A method for predicting potential hypothesis relationships based on a causal signed network

Abstract

The application discloses a potential hypothesis relationship prediction method based on a causal signed network, and comprises the following steps: scientific hypothesis variable identification and multi-dimensional causal relationship discrimination; construction of a causal signed network; hypothesis inspiration heuristic search; potential hypothesis relationship prediction; and potential hypothesis relationship credibility sorting. The application can effectively capture and utilize intermediary and regulation effects, thereby generating new hypotheses containing the complex mechanisms and being closer to scientific problems in the real world, and enabling scientific researchers to preferentially focus on the most likely established hypotheses and save the time for screening and verification.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, knowledge discovery and scientific literature mining, and particularly relates to a potential hypothesis relationship prediction method based on a causal symbolic network. BACKGROUND

[0002] AI for Science, which utilizes artificial intelligence technology to assist scientific discovery, has become a frontier trend in scientific research. Some existing computational methods attempt to discover potential scientific hypotheses with the aid of large language model technology. However, the existing technical solutions still have many limitations, including: ①Cognitive bias of training data. The ability of large language models comes from their vast training data. Since the data itself may contain outdated, incorrect, or even biased knowledge, the model will inevitably reproduce and even amplify these biases during the learning process, causing the hypotheses generated by the model to tend to be mainstream or popular, while ignoring more disruptive and counterintuitive innovative paths. ②Black box characteristics of reasoning mechanism. The reasoning process of large language models based on neural networks lacks explicit logic and causal paths. When the model generates new hypotheses, it is difficult to determine whether it is based on rigorous logical deduction or just shallow associations of data, making it difficult to guarantee the causal logic consistency of the scientific hypotheses generated by the model, and even potentially producing misleading pseudo-hypotheses that interfere with scientific research activities. SUMMARY

[0003] The purpose of the present application is to provide a potential hypothesis relationship prediction method based on a causal symbolic network, which automatically and structurally extracts scientific hypothesis information containing causal direction, effect property and complex mediation / modulation relationship from unstructured scientific literature, constructs a network model that can formally express the causal logic of scientific knowledge, and generates new and structurally complete hypothesis relationships to solve the problems in the background art.

[0004] The technical solution of the present application is a potential hypothesis relationship prediction method based on a causal symbolic network, which includes the following steps:

[0005] (1) Scientific hypothesis variable identification and multi-dimensional causal relationship discrimination: From unstructured scientific literature data sources, automatically identify and extract structured scientific hypothesis variables, and discriminate the causal relationships between variables in multiple dimensions;

[0006] (2) Constructing a causal symbolic network: based on the causal relationships discriminated as valid in step (1), constructing a directed and signed weight causal symbolic network, where the nodes represent scientific variables, the directed edges represent the direction of causal relationships, and the signs of the edges represent the properties of causal effects;

[0007] (3) Hypothesis-inspired search: for one or more target variables in the network, a heuristic random walk strategy is adopted to explore the causal symbolic network, generating a local subgraph containing potential hypothesis inspiration;

[0008] (4) Potential hypothesis relationship prediction: within the generated local subgraph, a set of new candidate hypothesis relationships are generated by reorganizing the knowledge of existing variables and their causal relationship paths;

[0009] (5) Potential hypothesis relationship credibility ranking: the comprehensive credibility score of each candidate hypothesis relationship is designed and calculated, and the candidate hypothesis relationships are ranked in descending order according to the score, and finally the list of potential hypothesis relationships is output.

[0010] Further, step (1) includes the following steps:

[0011] (11) A large language model fine-tuned on scientific literature in a specific field is used to perform a joint task of named entity recognition and relationship extraction on the input scientific paper full text or abstract text;

[0012] (12) The large language model automatically extracts and formalizes structured five-tuple hypothesis relationships from the text, and the structure of the five-tuple is: <independent variable, dependent variable, effect, mediator variable, moderator variable>, where the effect refers to the impact direction of the independent variable on the dependent variable, and its value is positive or negative;

[0013] (13) According to a set of preset causal discrimination dimensions, the variable relationships represented by the five-tuple are evaluated one by one; the preset causal discrimination includes the following dimensions: domain consistency, direction rationality, mechanism rationality, causal temporal sequence, mediation rationality, and adjustment appropriateness;

[0014] (14) The evaluation is performed by expert consultation or automated rule logic, and when a variable relationship meets at least two of the dimension criteria, the variable relationship is confirmed as an effective causal relationship and used for subsequent network construction.

