generating a scientific hypothesis graph based on hypothesis relationship recognition. Background Art

[0002] In the field of scientific research, the proposal and verification of scientific hypotheses are the core links to promote the development of disciplines. Scientific hypotheses are usually proposed on the basis of a large number of existing studies and theories, and need to be verified through systematic experiments and data analysis. However, with the rapid increase in the number of scientific literature, researchers are faced with a huge challenge, namely how to effectively extract, organize and analyze the hypothesis relationship from the massive literature to construct a scientific hypothesis map to help guide subsequent research.

[0003] In recent years, with the development of natural language processing technology, especially the progress of semantic understanding and relationship extraction technology, it has become possible to automatically extract valuable information from unstructured scientific literature. By constructing a scientific hypothesis map, it is possible to visualize complex scientific problems and help researchers more intuitively understand the connection between various hypotheses and their relationship with known facts. However, existing scientific hypothesis map construction methods mostly rely on manual annotation or can only handle closed knowledge systems in specific fields. They lack the ability to transfer domain hypothesis relationships, and the degree of automation is not high, making it difficult to adapt to the rapidly changing scientific research environment. Summary of the invention

[0004] Purpose of the invention: The purpose of the present invention is to provide a method for generating a scientific hypothesis map based on hypothesis relationship identification to solve the problems existing in the background technology.

[0005] Technical solution: The method for generating a scientific hypothesis map based on hypothesis relationship identification described in the present invention comprises the following steps:

[0006] (1) Collect academic literature in the target field: Collect large-scale academic literature in the target field and perform structured analysis on academic abstracts and full text content;

[0007] (2) Identification and screening of hypothetical relationship declarative sentences: Based on the large language model, few-shot learning technology, and large language model structured output tools, we automatically identify structured hypothetical relationship declarative sentences from academic literature; we build a classifier that combines grammatical features with rule templates to screen out sentences that meet the criteria from hypothetical relationship declarative sentences;

[0008] (3) Identifying multi-type structured hypothesis relations: Combining large language model fine-tuning technology and predefined hypothesis relation architecture patterns, we can identify multi-type, structured hypothesis relations from hypothesis relation statements.

[0009] (4) Identify multiple types of hypothetical relationships: Combine a large language model with customized rule templates to automatically identify multiple types of relationships between hypotheses through semantic parsing and logical reasoning mechanisms;

[0010] (5) Processing hypothesis relationship standardization: Disambiguation of hypothesis relationship entities, correction of logical relationships, and resolution of hypothesis conflicts based on data verification, abbreviation expansion, and word form restoration operations;

[0011] (6) Automatically generate scientific hypothesis graphs: Build domain scientific hypothesis graphs based on graph databases, with assumed variables as graph nodes and multiple assumed relationship types as graph edge attributes, to achieve visualization of causal relationships and structural logic of domain knowledge.

[0012] Furthermore, in step (1), the full text of the target document is obtained. For content in PDF format, Python libraries such as PyMuPDF are used to parse the content into text format. For full text content on a web page, the Elsevier API or other academic publishers' API is used to obtain the full text content. The abstract and full text of the academic text are cleaned and saved. The type of academic document is publicly published academic document.

[0013] Furthermore, in step (2), the structured hypothetical relationship statement sentence includes: the type of hypothetical relationship statement sentence, the specific content of the hypothetical relationship statement sentence, the ID of the academic text from which the hypothetical relationship statement sentence comes, and other related attributes; the grammatical features include: subject-verb-object structure analysis, tense analysis, and semantic role labeling of the sentence; the rule template refers to a template that uses regular expressions to extract the basic structure from the chapters of the academic text that potentially contain hypothetical relationships; the hypothetical relationship statement sentence is represented as:

[0014]

[0015] Among them, "type" is the type of the hypothetical relationship statement (hypothesis part, result part or conclusion part), "sentence" is the specific content, and "source" is the academic text ID;

[0016] The classifier is represented as:

[0017] ;

[0018] in, C(s) is the classification result. If the sentence s If the matching condition is met, the classifier marks it as 1, otherwise it marks it as 0.

