A cross-document opinion relation reasoning method, a related method and device

Abstract

This invention relates to a method, related methods, and apparatus for cross-document opinion relationship reasoning. The method includes: after labeling each opinion document with paragraph types, performing sentence-level argument node extraction to obtain an argument node set and a directed reasoning edge set; converting each opinion document into a proof tree; extracting issue fingerprint vectors and calculating similarity for each proof tree to obtain paired proof trees; constructing a node semantic matching matrix; combining preset node constraints to solve for the optimal argument node pairing scheme to obtain a shared argument node set and a unique argument node set; locating divergent nodes and divergence types; determining the relationship type of each pair of paired proof trees based on the shared argument node set, the unique argument node set, and the preset relationship type; and comprehensively obtaining the relationship reasoning result for each pair of opinion documents. This method achieves alignment comparison of two proof trees across documents and accurately locates argument nodes where reasoning diverges.

Description

Technical Field

[0001] This invention relates to a method, related method, and apparatus for cross-document viewpoint relationship reasoning. Background Technology

[0002] In today's information-intensive decision-making environment, high-value texts such as think tank reports and authoritative news commentaries carry in-depth analytical perspectives from a wealth of important individuals (such as domain experts and policy advisors) and key enterprises (such as industry leaders and research institutions). These perspectives often involve complex logical relationships; for example, one expert might elaborate on another's argument, supplement a policy proposal from a different angle, or arrive at diametrically opposed conclusions based on the same economic data. Accurately identifying and demonstrating these deep relationships between perspectives has significant practical value for policy analysis, public opinion monitoring, and business decision-making.

[0003] Relation-based Argument Mining (RbAM) is the mainstream method in the field of argument analysis. This method uses a large language model combined with cue engineering (a technique that guides the model output by designing input cue templates) to classify the argument relationship between two texts into binary ("support" or "oppose") or ternary ("support", "oppose", or "neutral").

[0004] The Chain-of-Syllogisms (CoS) method simulates human syllogistic reasoning by breaking down a complete argument in a document into a series of interlocking "atomic syllogisms" (i.e., the smallest units of reasoning, each containing a major premise, a minor premise, and a local conclusion). The key design is to uniformly fold and classify the local conclusions of all intermediate reasoning steps into "analysis" nodes, and use them as input premises for the next level of reasoning, ultimately forming a proof tree in the form of a Directed Acyclic Graph (DAG). Summary of the Invention

[0005] To perform structural alignment and comparison of two proof trees from different documents and to accurately locate the argument nodes where the reasoning diverges, this invention provides a cross-document viewpoint relationship reasoning method, related methods, and apparatus.

[0006] In a first aspect, the present invention provides a method for cross-document viewpoint relationship reasoning, the method comprising:

[0007] Obtain multiple opinion documents and annotate each paragraph in each opinion document according to preset paragraph types to obtain opinion documents with annotated paragraph types;

[0008] For each paragraph of an opinion document after labeling the paragraph type, a pre-trained language model is used to extract sentence-level argument nodes according to the preset argument node type to obtain an argument node set for each opinion document. Then, a natural language inference model is used to determine the reasoning support relationship between any two argument nodes in the argument node set to obtain a set of directed reasoning edges for each opinion document. Based on the argument node set and the set of directed reasoning edges, each opinion document is converted into a proof tree.

[0009] Based on the set of argument nodes in each proof tree, an issue fingerprint vector is extracted from each proof tree, and the similarity between the issue fingerprint vectors of any two proof trees is calculated. Each pair of proof trees whose issue fingerprint vector similarity is greater than the set issue matching threshold is determined as a paired proof tree.

[0010] Based on preset rules, the matching score between any argument node of one proof tree and any argument node of the other proof tree in each pair of paired proof trees is calculated, and the node semantic matching matrix of each pair of paired proof trees is constructed.

[0011] Based on the node semantic matching matrix and the preset node constraints, the optimal argument node pairing scheme for each pair of paired proof trees is solved, resulting in the shared argument node set for each pair of paired proof trees and the unique argument node set for each proof tree in each pair of paired proof trees.

[0012] Based on the shared set of argument nodes and the unique set of argument nodes, the divergence nodes are located, and the set of divergence nodes is obtained.

[0013] Based on the set of branch nodes and the preset branch types, determine the branch type of all branch nodes;

[0014] Based on the shared set of argument nodes, the unique set of argument nodes, and the preset relationship type, determine the relationship type of each pair of paired proof trees;

[0015] Based on the optimal argument node pairing scheme, the shared argument node set, the unique argument node set, the divergence node set, the divergence type, and the relationship type, the relationship reasoning result for each pair of opinion documents is obtained.