[0015] Further, step (2) includes the following steps:

[0016] (21) All identified independent variables and dependent variables are instantiated as nodes in the network;

[0017] (22) If there is an effective causal relationship between variable A and variable B, where A is an independent variable and B is a dependent variable, a directed edge is established between the node representing A and the node representing B, with the edge direction pointing from A to B;

[0018] (23) According to the effect information in the five-tuple, a signed weight is assigned to the directed edge: if the effect is positive, the sign of the edge is defined as positive +; if the effect is negative, the sign of the edge is defined as negative -;

[0019] (24) The information of the intermediary variable and the regulatory variable contained in the five-tuple is attached and stored as an attribute on the directed edge.

[0020] Further, step (3) comprises the following steps:

[0021] (31) By calculating and comparing at least one network centrality index of each node in the network, one or more target variables with high exploration value are automatically identified and selected, denoted as v ego ;

[0022] (32) Starting from v ego , a heuristic random walk algorithm is performed on the network, wherein the next node to be visited v next in the walk process is randomly selected from the bidirectional neighborhood of the current node v current ;

[0023] (33) The bidirectional neighborhood of the current node N bidirectional ( v current ) is defined as the union set of the predecessor node set and the successor node set of the node, and the mathematical expression is:

[0024] ;

[0025] Wherein, Pre( v ) represents the incoming neighborhood set of node v , and Suc( v ) represents the outgoing neighborhood set of node v .

[0026] Further, the termination condition of the heuristic random walk is set to satisfy any of the following conditions: the depth of the walk path, i.e. the number of edges passed from the starting point v ego , reaches a pre-set maximum depth threshold d max ; the total number of unique nodes visited in the walk process exceeds a pre-specified node number threshold n max ; when the walk terminates, all visited nodes and the original edges between them collectively form a local subgraph that maintains the original network structure G sub= G [ V visited ]。

[0027] Further, in step (4), the potential hypothetical relationship prediction is performed in a local subgraph v ego centered around G sub , including the following steps:

[0028] (41) Identify all candidate variable nodes G sub that have a path connection but are not directly adjacent to v ego , forming a candidate set U={ v i , v 1 , v 2 , …, v i};

[0029] (42) For each candidate variable v i U, infer the most likely causal direction between v ego and v i by computing the node directional measure on all paths between them, and quantify it as a directional score;

[0030] (43) Determine the sign of the causal relationship in the new hypothesis, i.e., positive or negative, by calculating the cumulative effect of the path between them by multiplying the signed effect of all edges on the path;

[0031] (44) Check if there are edges with intermediary or regulatory properties on the path connecting v ego and v i to identify potential intermediary variables and regulatory variables in the new hypothesis.

[0032] Further, the calculation of the directional score follows the formula:

[0033] ;

[0034] where represents the node directional measure between G sub and v ego and v iThe set of intermediate nodes on all paths; It is the size of the set. d ( x ) is a predefined node orientation metric function, whose positive or negative value indicates the orientation of the intermediate node relative to the path. v ego The direction of forward or backward movement.

[0035] Furthermore, the cumulative effect is calculated according to the following formula:

[0036] ;

[0037] in, e j It is the first on a specific path that is connected. j Edge; effect e j ) is the symbolic network of that edge. k It is the length of the path.

[0038] Furthermore, step (5) includes the following steps:

[0039] (51) The number of common neighbors is used as the main criterion for evaluating the structural proximity of two variables, and this number is quantified as the target variable. v ego Neighbor set and candidate variables v i The size of the intersection of the neighbor sets, i.e. ,in Represents a node v The set of direct neighbors;

[0040] (52) The frequency of occurrence of potential mediating and moderating variables identified by the candidate hypothesis is used as an additional evaluation indicator for statistical analysis;

[0041] (53) The number of common neighbors and the frequencies of mediators and moderating variables are combined into a comprehensive credibility score through a weighted summation function;

[0042] (54) All candidate hypotheses h i Candidate hypotheses are sorted in descending order based on their overall credibility scores to generate a final ordered set of hypotheses. H ={ h 1 , h 1 ,…, h i}

[0043] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0044] (1) Improve the reliability of hypothesis prediction and the consistency of causal logic: The present application is not based on simple correlation, but extracts causal relationships with direction and nature from literature, and screens through multi-dimensional discriminant criteria, ensuring knowledge quality from the source. The causal symbolic network constructed makes all subsequent predictions based on causal logic chains, significantly improving the scientificity and reliability of generated hypotheses;

[0045] (2) Enhance the explainability of the prediction process and results: Unlike the black box mechanism defects of large language model driven hypothesis prediction, each potential hypothesis generated by the present application is accompanied by a clear evidence path, which can be traced back to which local subgraph, based on which existing causal paths, and how to generate through recombination and inference rules.