[0019] Furthermore, in step (3), the predefined hypothesis relationship architecture pattern refers to the architecture used to extract multiple types of hypothesis relationship frameworks and their corresponding data organization structures from the hypothesis relationship statement, wherein the architecture pattern is organized in a nested JSON format to ensure multi-dimensional and hierarchical extraction of hypothesis relationships; the multi-type and structured hypothesis relationships include: independent variables, dependent variables, effects, mediating variables, and moderating variables; specifically, as follows:

[0020] set up V Represents a set of assumed variable types, the set V It is divided into five components: independent variables V ind , dependent variable V dep ,effect V eff , mediating variables V med and moderating variables V mod ,but:

[0021] ;

[0022] Among them, the independent variable V ind Considered as a potential cause of variation in other variables; dependent variable V dep will respond to changes in the independent variable, while the effect V eff The mediating variables V med The role of the moderating variable is to convey or explain the relationship between the independent variable and the dependent variable; V mod It shows how external conditions affect the strength or direction of the relationship between the independent and dependent variables; R Represents a collection of relations, a collection R is constructed as a nested JSON object:

[0023] ;

[0024] exist R In each V n Each has a paperID that links to the corresponding academic paper; then, supervised fine-tuning is used to train a model that can extract variables from academic texts.

[0025] Furthermore, in step (4), there are multiple types of relationships, including: positive relationship, negative relationship, moderating effect, mediating effect, and causal effect.

[0026] The scientific hypothesis graph generation system based on hypothesis relationship recognition described in the present invention comprises:

[0027] Collection module (used to collect academic literature in the target field): collect large-scale academic literature in the target field and perform structured analysis on academic abstracts and full text content;

[0028] Declarative sentence module (used to identify and screen hypothetical relationship declarative sentences): Based on the large language model, few-shot learning technology and large language model structured output tools, it automatically identifies structured hypothetical relationship declarative sentences from academic full texts; it builds a classifier that combines grammatical features with rule templates to screen out qualified sentences from hypothetical relationship declarative sentences;

[0029] Multi-type structured hypothesis relationship extraction module (used to identify multi-type structured hypothesis relationships): Combines large language model fine-tuning technology and predefined hypothesis relationship architecture patterns to identify multi-type, structured hypothesis relationships from hypothesis relationship statements;

[0030] Discrimination module (used to distinguish multiple types of hypothetical relationships): Combines a large language model with customized rule templates to automatically identify multiple types of relationships between hypotheses through semantic analysis and logical reasoning mechanisms;

[0031] Standardization module (used to handle hypothesis relationship standardization): performs hypothesis relationship entity disambiguation, logical relationship error correction, and hypothesis conflict resolution based on data verification, abbreviation expansion, and morphological restoration operations;

[0032] Scientific hypothesis graph module (used to automatically generate scientific hypothesis graphs): Build a domain scientific hypothesis graph based on a graph database, assume variables as graph nodes, and multiple assumed relationship types as graph edge attributes to achieve causal association and structural logic visualization of domain knowledge.

[0033] Furthermore, in the collection module, the abstract and full-text content of the target document are obtained. For PDF format content, Python libraries such as PyMuPDF are used to parse it into text format; for the full-text content of the web page, the Elsevier API or other academic publishers' API is used to obtain the full-text content; the abstract and full-text of the academic text are cleaned and saved; among them, the type of academic document is publicly published academic document.

[0034] Furthermore, in the declarative sentence module, the structured hypothetical relationship declarative sentence includes: the type of hypothetical relationship declarative sentence, the specific content of the hypothetical relationship declarative sentence, the academic text ID from which the hypothetical relationship declarative sentence comes, and other related attributes; the grammatical features include: the subject-verb-object structure analysis, tense analysis, and semantic role labeling of the sentence; the rule template refers to a template that uses regular expressions to extract the basic structure from the chapters of the academic text that potentially contain hypothetical relationships. The hypothetical relationship declarative sentence is represented as:

[0035]

[0036] Among them, "type" is the type of the hypothetical relationship statement (hypothesis part, result part or conclusion part), "sentence" is the specific content, and "source" is the academic text ID;

[0037] The classifier is represented as:

[0038] ;

[0039] in, C(s) is the classification result. If the sentence s If the matching condition is met, the classifier marks it as 1, otherwise it marks it as 0.