[0016] In one or more optional embodiments of this application, the following further includes:

[0017] For the pre-trained language model, argument nodes are extracted independently multiple times from the same opinion document by using different random seeds or prompt templates, resulting in multiple argument node sets. The reasoning support relationship between any two argument nodes in each argument node set is determined by a natural language inference model, resulting in multiple directed reasoning edge sets. Based on the multiple argument node sets and the corresponding multiple directed reasoning edge sets, each opinion document is converted into multiple candidate proof trees.

[0018] For each argument node in each candidate proof tree, the consistency score of the argument node is calculated using the following formula:

[0019] ,

[0020] In the formula, This indicates the consistency score of the argument nodes. , These represent the argument nodes in two different candidate proof trees. , Representing the argument nodes respectively , semantic embedding vector, Represents the cosine similarity function. This represents the total number of candidate proof trees. This represents the consensus threshold for the argument nodes;

[0021] For each directed inference edge in each candidate proof tree, the consistency score of the directed inference edge is calculated using the following formula:

[0022] ,

[0023] In the formula, This represents the consistency score of directed inference edges. , Let represent the directed inference edges in two different candidate proof trees. , , , Representing the argument nodes respectively , , , semantic embedding vector, This represents the consistency threshold for directed inference edges;

[0024] Based on the set of argument nodes whose consistency scores reach the first threshold and the set of directed inference edges whose consistency scores reach the second threshold, each opinion document is transformed into a proof tree.

[0025] In one or more optional embodiments of this application, the argument node type includes premise node, data citation node, authoritative citation node, analysis node, and conclusion node.

[0026] Based on the set of argument nodes in each proof tree, an issue fingerprint vector is extracted from each proof tree, and the similarity between the issue fingerprint vectors of any two proof trees is calculated. Each pair of proof trees whose issue fingerprint vector similarity is greater than a set issue matching threshold is identified as a paired proof tree, including:

[0027] Based on the set of argument nodes in each proof tree, the issue fingerprint vector is extracted for each proof tree using the following formula:

[0028] ,

[0029] In the formula, This represents the issue fingerprint vector of any proof tree. This represents the set of conclusion nodes within the set of argument nodes. This represents the set of argument nodes in the argument node, which includes premise nodes, data citation nodes, and authoritative citation nodes. Indicates the argument node semantic embedding vector, Indicates the fusion weights;

[0030] The similarity of the issue fingerprint vectors of any two proof trees is calculated using the following formula. Each pair of proof trees whose issue fingerprint vector similarity is greater than a set issue matching threshold is identified as a paired proof tree:

[0031] ,

[0032] In the formula, This represents the similarity of the issue fingerprint vectors of any two proof trees. , This represents the issue fingerprint vector of any two proof trees.

[0033] In one or more optional embodiments of this application, the matching score between any argument node of one proof tree and any argument node of the other proof tree in each pair of paired proof trees is calculated based on preset rules, and a node semantic matching matrix for each pair of paired proof trees is constructed, including:

[0034] For any pair of paired proof trees ,in , , , , , Represent the proof tree respectively , The set of argument nodes, , Represent the proof tree respectively , Given a set of directed inference edges, calculate based on the following preset rules. Any argument node and Construct a pairwise proof tree based on the matching scores between any argument nodes. of Semantic matching matrix of order nodes :

[0035] Rule 1: If The i-th argument node and The Middle One argument node If all of them are data citation nodes or authoritative citation nodes, and cite the same data source or authoritative source, then... , express The i-th argument node and The Middle One argument node Match score between them;

[0036] Rule 2: If The i-th argument node and The Middle One argument node If the argument nodes are of the same type but do not satisfy the conditions of Rule 1, then... ,in, , Representing the argument nodes respectively , semantic embedding vector;

[0037] Rule 3: If The i-th argument node and The Middle One argument node The types of argument nodes are different, then ,in, Indicates the type attenuation factor.

[0038] In one or more optional embodiments of this application, based on the node semantic matching matrix and preset node constraints, the optimal proof node pairing scheme for each pair of proof trees is solved, including:

[0039] Define any pair of paired proof trees The set of argument node pairings is: ,in and It can appear in at most one argument node pairing;

[0040] Solve any pair of proof trees based on the following formula and node constraints. The optimal pairing scheme for argument nodes is obtained for each pair of paired proof trees:

[0041] ,

[0042] Constraint 1: When At that time, the argument node pairing Included in the matching scheme, among which The pairing threshold;

[0043] Constraint 2: If there are two argument nodes paired in the pairing scheme. and And in There exists a directed reasoning edge. Then in There is also a source from arrive Directed reasoning edges;

[0044] Constraint 3: Argument node pairings that conform to Rule 1 are included in the pairing scheme.