[0046] (3) Realize the modeling and prediction of complex scientific hypotheses: Through the storage of five-tuple structure and edge attributes, the present application can effectively capture and utilize intermediates and regulatory effects, thereby generating new hypotheses containing these complex mechanisms, which are closer to real-world scientific problems.

[0047] (4) Provide a diversified and quantifiable evaluation system: The present application not only generates hypotheses, but also sorts the hypothesis relationships by combining the comprehensive credibility scores of network topology structure and relationship attributes, so that researchers can focus on the most likely established hypotheses first, saving time for screening and verification. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 is a schematic diagram of the present application. DETAILED DESCRIPTION

[0049] The technical solutions of the present application will be further described below in conjunction with the drawings.

[0050] As Figure 1 shown, the present application provides a potential hypothesis relationship prediction method based on causal symbolic network, including the following steps:

[0051] S1, Targeted domain academic literature collection and preprocessing. First, the target domain is selected as enterprise digital transformation. This domain is the intersection of disciplines such as econometrics, information technology, business management, and organizational behavior, and its academic papers usually contain explicit hypothesis and empirical research process, which is suitable as the data source of this method. Second, through manual retrieval combined with automated crawler program, relevant academic literature in this field is captured from mainstream academic databases such as Web of Science and Scopus. The title information (such as title, author, publication year, publication name, literature ID, etc.) and abstract of the paper are collected. To obtain more complete context information, the API interface of open academic platforms such as OpenAlex or Semantic Scholar is called to match and obtain the full text content of the target literature using the obtained title information. For PDF format full text, PyMuPDF or other Python libraries are used for parsing and converting it into pure text format. For HTML format web page full text, its text content is directly extracted through Elsevier API or other academic publisher API. Finally, all collected academic texts (abstracts and full texts) are cleaned to remove irrelevant headers, footers, reference lists, chart titles, and other noise data, and the cleaned texts are stored in a structured manner.

[0052] S2, Variable Hypothesis Relation Extraction Module Based on Large Language Model. Extracting hypothesis relations from academic literature requires first identifying the declarative sentences that may contain hypotheses. Academic text structures typically include an introduction, literature review (or theory & hypothesis), methods, results, and conclusions. The declarative sentences containing hypotheses are usually found in the hypothesis section, results section, and conclusion section of academic papers. The hypothesis section typically contains specific statements based on existing theories or logical reasoning about the expected relationships between variables, while the results section and conclusion section are more general, highlighting the assumed results and main findings of the academic text. The Qwen-max large language model from Alibaba is used to identify hypothesis relation declarative sentences in academic texts. To improve recognition performance, examples are added to the input prompt, and then LangChain is used for output type verification to ensure that the return is in structured JSON format. The structured hypothesis relation data generated from the previous step is input into the large language model, which uses semantic parsing capabilities to analyze the text content of each hypothesis relation in depth. The results are initially classified into positive relationships, negative relationships, moderating effects, and mediating effects, etc. according to the effect type. Subsequently, pre-designed customized rule templates are applied to verify and adjust the preliminary classification results. These templates are based on domain knowledge and common patterns of hypothesis relations, aiming to identify the characteristics and logical structures of specific types of relationships. Then, a logical reasoning module is used to further verify the relationship types, ensuring the logical consistency of the hypothesis relations and the rationality of the reasoning chain in the overall context. After semantic analysis, rule template matching, and logical reasoning verification, the system makes a final determination of the type of each hypothesis relation and outputs it in the form of a multi-tuple, with content including specific relationship types and related variable information.

[0053] S3, Multi-dimensional Causal Relation Discrimination. Each five-tuple relation extracted is subjected to validity discrimination. For example, for the <technological proximity, digital transformation maturity, positive, knowledge diversity, technological turbulence> five-tuple, the discrimination process is as follows: ① Domain consistency: Enterprises use existing technology to promote transformation, consistent with the Resource-Based View theory, discrimination passed. ② Direction rationality: Technological proximity is the cause, and transformation maturity is the effect, the direction is reasonable, discrimination passed. ③ Mediation rationality: Technological proximity → knowledge diversity → transformation maturity forms a logical and smooth causal chain (i.e., it is easier to absorb new knowledge with existing technology, and abundant knowledge promotes transformation), discrimination passed.