[0040] Furthermore, in the multi-type structured hypothesis relationship module, the predefined hypothesis relationship architecture pattern refers to the use of the nested JSON format to organize various types of hypothesis relationship frameworks and their corresponding data organization structures from the hypothesis relationship statement, so as to ensure the multi-dimensional and hierarchical extraction of the hypothesis relationship; the multi-type and structured hypothesis relationship includes: independent variables, dependent variables, effects, mediating variables, and moderating variables; specifically, as follows:

[0041] set up V Represents a set of assumed variable types, the set V It is divided into five components: independent variables V ind , dependent variable V dep ,effect V eff , mediating variables V med and moderating variables V mod ,but:

[0042]

[0043] Among them, the independent variable V ind Considered as a potential cause of variation in other variables; dependent variable V dep will respond to changes in the independent variable, while the effect V eff The mediating variables V med The role of the moderating variable is to convey or explain the relationship between the independent variable and the dependent variable; V mod It shows how external conditions affect the strength or direction of the relationship between the independent and dependent variables;R Represents a collection of relations, a collection R is constructed as a nested JSON object:

[0044]

[0045] exist R In each V n Each has a paperID that links to the corresponding academic paper; then, supervised fine-tuning is used to train a model that can extract variables from academic texts.

[0046] Furthermore, in the discrimination module, there are multiple types of relationships, including positive relationships, negative relationships, moderating effects, mediating effects, and causal effects.

[0047] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: by extracting multidimensional variable relationships in academic texts, establishing hypothetical relationships among structured variables such as independent variables, dependent variables, effects, mediating variables, and moderating variables, and performing hypothetical relationship entity disambiguation, logical relationship error correction, and hypothesis conflict resolution based on operations such as data verification, abbreviation expansion, and word form restoration, the recognition accuracy and automation level of hypothetical relationships in academic texts are comprehensively improved, and a systematic and structured scientific hypothesis map is constructed. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a framework diagram of the present invention. DETAILED DESCRIPTION

[0049] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0050] like Figure 1 As shown, an embodiment of the present invention provides a method for generating a scientific hypothesis graph based on hypothesis relationship identification, comprising the following steps:

[0051] S1. Collection of academic literature in the target field. First, select a target field. The requirement of the target field is that the typical research paradigm in this field needs to meet hypothesis testing and variable verification, such as the field of "econometrics". Secondly, use manual methods to crawl academic literature in the target field from Web of Science. For each paper, collect its bibliographic information (such as title, author, publication year, publication name, document ID, etc.), as well as the abstract. Then, call the API interface of academic platforms such as OpenAlex or SemanticScholar to obtain the full text content of the target document. For PDF format content, use Python libraries such as PyMuPDF to parse it into text format; for the full-text content of the web page, use Elsevier API or other academic publishers API to obtain the full text content. Finally, the abstract and full text of the academic text are cleaned and saved.

[0052] S2. Extraction and screening of hypothetical relationship declarative sentences. To extract hypothetical relationships from academic literature, we must first identify declarative sentences that may contain hypotheses. The structure of academic texts usually includes an introduction, a literature review (or theory & hypothesis), methods, results, and conclusions. Declarative sentences containing hypotheses usually appear in the hypothesis section, results section, and conclusion section of academic papers. The hypothesis section is usually a specific statement based on existing theories or logical reasoning about the relationship between expected variables. The results section and the conclusion section are general and explain the hypothetical results and main findings of the academic text. Use the Qwen-max large language model to identify hypothetical relationship declarative sentences in academic texts. To improve recognition performance, examples are added to the input prompts, and then LangChain is used to verify the output type to ensure that a structured JSON format is returned. A hypothetical relationship declarative sentence can be expressed as:

[0053] ;

[0054] Among them, "type" is the type of the hypothetical relationship statement (hypothesis part, result part or conclusion part), "sentence" is the specific content, and "source" is the academic text ID. There are some problems when processing some hypothetical relationship statements, such as the lack of hypothetical relationship, or interrogative sentences or negative sentences. To solve this problem, the present invention designs a classifier based on pattern matching to determine whether a given hypothetical relationship statement meets the standard. By combining syntactic analysis and predefined keywords, a joint regular expression method is used to match statements; if the match is successful, the statement is retained, and if the match fails, the statement is discarded. The classifier is expressed as:

[0055] ;

[0056] in, C(s) is the classification result. If the sentences If the matching condition is met, the classifier marks it as 1, otherwise it marks it as 0.