[0045] Secondly, the present invention provides a method for constructing a cross-document viewpoint relationship graph, comprising:

[0046] Obtain the relational reasoning results between every two opinion documents obtained according to the method described in the first aspect above;

[0047] Based on the aforementioned relational reasoning results, a three-layer nested viewpoint relational graph is constructed, wherein:

[0048] The first layer is the entity relationship layer. Nodes are the personnel or organizational entities corresponding to the opinion document, and edges are the overall stance relationships between entities. Based on the relationship reasoning results of all involved entities, the overall stance relationship is scored. When the score of the overall stance relationship is greater than zero, it is marked as overall consistency; otherwise, it is marked as overall opposition.

[0049] The second layer is the viewpoint relationship layer, where nodes are the proof trees corresponding to each entity, and edges represent the relationship types between the proof trees.

[0050] The third layer is the argument structure layer. The nodes are the argument nodes in the proof tree, and the edges include directed reasoning edges of the same proof tree and paired edges of argument nodes of different proof trees obtained based on the shared set of argument nodes.

[0051] Thirdly, the present invention provides a cross-document viewpoint relationship reasoning device, comprising:

[0052] The paragraph type annotation module is used to acquire multiple opinion documents and annotate each paragraph in each opinion document according to preset paragraph types, so as to obtain opinion documents with annotated paragraph types.

[0053] The proof tree conversion module is used to extract sentence-level argument nodes for each paragraph of each opinion document after each paragraph is labeled with a paragraph type, using a pre-trained language model according to the preset argument node type, to obtain an argument node set for each opinion document. Then, it uses a natural language inference model to determine the reasoning support relationship between any two argument nodes in the argument node set, to obtain a set of directed reasoning edges for each opinion document. Based on the argument node set and the set of directed reasoning edges, each opinion document is converted into a proof tree.

[0054] The proof tree pairing module is used to extract the issue fingerprint vector of each proof tree based on the set of argument nodes in each proof tree, calculate the similarity of the issue fingerprint vectors of any two proof trees, and determine each pair of proof trees whose issue fingerprint vector similarity is greater than a set issue matching threshold as a paired proof tree.

[0055] The node semantic matching matrix construction module is used to calculate the matching score between any argument node of one proof tree and any argument node of the other proof tree in each pair of paired proof trees based on preset rules, and to construct the node semantic matching matrix for each pair of paired proof trees.

[0056] The optimal argument node pairing solution module is used to solve the optimal argument node pairing solution for each pair of paired proof trees based on the node semantic matching matrix and preset node constraints, so as to obtain the shared argument node set of each pair of paired proof trees and the unique argument node set of each proof tree in each pair of paired proof trees.

[0057] The divergence node location module is used to locate divergence nodes and obtain a set of divergence nodes based on the shared set of argument nodes and the unique set of argument nodes.

[0058] The branching type determination module is used to determine the branching type of all branching nodes based on the set of branching nodes and a preset branching type.

[0059] The relationship type determination module is used to determine the relationship type of each pair of paired proof trees based on the shared set of proof nodes, the unique set of proof nodes, and the preset relationship type.

[0060] The relational reasoning result module is used to obtain the relational reasoning result for each pair of opinion documents based on the optimal argument node pairing scheme, the shared argument node set, the unique argument node set, the divergence node set, the divergence type, and the relation type.

[0061] Fourthly, the present invention provides a cross-document viewpoint relationship graph construction apparatus, comprising:

[0062] The relational reasoning result acquisition module is used to acquire the relational reasoning result between every two opinion documents obtained according to the method described in the first aspect above.

[0063] The relationship graph construction module is used to construct a three-level nested viewpoint relationship graph based on the relationship reasoning results, wherein:

[0064] The first layer is the entity relationship layer. Nodes are the personnel or organizational entities corresponding to the opinion document, and edges are the overall stance relationships between entities. Based on the relationship reasoning results of all involved entities, the overall stance relationship is scored. When the score of the overall stance relationship is greater than zero, it is marked as overall consistency; otherwise, it is marked as overall opposition.

[0065] The second layer is the viewpoint relationship layer, where nodes are the proof trees corresponding to each entity, and edges represent the relationship types between the proof trees.

[0066] The third layer is the argument structure layer. The nodes are the argument nodes in the proof tree, and the edges include directed reasoning edges of the same proof tree and paired edges of argument nodes of different proof trees obtained based on the shared set of argument nodes.

[0067] Fifthly, embodiments of the present invention provide a computer-readable storage medium having a computer program / instruction stored thereon, which, when executed by a processor, implements the cross-document viewpoint relationship reasoning method as described in the first aspect above, and / or the cross-document viewpoint relationship graph construction method as described in the second aspect above.