[0054] S4, Causal Symbol Network Construction. All causal relations in the digital transformation field of the enterprise that pass the verification are integrated into a global knowledge network GAll identified variables, such as AI technology investment, market response speed, supply chain management, data analysis capability, enterprise performance, organizational agility, etc., are instantiated as nodes in the network. According to the above example, a directed edge will be established between the AI technology investment node and the market response speed node, with the former pointing to the latter. Since the effect is positive, the symbol of this edge is set to +. Supply chain management (intermediate variable) and data analysis capability (moderating variable) are attached to this + directed edge as metadata. Write a program to repeat this process and finally build a symbolic network that describes the causal relationships between various variables in the field of enterprise digital transformation.

[0055] S5, Hypothesis-inspired search. In the causal symbolic network of enterprise digital transformation, an inspiration subgraph for generating new hypotheses is circled. The betweenness centrality of each node in the network is calculated. It may be found that the node centrality score of data-driven decision-making or customer experience is very high because they are key bridges connecting many different research topics. Select one of them, such as data-driven decision-making, as the target variable v ego . From the data-driven decision-making node, perform bidirectional neighborhood random walk. The path that the walk process may pass through. Stop when the walk reaches the preset depth or node number limit. All visited nodes, such as cloud computing adoption, business process automation, employee skill training, executive support, and data-driven decision-making, and their original edges together form a local inspiration subgraph about data-driven decision-making G sub .

[0056] S6, Potential hypothesis relationship prediction. In the local subgraph centered on v ego G sub is executed. First, all candidate variable nodes that are path-connected but not directly adjacent to G sub v ego are identified from v i , forming a candidate set U={ v 1, v 2, …, v i}; For each candidate variable v i U, the causal relationship between v ego and v i ​​the direction score between the two, and further, by multiplying the sign effect of all edges on the path, the cumulative effect of the path between the two is computed, which determines the sign (positive or negative) of the causal relationship in the new hypothesis, the computation of the cumulative effect follows the formula:

[0057] ;

[0058] where, e j is the j th edge on a certain specific path of the connection; effect( e j ) is the sign network of the edge, k is the length of the path. Finally, check whether there exists an edge with a mediating or moderating property on the path connecting v ego and v i to identify potential mediating and moderating variables in the new hypothesis, the computation of the direction score follows the formula:

[0059] ;

[0060] where, G sub denotes the set of intermediate nodes on all paths connecting v ego and v i ; is the size of the set, d ( x ) is a predefined node direction metric function, whose value indicates the advancing or retreating direction of the intermediate node in the path relative to v ego .

[0061] ​S7, Hypothesis relationship potential credibility ranking. Rank all the generated new hypotheses, for example, calculate the intersection of the neighbor set of data-driven decision-making and the neighbor set of organizational learning capability, it may be found that they are both directly connected to the two nodes of high-level management support and IT infrastructure, and the number of common neighbors is 2. If the decision quality repeatedly appears as a mediator in multiple different inference paths, it will have a higher frequency score as a mediator variable. The number of common neighbors (for example, counted as 2 points) and the frequency score of the mediator / moderator variable (for example, counted as 1.5 points) are integrated into a total score through a weighted formula (such as Score = 0.6 * CommonNeighbors + 0.4 * MetaScore). Compare the score of this hypothesis with other generated new hypotheses, and rank them from high to low according to the score.