[0057] S3. Multi-type structured hypothesis relationship recognition. Hypothesis relationship recognition is regarded as a joint task of entity recognition and relation extraction in the field of natural language processing. To represent the extraction of hypotheses and their corresponding relation patterns from hypothesis relationship statements, the following definition is used: V A collection representing hypothetical variable types. V It is divided into five components: independent variables ( V ind ), dependent variable ( V dep ),effect( V eff ), mediating variables ( V med ) and the moderating variable ( V mod ).

[0058]

[0059] Among them, the independent variable ( V ind ) is considered as a potential cause of changes in other variables. The dependent variable ( V dep ) responds to changes in the independent variable, while the effect ( V eff ) then explain the positive or negative mechanisms of these effects. V med ) is used to convey or explain the relationship between the independent variable and the dependent variable. V mod ) shows how external conditions affect the strength or direction of the relationship between the independent and dependent variables. R Represents a collection of relations. R is constructed as a nested JSON object:

[0060] ;

[0061] exist R In each V n Each of them has a paperID, which links to the corresponding academic paper. Next, supervised fine-tuning is used to train a model that can extract variables from academic texts. The fine-tuned model not only extracts complex variable relationships, but also outputs them in a structured format. Several data pieces are manually annotated from the hypothetical relationship statements and used for training and testing in a ratio of 7:3.

[0062] S4. Automatic identification of multi-type structured hypothetical relationships. The structured hypothetical relationship data generated from the previous step is input into the large language model, and the text content of each hypothetical relationship is deeply analyzed through semantic parsing capabilities. According to the effect type, it is preliminarily divided into positive relationships, negative relationships, regulatory effects, mediating effects, or causal effects. Subsequently, the preliminary classification results are verified and adjusted using pre-designed customized rule templates. These templates are based on domain knowledge and common patterns of hypothetical relationships, and are designed to identify the characteristics and logical structure of specific types of relationships. The logical reasoning module is then used to further verify the identified relationship type to ensure the logical consistency of the hypothetical relationship in the overall context and the rationality of the reasoning chain. After semantic parsing, rule template matching, and logical reasoning verification, the system makes a final judgment on the type of each hypothetical relationship and outputs it in the form of a multi-tuple. The annotation content includes the specific relationship type and related variable information.

[0063] S5. Standardization of hypothesis relations. Based on data validation, abbreviation expansion, and lemmatization, hypothesis relation entity disambiguation, logical relation error correction, and hypothesis conflict abbreviated are performed. Data validation screens out hypothesis relations that do not conform to predefined structural patterns or have obvious logical errors. Abbreviation expansion is responsible for converting variable abbreviations into full names. Each variable is associated with its corresponding source data through paperID. Then the full names of variables containing abbreviations are detected using a large language model. The specific process is as follows: ① Use regular expressions to match all-capitalized words that appear at least twice in a row and identify them as abbreviations; ② Collect and summarize the occurrence of abbreviations and their IDs; ③ Use the Qwen-max large language model and LangChain tool to identify and standardize the full names of abbreviations and map them to an abbreviation-full name dictionary; ④ Use the dictionary to replace all abbreviations with full names. Finally, the results will be sorted and merged. Finally, SpaCy is used to lemmatize the hypothesis variables and convert the identified variables into their standard forms to standardize the vocabulary and eliminate inconsistencies that may be caused by tense and other morphological changes.

[0064] S6. Automatic generation of scientific hypothesis maps. The basic structure in the graph database is constructed using the hypothesis relationship data after standardization in the previous link. The variables in each hypothesis will be mapped to nodes in the graph, and these nodes are connected by relationship edges. The attributes of the relationship edges will describe in detail the various types of relationships between the hypotheses, including positive relationships, negative relationships, regulatory effects, mediating effects, and causal effects. The graph database can efficiently store and query these complex relationship structures to ensure that the associations and logical structures between the data can be accurately presented. Secondly, the domain scientific hypothesis map is automatically generated through the graph algorithm and query function of the graph database. Displaying the hypothesis relationship between variables reflects more complex causal chains, regulatory effects, mediating effects, and other multi-level relationships, helping researchers to intuitively understand the causal relationship and structural logic of knowledge in the field. In addition, the present invention can also display the generated map through a visualization tool, so that researchers can intuitively browse and analyze the hypothesis relationship through an interactive graphical interface.