[0068] In a sixth aspect, embodiments of the present invention provide a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the cross-document viewpoint relationship reasoning method as described in the first aspect above, and / or the cross-document viewpoint relationship graph construction method as described in the second aspect above.

[0069] The beneficial effects of the above-described technical solutions provided in the embodiments of the present invention include at least the following:

[0070] This invention provides a cross-document opinion relationship reasoning method. Each opinion document is labeled with paragraph types. For each paragraph of the labeled opinion document, sentence-level argument nodes are extracted according to preset argument node types to obtain an argument node set for each opinion document. The reasoning support relationship between any two argument nodes is determined to obtain a set of directed reasoning edges for each opinion document. Based on the argument node set and the directed reasoning edge set, each opinion document is converted into a proof tree. An issue fingerprint vector is extracted from each proof tree, and the similarity between the issue fingerprint vectors of any two proof trees is calculated. Proof trees with a similarity greater than a similarity threshold are identified as paired proof trees. Based on the matching score between the two argument nodes in each paired proof tree, a method is constructed for each paired proof tree. The method uses the semantic matching matrix of the nodes of the proof tree, combined with preset node constraints, to solve for the optimal argument node pairing scheme (i.e., the maximum common reasoning subgraph) for each pair of proof trees. This yields the shared argument node set for each pair of proof trees and the unique argument node set for each proof tree within each pair. It then locates divergent nodes and obtains their sets. Combining preset divergence types, it determines the divergence type of all divergent nodes. Based on the shared argument node set, the unique argument node set, and the preset relationship type, it determines the relationship type of each pair of proof trees. Finally, it integrates the optimal argument node pairing scheme, the shared argument node set, the unique argument node set, the divergence node set, the divergence type, and the relationship type to obtain the relationship reasoning result for each pair of opinion documents. This invention enables structural-level alignment comparison of two proof trees from different documents and accurately locates argument nodes where reasoning diverges. It solves the problems of existing opinion reasoning methods, such as coarse "black box" classification and lack of cross-document proof tree alignment and relationship reasoning functions. It achieves fine-grained (sentence-level) deconstruction and high-precision alignment of complex cross-document opinion trees. Specifically, this invention has the following beneficial effects:

[0071] (1) By constructing a node semantic matching matrix for the pairing proof tree, the optimal argument node pairing scheme (maximum common reasoning subgraph) is solved under the structure preservation constraint, realizing the leap from "black box relationship label" to "structured divergence location", which can accurately output the specific reasoning node location and divergence type (such as factual divergence, reasoning divergence, etc.) of two viewpoints that diverge.

[0072] (2) By extracting the issue fingerprint vector for similarity calculation and pairing the argument nodes, the single-document argument structure is extended to cross-document structural comparison, which fills the technical gap of single-document proof trees (such as the CoS framework) lacking cross-document structural comparison capability. Attached Figure Description

[0073] The accompanying drawings are provided to further illustrate the technical solutions of this application and constitute a part of the specification. They are used together with the embodiments of this application to explain the technical solutions of this application, and do not constitute a limitation on the technical solutions of this application. In the accompanying drawings:

[0074] Figure 1 A schematic flowchart of the cross-document viewpoint relationship reasoning method provided in this application embodiment;

[0075] Figure 2 A schematic diagram illustrating the structure of converting a single document into a proof tree, as provided in this application embodiment;

[0076] Figure 3 A schematic diagram of cross-document proof tree alignment and branch node positioning provided for embodiments of this application;

[0077] Figure 4 A schematic diagram of the method for constructing a cross-document viewpoint relationship reasoning graph provided in this application embodiment;

[0078] Figure 5 A schematic diagram of a three-layer nested viewpoint relationship provided for embodiments of this application;

[0079] Figure 6 This application provides a schematic diagram illustrating the evolution of viewpoint camps and tracing chains in its embodiments.

[0080] Figure 7 A schematic diagram of a cross-document viewpoint relationship reasoning device provided in an embodiment of this application;

[0081] Figure 8 A schematic diagram of a cross-document viewpoint relationship graph construction device provided in an embodiment of this application. Detailed Implementation

[0082] Embodiments of this application will now be described in more detail with reference to the accompanying drawings. However, this description is exemplary and not restrictive, and those skilled in the art can implement this application in various forms without being limited to the embodiments set forth herein. These embodiments are provided to enable a more thorough understanding of this disclosure and to fully convey the scope of this disclosure to those skilled in the art.

[0083] The embodiments, features, and elements disclosed in this application can also be combined with any conventional features or elements to form a unique inventive scheme as defined by the claims. Any feature or element of any embodiment can also be combined with features or elements from other inventive schemes to form another unique inventive scheme as defined by the claims. Therefore, it should be understood that any feature shown and / or discussed in this application can be implemented individually or in any suitable combination.