Claims

1. A method for predicting potential hypothetical relationships based on a causal signed network, characterized by, Comprising the following steps: (1) Scientific hypothesis variable identification and multi-dimensional causal relationship discrimination: from unstructured scientific literature data sources, automatically identify and extract structured scientific hypothesis variables, and discriminate the causal relationship between variables in multiple dimensions; (2) Building a causal signed network: based on the effective causal relationship discriminated in step (1), a directed and signed causal signed network is constructed, in which the nodes represent scientific variables, the directed edges represent the direction of the causal relationship, and the sign of the edge represents the nature of the causal effect; (3) Hypothesis inspiration heuristic search: for one or more target variables in the network, a heuristic random walk strategy is used to explore the causal signed network, generating a local subgraph containing potential hypothesis inspiration; (4) Prediction of potential hypothesis relationship: within the generated local subgraph, a set of new candidate hypothesis relationships is generated by reorganizing the existing variables and their causal relationship paths; (5) Ranking the credibility of potential hypothesis relationships: design and calculate the comprehensive credibility score of each candidate hypothesis relationship, and sort the candidate hypothesis relationships in descending order according to the score, and finally output the list of potential hypothesis relationships. 2.The method of claim 1, wherein, Step (1) includes the following steps: (11) Use a large language model fine-tuned on scientific literature in a specific field to perform named entity recognition and relationship extraction on the input scientific paper full text or abstract text; (12) Automatically extract and formalize the structured five-tuple hypothesis relationship from the text by the large language model, the structure of the five-tuple is: <independent variable, dependent variable, effect, mediator variable, moderator variable>, where the effect refers to the impact of the independent variable on the dependent variable, and its value is positive or negative; (13) According to a set of preset causal discrimination dimensions, evaluate each variable relationship represented by the five-tuple; The preset causal discrimination includes the following dimensions: domain consistency, direction rationality, mechanism rationality, causal temporal sequence, mediation rationality, and adjustment appropriateness; (14) Evaluation is carried out by expert consultation or automatic rule logic, when a variable relationship meets at least two of the dimension standards, the variable relationship is confirmed as an effective causal relationship and used for subsequent network construction. 3.The method of claim 1, wherein, Step (2) includes the following steps: (21) All identified independent variables and dependent variables are instantiated as nodes in the network; (22) If there is an effective causal relationship between variables A and B, where A is the independent variable and B is the dependent variable, a directed edge is established between the node representing A and the node representing B, and the direction of the edge is from A to B; (23) According to the effect information in the five-tuple, assign a signed weight to the directed edge: if the effect is positive, the sign of the edge is defined as positive +; if the effect is negative, the sign of the edge is defined as negative -; (24) The information of the mediator variable and the moderator variable contained in the five-tuple is attached and stored as an attribute on the directed edge. 4.The method of claim 1, wherein, Step (3) includes the following steps: (31) by calculating and comparing at least one network centrality indicator of each node in the network, automatically identifying and selecting one or more target variables with high exploration value, denoted as v ego ; (32) from v ego starting, a heuristic random walk algorithm is performed on the network, wherein the next node v next to be visited in the walk is randomly selected from the bidirectional neighborhood of the current node v current ​ (33) the bidirectional neighborhood of the current node N bidirectional ( v current ) is defined as the union of the predecessor node set and the successor node set of the node, mathematically expressed as: ； Among them, Pre( v ) represents a node v The set of the ingress neighborhood, Suc( v ) represents a node v The set of out-neighbors.

5. The method of claim 4, wherein the method further comprises: The termination condition of the heuristic random walk is set to satisfy any one of the following conditions: the depth of the walk path, i.e. the number of edges passed from the starting point v ego , reaches a pre-set maximum depth threshold d max ; the total number of unique nodes visited during the walk process exceeds a pre-specified node number threshold n max ; when the walk terminates, all visited nodes and the original edges between them collectively form a local and structure-preserving subgraph G sub = G [ V visited ] 6. The method of claim 1, wherein the method is based on a causal signed network of latent hypothetical relationships. In step (4), the potential hypothesis relationship prediction is centered on a local subgraph v ego a local subgraph G sub is executed, including the following steps: (41) from G sub all candidate variable nodes that are connected by a path, but not directly adjacent v ego are identified v i , form a candidate set U={ v 1 , v 2 , …, v i} (42) For each candidate variable v i U, through calculation v ego and v i The directional metrics of nodes on all paths between them are used to infer the causal direction between them and quantify it into a direction score. (43) Calculate the cumulative effect of the path between the two by multiplying the signed effect of all edges on the path, thereby determining the sign of the causal relationship in the new hypothesis, i.e. positive or negative; (44) Check connections v ego With v i whether there are edges with mediating or moderating properties on the path of to identify potential mediating and moderating variables in the new hypothesis.

7. The method of claim 6, wherein the method further comprises: The calculation of the direction score follows the formula: ； wherein, denotes G sub intermediate nodes on all paths of v ego and v i ; is the size of the set, d ( x ) is a predefined node direction metric function whose value's sign indicates the intermediate node's advancing or retreating direction in the path with respect to v ego .

8. The method of claim 6, wherein the method further comprises: The calculation of the cumulative effect follows the formula: ； wherein, e j is the length of the path on which the connection is made; j is the length of the path on which the connection is made; e j is the symbolic network of edges, k is the length of the path on which the connection is made. 9.The method of claim 1, wherein, Step (5) includes the following steps: (51) Using the number of common neighbors as the primary criterion for assessing the structural proximity of two variables, the number is quantified as the target variable v ego of the candidate variable v i The intersection size of the neighbor sets of the target variable , where denotes the direct neighbor set of the node v ; (52) The frequency of the occurrence of the potential mediator and moderator variables identified by the candidate hypothesis is counted as an additional evaluation indicator; (53) The number of common neighbors and the frequency of the mediator and moderator variables are integrated into a comprehensive credibility score through a weighted summation function; (54) all candidate hypotheses h i ranked in descending order of their integrated credibility scores H ={ h 1 , h 1 ,…, h i}.

Images