[0065] The embodiment of the present invention also provides a scientific hypothesis graph generation system based on hypothesis relationship recognition, including:

[0066] Collection module (used to collect academic literature in the target field): collect large-scale academic literature in the target field, and perform structured analysis on academic abstracts and full-text content. In particular, obtain the abstract and full-text content of the target document. For PDF format content, use Python libraries such as PyMuPDF to parse it into text format; for full-text content of web pages, use ElsevierAPI or other academic publishers' APIs to obtain full-text content; clean the abstract and full-text of the academic text and save it; the type of academic literature is publicly published academic literature.

[0067] Declarative sentence module (used to identify and screen hypothetical relationship declarative sentences): Based on the large language model, few-shot learning technology and large language model structured output tools, it automatically identifies structured hypothetical relationship declarative sentences from academic full texts; builds a classifier that combines grammatical features with rule templates to screen qualified sentences from hypothetical relationship declarative sentences; among them, structured hypothetical relationship declarative sentences include: hypothetical relationship declarative sentence type, hypothetical relationship declarative sentence specific content, hypothetical relationship declarative sentence source academic text ID, and other related attributes; grammatical features include: sentence subject-verb-object structure analysis, tense analysis, and semantic role labeling; rule templates refer to templates that use regular expressions to extract basic structures from chapters in academic texts that potentially contain hypothetical relationships. Hypothetical relationship declarative sentences are represented as:

[0068]

[0069] Among them, "type" is the type of the hypothetical relationship statement (hypothesis part, result part or conclusion part), "sentence" is the specific content, and "source" is the academic text ID;

[0070] The classifier is represented as:

[0071] ;

[0072] in, C(s) is the classification result. If the sentence s If the matching condition is met, the classifier marks it as 1, otherwise it marks it as 0.

[0073] Multi-type structured hypothesis relationship extraction module (used to identify multi-type structured hypothesis relationships): Combines large language model fine-tuning technology and predefined hypothesis relationship architecture patterns to identify multi-type, structured hypothesis relationships from hypothesis relationship statements; predefined hypothesis relationship architecture patterns refer to the use of hypothesis relationship frameworks and their corresponding data organization structures to extract multiple types of hypothesis relationship frameworks from hypothesis relationship statements, where the architecture pattern is organized in a nested JSON format to ensure multi-dimensional and hierarchical extraction of hypothesis relationships; multi-type, structured hypothesis relationships include: independent variables, dependent variables, effects, mediating variables, and moderating variables; the details are as follows:

[0074] set up V Represents a set of assumed variable types, the set V It is divided into five components: independent variables V ind , dependent variable V dep ,effect V eff , mediating variables V med and moderating variables V mod ,but:

[0075]

[0076] Among them, the independent variable V ind Considered as a potential cause of variation in other variables; dependent variable V dep will respond to changes in the independent variable, while the effect V eff The mediating variables V med The role of the moderating variable is to convey or explain the relationship between the independent variable and the dependent variable; V modIt shows how external conditions affect the strength or direction of the relationship between the independent and dependent variables; R Represents a collection of relations, a collection R is constructed as a nested JSON object:

[0077]

[0078] exist R In each V n Each has a paperID that links to the corresponding academic paper; then, supervised fine-tuning is used to train a model that can extract variables from academic texts.

[0079] Discrimination module (used to distinguish multiple types of hypothetical relationships): Combining a large language model with customized rule templates, it automatically identifies multiple types of relationships between hypotheses through semantic analysis and logical reasoning mechanisms; multiple relationship types, including: positive relationships, negative relationships, regulatory effects, mediating effects, and causal effects.