[0084] To facilitate understanding of the technical solutions of the embodiments of this application, some terms or concepts involved in the embodiments of this application will be briefly described first.

[0085] 1. Proof Tree: A data model consisting of proof nodes and directed reasoning edges connecting the proof nodes, used to represent the logical proof process in a structured and visual way.

[0086] 2. Pre-trained Language Model (PLM): This refers to a type of language representation model that is pre-trained on a large-scale unlabeled text corpus, such as BERT, RoBERTa, or ChatGPT models.

[0087] 3. Natural Language Inference (NLI) model: This is an AI model used to determine the logical relationship between two texts, including implication, contradiction, or irrelevant relationships.

[0088] 4. Issue fingerprint vector: It is a digital vector representation extracted from the argument nodes to accurately characterize and identify the semantic features of the core issue. It is the unique identity ID of the issue.

[0089] 5. Data source identifier: This is a metadata tag attached to each argument node (D-type or A-type node) that uniquely identifies the specific data source referenced by that node.

[0090] 6. Maximum Common Inference Subgraph: Find a subgraph in each of the two proof trees such that the two subgraphs are as consistent as possible in terms of node matching and edge structure, and the total matching score is maximized.

[0091] 7. Branch node: An argument node that belongs to the shared set of argument nodes, and in at least one proof tree, the direct successor argument node of the argument node belongs to the unique set of argument nodes of that tree.

[0092] The inventor referenced existing methods for opinion reasoning, including Relational Argument Mining (RbAM) and Syllogism Chain Method. Relational Argument Mining Method treats two opinion segments as an inseparable whole text for end-to-end classification, which is a "black box" classification. It can only output the final relationship label (such as "support" or "oppose") and cannot reveal at which specific reasoning stage the two opinion segments agree or disagree. For example, when two think tank experts hold opposing views on the same economic policy, RbAM can only determine that they are "opposing" but cannot analyze whether the root cause of the opposition is based on different economic data or a difference in the interpretation of the same data. Moreover, the relationship types of this method are too coarse (only 2 to 3 types), which cannot meet the needs of think tank analysis scenarios for fine characterization of opinion relationships. In the syllogistic chain method, the CoS framework only constructs proof trees for single documents, structuring the argumentation within a document into a DAG form. When it is necessary to analyze the relationship between the opinions of multiple experts or institutions in different documents, the CoS framework lacks the ability to perform structural comparison of two or more proof trees from different documents. In other words, CoS can achieve high-quality single-document argumentation structure representation, but it does not have the function of cross-document proof tree alignment and relational reasoning. To at least partially address these issues, the inventors attempted to transform a single document into a proof tree and align and compare two proof trees across documents to evaluate the relationship type of cross-document viewpoints and locate divergence nodes. The research found that by extracting issue fingerprint vectors and calculating similarity for each tree, paired proof trees with the same issue can be found. A node semantic matching matrix is ​​constructed for each pair of paired proof trees. Combined with preset node constraints, the optimal argument node pairing scheme (maximum common reasoning subgraph) for each pair of paired proof trees can be solved, resulting in a shared set of argument nodes for each pair of paired proof trees and a unique set of argument nodes for each proof tree within each pair of paired proof trees. This allows for the location of divergence nodes and divergence types. Based on the shared set of argument nodes, the unique set of argument nodes, and the preset relationship type, the relationship type of each pair of paired proof trees is determined, and the relationship reasoning result for each pair of viewpoint documents is obtained. Based on this, the present invention provides a cross-document viewpoint relationship reasoning method, related methods, and apparatus.

[0093] Example 1

[0094] Embodiment 1 of this invention provides a method for cross-document viewpoint relationship reasoning, referring to... Figure 1 As shown, the method includes the following steps:

[0095] S101: Obtain multiple opinion documents and annotate each paragraph in each opinion document according to the preset paragraph types to obtain opinion documents with annotated paragraph types.

[0096] S102: For each paragraph of the opinion document after each paragraph type is marked, sentence-level argument nodes are extracted using a pre-trained language model according to the preset argument node type to obtain the argument node set of each opinion document. The reasoning support relationship between any two argument nodes in the argument node set is determined by a natural language inference model to obtain the directed reasoning edge set of each opinion document. Based on the argument node set and the directed reasoning edge set, each opinion document is converted into a proof tree.

[0097] S103: Based on the set of argument nodes in each proof tree, extract the issue fingerprint vector for each proof tree, calculate the similarity of the issue fingerprint vectors of any two proof trees, and determine each pair of proof trees whose issue fingerprint vector similarity is greater than the set issue matching threshold as a paired proof tree.

[0098] S104: Calculate the matching score between any argument node of one proof tree and any argument node of the other proof tree in each pair of paired proof trees based on preset rules, and construct the node semantic matching matrix of each pair of paired proof trees.

[0099] S105: Based on the node semantic matching matrix and the preset node constraints, solve the optimal argument node pairing scheme for each pair of paired proof trees to obtain the shared argument node set for each pair of paired proof trees and the unique argument node set for each proof tree in each pair of paired proof trees.

[0100] S106: Based on the shared set of argument nodes and the unique set of argument nodes, locate the divergent nodes and obtain the set of divergent nodes.

[0101] S107: Based on the set of branch nodes and the preset branch types, determine the branch type of all branch nodes.

[0102] S108: Based on the shared set of argument nodes, the unique set of argument nodes, and the preset relationship type, determine the relationship type of each pair of paired proof trees.

[0103] S109: Based on the optimal argument node pairing scheme, the shared argument node set, the unique argument node set, the divergence node set, the divergence type, and the relationship type, the relationship reasoning result for each pair of opinion documents is obtained.

[0104] The method of this invention can perform structural alignment comparison of two proof trees from different documents and accurately locate the argument nodes where the reasoning diverges. It solves the problems of the existing viewpoint reasoning, such as the coarse "black box" classification and lack of cross-document proof tree alignment and relational reasoning functions. It realizes fine-grained (sentence-level) deconstruction and high-precision alignment of complex viewpoints across documents. Specifically, it includes: (1) constructing a node semantic matching matrix for paired proof trees and solving the optimal argument node pairing scheme (maximum common reasoning subgraph) under the structural preservation constraint. This realizes the leap from "black box relation label" to "structured divergence location" and can accurately output the specific reasoning node positions and divergence types (such as factual divergence, reasoning divergence, etc.) where two viewpoints diverge; (2) extracting issue fingerprint vectors for similarity calculation and performing argument node matching. This realizes the expansion of the single-document argument structure to cross-document structural comparison and fills the technical gap of single-document proof trees (such as the CoS framework) lacking cross-document structural comparison capabilities.

[0105] In this embodiment, the preset paragraph types in step S101 include background statement type, factual data type, analysis and reasoning type, opinion and claim type, and policy recommendation type. Multiple opinion documents are usually obtained from think tank reports and authoritative news commentaries. For each opinion document, the chapter structure is first identified. Opinion documents in think tank reports usually have a structure of "abstract-background-analysis-conclusion-recommendation". Opinion documents in authoritative news commentaries usually have a structure of "introduction-background-viewpoints-comment". Then, a rule-based paragraph function classifier is used to label each paragraph in each opinion document as one of the above paragraph types, so as to obtain each opinion document after labeling the paragraph types.

[0106] In this embodiment, the argument node types in step S102 include Premise Node (P-type node), Data Reference Node (D-type node), Authority Reference Node (A-type node), Analysis Node (N-type node), and Conclusion Node (C-type node). P-type nodes express a factual assumption or widely accepted background knowledge as the starting point for reasoning, such as: "The current global supply chain landscape is undergoing profound restructuring"; D-type nodes cite specific statistical data, economic indicators, or quantitative facts, such as: "The year-on-year GDP growth rate in the third quarter of 2025 was 6.2%"; A-type nodes cite the views of other experts, institutions, or literature as arguments, such as: "The International Monetary Fund pointed out in its latest report..."; N-type nodes are inferences, analyses, or interpretations based on premises and evidence, serving as crucial nodes in the proof tree, transforming premises and evidence into intermediate judgments that support the conclusion through logical deduction, such as: "This indicates that the domestic demand-driven growth model has been initially established"; C-type nodes express the document's final core proposition or policy recommendation, such as: "Therefore, a prudent monetary policy stance should be maintained at this stage."

[0107] The pre-trained language model can be BERT, RoBERTa, ChatGPT, etc., capable of extracting semantically complete sentences from paragraphs as argument nodes and matching them as P-type, D-type, A-type, N-type, or C-type nodes. Typically, background statement paragraphs mainly extract P-type nodes, factual data paragraphs mainly extract D-type nodes, and analytical reasoning, opinion / claim, and policy recommendation paragraphs (collectively referred to as opinion-carrying paragraphs) can extract all types of nodes: P-type, D-type, A-type, N-type, and C-type. The natural language inference model can be RoBERTa, DeBERTa, etc., capable of determining whether there is an implicit, contradictory, or irrelevant relationship between two texts, and for any opinion document's set of argument nodes... Any two argument nodes and If the inference model determines for Provided logical support (i.e.) The content implies or supports (from the derivation), then a directed reasoning edge is established. This leads to the set of directed reasoning edges. Convert each opinion document into a proof tree. .