[0080] Standardization module (used to handle hypothesis relationship standardization): performs hypothesis relationship entity disambiguation, logical relationship error correction, and hypothesis conflict resolution based on data verification, abbreviation expansion, and morphological restoration operations;

[0081] Scientific hypothesis graph module (used to automatically generate scientific hypothesis graphs): Build a domain scientific hypothesis graph based on a graph database, assume variables as graph nodes, and multiple assumed relationship types as graph edge attributes to achieve causal association and structural logic visualization of domain knowledge.

Claims

1. A method for generating a scientific hypothesis graph based on hypothesis relationship identification, characterized in that: The following steps are involved: (1) Collect academic literature in the target field: Collect large-scale academic literature in the target field and perform structured analysis on the abstracts and full text contents of the academic literature; (2) Identification and screening of hypothetical relationship statements: Based on the large language model, few-shot learning technology and large language model structured output tools, structured hypothetical relationship statements are automatically identified from academic texts; a classifier that combines grammatical features with rule templates is constructed to automatically identify qualified sentences from hypothetical relationship statements; structured hypothetical relationship statements include: type of hypothetical relationship statement, specific content of hypothetical relationship statement, and ID of the academic text from which the hypothetical relationship statement comes; grammatical features include: subject-verb-object structure analysis, tense analysis, and semantic role labeling of sentences; rule templates refer to templates that use regular expressions to extract basic structures from chapters in academic texts that potentially contain hypothetical relationships; hypothetical relationship statements are represented as: ； Among them, "type" is the type of the hypothetical relationship statement, i.e., the hypothesis part, result part or conclusion part, "sentence" is the specific content, and "source" is the academic text ID; The classifier is represented as: ； in, C(s) is the classification result. If the sentence s If the matching condition is met, the classifier is marked as 1, otherwise it is marked as 0; (3) Identify multi-type structured hypothesis relationships: Combine the large language model fine-tuning technology and the predefined hypothesis relationship architecture pattern to identify multi-type and structured hypothesis relationships from hypothesis relationship statements; the predefined hypothesis relationship architecture pattern refers to the framework for extracting multiple types of hypothesis relationship frameworks and their corresponding data organization structures from hypothesis relationship statements, and the architecture pattern is organized in a nested JSON format to ensure multi-dimensional and hierarchical extraction of hypothesis relationships; the multi-type and structured hypothesis relationships include: independent variables, dependent variables, effects, mediating variables, and moderating variables; the details are as follows: set up V Represents a set of assumed variable types, the set V It is divided into five components: independent variables V ind , dependent variable V dep ,effect V eff , mediating variables V med and moderating variables V mod ,but: ； Among them, the independent variable V ind Considered as a potential cause of variation in other variables; dependent variable V dep will respond to changes in the independent variable, while the effect V eff The mediating variables V med The role of the moderating variable is to convey or explain the relationship between the independent variable and the dependent variable; V mod It shows how external conditions affect the strength or direction of the relationship between the independent and dependent variables; R Represents a collection of relations, a collection R is constructed as a nested JSON object: ； exist R In each V n Each of them has a paperID, which links to the corresponding academic paper. Next, supervised fine-tuning is used to train a model that can extract structured variables from academic texts. (4) Identify multiple types of hypothetical relationships: Combine a large language model with customized rule templates to automatically identify multiple types of relationships between hypotheses through semantic parsing and logical reasoning mechanisms; (5) Processing hypothesis relationship standardization: Disambiguation of hypothesis relationship entities, correction of logical relationships, and resolution of hypothesis conflicts based on data verification, abbreviation expansion, and word form restoration operations; (6) Automatically generate scientific hypothesis graphs: Build domain scientific hypothesis graphs based on graph databases, with assumed variables as graph nodes and multiple assumed relationship types as graph edge attributes, to achieve visualization of causal relationships and structural logic of domain knowledge.

2. A method for generating a scientific hypothesis graph based on hypothesis relationship identification according to claim 1, characterized in that: In step (1), the full text of the target document is obtained. For PDF format content, the PyMuPDF library is used to parse it into text format; for the full text content of the web page, the Elsevier API is used to obtain the full text content; the abstract and full text of the academic text are cleaned and saved; among them, the type of academic document is publicly published academic document.

3. The method for generating a scientific hypothesis graph based on hypothesis relationship identification according to claim 1, characterized in that: In step (4), there are multiple types of relationships, including: positive relationships, negative relationships, moderating effects, mediating effects, and causal effects.