[0108] Since the extraction of argument nodes and the determination of directed inference edges in the proof tree mainly rely on the output of the language model, and the language model may generate content that is not faithful to the original text (i.e., the "illusion" phenomenon), directly using an unverified proof tree for subsequent cross-document alignment may lead to error propagation. Therefore, in this embodiment, a self-consistency verification step is added after the proof tree is constructed to ensure the structural reliability of the proof tree. The specific self-consistency verification process is as follows:

[0109] For the pre-trained language model, argument nodes are extracted independently multiple times from the same opinion document by using different random seeds or prompt templates, resulting in multiple argument node sets. The reasoning support relationship between any two argument nodes in each argument node set is determined by a natural language inference model, resulting in multiple directed reasoning edge sets. Based on the multiple argument node sets and the corresponding multiple directed reasoning edge sets, each opinion document is converted into multiple candidate proof trees.

[0110] For each argument node in each candidate proof tree, the consistency score of the argument node is calculated using the following formula:

[0111] ,

[0112] In the formula, This indicates the consistency score of the argument nodes. , These represent the argument nodes in two different candidate proof trees. , Representing the argument nodes respectively , semantic embedding vector, Represents the cosine similarity function. This represents the total number of candidate proof trees. This represents the argument node consistency threshold, used to determine whether the argument nodes of two candidate proof trees are the same argument node, and can take a value of 0.85.

[0113] For each directed inference edge in each candidate proof tree, the consistency score of the directed inference edge is calculated using the following formula:

[0114] ,

[0115] In the formula, This represents the consistency score of directed inference edges. , Let represent the directed inference edges in two different candidate proof trees. , , , Representing the argument nodes respectively , , , semantic embedding vector, This represents the directed reasoning edge consistency threshold, used to determine whether the directed reasoning edges of two candidate proof trees are the same directed reasoning edge. It can take a value of 0.85.

[0116] Based on the set of argument nodes whose consistency scores reach the first threshold and the set of directed inference edges whose consistency scores reach the second threshold, each opinion document is transformed into a proof tree.

[0117] The above self-consistency verification process can be understood as: in In the independently constructed candidate proof trees, the candidate proof trees are pruned by calculating the consistency scores of the proof nodes and the consistency scores of the directed reasoning edges, retaining only the nodes that are not explicitly defined. Argumentation nodes and Directed reasoning edge ( , The first and second score thresholds are respectively (which can be 0.6), meaning that only argument nodes and directed reasoning edges that are reproduced in at least 60% of the independent constructions are retained. Finally, a proof tree is constructed using the set of retained argument nodes and directed reasoning edges.

[0118] In this embodiment, in step S103, the issue fingerprint vector can be extracted for each proof tree using the following formula:

[0119] ,

[0120] In the formula, This represents the issue fingerprint vector of any proof tree. This represents the set of conclusion nodes within the set of argument nodes. This represents the set of argument nodes in the argument node, which includes premise nodes, data citation nodes, and authoritative citation nodes. Indicates the argument node semantic embedding vector, This represents the fusion weight, with a value between 0 and 1, used to control the proportion of conclusion semantics and argument semantics in the issue fingerprint.

[0121] The similarity of the issue fingerprint vectors of any two proof trees can be calculated using the following formula:

[0122] ,

[0123] In the formula, This represents the similarity of the issue fingerprint vectors of any two proof trees. , This represents the issue fingerprint vector of any two proof trees.

[0124] when hour( A threshold is set for the issue (a value of 0.65 is acceptable). Add to the set of paired proof trees .

[0125] In this embodiment, the node semantic matching matrix of each pair of paired proof trees in step S104 is constructed in the following manner:

[0126] For any pair of paired proof trees ,in , , , , , Represent the proof tree respectively , The set of argument nodes, , Represent the proof tree respectively , Given a set of directed inference edges, calculate based on the following preset rules. Any argument node and Construct a pairwise proof tree based on the matching scores between any argument nodes. of Semantic matching matrix of order nodes:

[0127] Rule 1: If The i-th argument node and The Middle One argument node If all of them are data citation nodes or authoritative citation nodes, and cite the same data source or authoritative source, then... , express The i-th argument node and The Middle One argument node Match score between them;

[0128] Rule 2: If The i-th argument node and The Middle One argument node If the argument nodes are of the same type but do not satisfy the conditions of Rule 1, then... ,in, , Representing the argument nodes respectively , semantic embedding vector;

[0129] Rule 3: If The i-th argument node and The Middle One argument node The types of argument nodes are different, then ,in, This represents the type attenuation factor, which takes a value between 0 and 1, with a possible value of 0.3.