4. A scientific hypothesis graph generation system based on hypothesis relationship recognition, characterized in that: include: The collection module is used to collect academic literature in the target field: collect large-scale academic literature in the target field, and perform structured analysis on academic abstracts and full text content; The declarative sentence module is used to identify and screen hypothetical relationship declarative sentences: based on the large language model, few-sample learning technology and large language model structured output tool, it automatically identifies structured hypothetical relationship declarative sentences from the academic full text; constructs a classifier that combines grammatical features and rule templates to screen qualified sentences from hypothetical relationship declarative sentences; among them, structured hypothetical relationship declarative sentences include: hypothetical relationship declarative sentence type, hypothetical relationship declarative sentence specific content, hypothetical relationship declarative sentence source academic text ID; grammatical features include: sentence subject-verb-object structure analysis, tense analysis, and semantic role labeling; rule templates refer to templates that use regular expressions to extract basic structures from chapters in academic texts that potentially contain hypothetical relationships; hypothetical relationship declarative sentences are represented as: ； Among them, "type" is the type of the hypothetical relationship statement, i.e., the hypothesis part, result part or conclusion part, "sentence" is the specific content, and "source" is the academic text ID; The classifier is represented as: ； in, C(s) is the classification result. If the sentence s If the matching condition is met, the classifier marks it as 1, otherwise it marks it as 0 The multi-type structured hypothesis relationship extraction module is used to identify multi-type structured hypothesis relationships: combining large language model fine-tuning technology and predefined hypothesis relationship architecture patterns to identify multi-type, structured hypothesis relationships from hypothesis relationship statements; wherein the predefined hypothesis relationship architecture pattern refers to the use of multiple types of hypothesis relationship frameworks and their corresponding data organization structures to extract from hypothesis relationship statements, wherein the architecture pattern is organized in a nested JSON format to ensure multi-dimensional, hierarchical extraction of hypothesis relationships; multi-type, structured hypothesis relationships include: independent variables, dependent variables, effects, mediating variables, and moderating variables; specifically as follows: set up V Represents a set of assumed variable types, the set V It is divided into five components: independent variables V ind , dependent variable V dep ,effect V eff , mediating variables V med and moderating variables V mod ,but: ； Among them, the independent variable V ind Considered as a potential cause of variation in other variables; dependent variable V dep will respond to changes in the independent variable, while the effect V eff The mediating variables V med The role of the moderating variable is to convey or explain the relationship between the independent variable and the dependent variable; V mod It shows how external conditions affect the strength or direction of the relationship between the independent and dependent variables; R Represents a collection of relations, a collection R is constructed as a nested JSON object: ； exist R In each V n Each of them has a paperID, which links to the corresponding academic paper. Next, supervised fine-tuning is used to train a model that can extract variables from academic texts. The discrimination module is used to discriminate multiple types of hypothetical relationships: it combines a large language model with customized rule templates to automatically identify multiple types of relationships between hypotheses through semantic analysis and logical reasoning mechanisms; The standardization module is used to process the standardization of hypothesis relations: based on data verification, abbreviation expansion and morphological restoration operations, hypothesis relation entity disambiguation, logical relation error correction and hypothesis conflict resolution are performed; The scientific hypothesis graph module is used to automatically generate scientific hypothesis graphs: construct a domain scientific hypothesis graph based on a graph database, assume variables as graph nodes, and use multiple assumed relationship types as graph edge attributes to achieve causal associations and structural logic visualization of domain knowledge.

5. A scientific hypothesis graph generation system based on hypothesis relationship identification according to claim 4, characterized in that: In the collection module, the abstract and full-text content of the target document are obtained. For PDF format content, the PyMuPDF library is used to parse it into text format; for the full-text content of the web page, the Elsevier API is used to obtain the full-text content; the abstract and full-text of the academic text are cleaned and saved; among them, the type of academic document is publicly published academic document.

6. A scientific hypothesis graph generation system based on hypothesis relationship identification according to claim 4, characterized in that: In the discrimination module, there are multiple types of relationships, including positive relationships, negative relationships, moderating effects, mediating effects, and causal effects.

Images