[0130] Rule 1 above can be understood as: if a pairing proof tree If two argument nodes are a citation anchor pair, then the element of the semantic matching matrix corresponding to the citation anchor pair is assigned a value of 1. The specific process for determining the citation anchor pair is as follows: traverse... and All D-type nodes (data citation nodes) and A-type nodes (authoritative citation nodes), if and If two argument nodes in a document cite the same data source or authoritative source (determined by exact matching of the data source identifier, which remains consistent across different opinion documents, such as using a standardized source name), then this pair of argument nodes is marked as a citation anchor pair. The set of citation anchor pairs can be represented as: , Represents a node The data source identifier.

[0131] The significance of setting citation anchor pairs lies in the fact that think tank reports extensively cite common public data (such as government statistical bulletins and international organization reports). These common citations serve as natural structural alignment reference points. Using citation anchor pairs as matching rules for solving the optimal argument node pairing scheme (maximum common reasoning subgraph) can significantly improve the accuracy of aligning different proof trees and reduce computational complexity. The citation anchor alignment mechanism designed in this invention utilizes extracted underlying factual data sources and authoritative citation identifiers as hard constraints for structural alignment. This solves the matching uncertainty and inaccuracy problems of pure semantic matching when facing complex textual expressions, effectively avoiding semantic mismatches caused by changes in textual wording, and significantly improving the accuracy and computational efficiency of the alignment algorithm in actual think tank report analysis.

[0132] Rule 1 is a hard matching rule for citation anchors. Its function is to treat two paired proof trees as a deterministic match when they cite the same data source or authoritative source, regardless of the wording of the argument node text. This ensures that the commonly cited objective facts and authoritative evidence are correctly aligned, providing a reliable anchor for subsequent structural comparisons of proof trees. Rule 2 is a semantic matching rule for nodes of the same type. Its function is to perform a fine-grained semantic comparison of argument nodes of the same type, identifying argument nodes with similar semantics despite different wording. Rule 3 is a cross-type node attenuation matching rule. Its function is to significantly reduce the probability of incorrect matching of argument nodes of different types while not completely eliminating the possibility of cross-type argument node matching. For example, a statement used as a premise in one tree may be used as an analytical judgment in another tree. Allowing a certain degree of cross-node type matching helps to capture the shift in reasoning roles.

[0133] In this embodiment, solving the maximum common inference subgraph for each pair of paired proof trees can be transformed into solving the optimal proof node pairing scheme for each pair of paired proof trees in step S105, and is solved in the following manner:

[0134] Define any pair of paired proof trees The set of argument node pairings is: ,in and It can appear in at most one argument node pairing;

[0135] Solve any pair of proof trees based on the following formula and node constraints. Optimal argument node pairing scheme:

[0136] ,

[0137] Constraint 1: When At that time, the argument node pairing Included in the matching scheme, among which This is the pairing threshold, used to filter out paired argument nodes with insufficient semantic similarity; it can take a value of 0.55.

[0138] Constraint 2: If there are two argument nodes paired in the pairing scheme. and And in There exists a directed reasoning edge. Then in There is also a source from arrive Directed reasoning edges (i.e.) It can be reached through one or more steps of directed reasoning. );

[0139] Constraint 3: Argument node pairings that conform to Rule 1 are included in the pairing scheme.

[0140] The first constraint is a matching threshold constraint, which ensures that only semantically similar argument nodes are considered aligned (paired) nodes, avoiding incorrect matching of irrelevant content. The second constraint is a structure preservation constraint, expressed as follows: ,in In the proof tree From the node The set of all argument nodes reachable along the directed reasoning edge from the starting point serves to ensure that the alignment (pairing) result follows the reasoning direction and logical hierarchy of the two proof trees. That is, if argument node A supports argument node B in one tree, then the argument node paired with A in the other tree should also be able to directly or indirectly support the argument node paired with B. The third constraint mentioned above is a mandatory constraint on citation anchors, which ensures that commonly cited objective data and authoritative sources are not omitted during the alignment (matching) process.

[0141] The specific solution process for the optimal argument node pairing scheme (maximum common reasoning subgraph) is as follows:

[0142] First, construct a node compatibility graph. , For a set of node pairs, Let this be the set of reasoning edges. Nodes in the node-compatible graph correspond to pairs of argument nodes that satisfy constraint condition one, i.e. The node weight is the corresponding matching score. A node is compatible with two nodes in the graph. and There exists a reasoning edge if and only if: (a) and (a) One-to-one matching; (b) Satisfying constraint condition two (structural preservation constraint).

[0143] Then, the maximum weight clique is solved on the node compatibility graph. The maximum weight clique is the optimal proof node pairing scheme that satisfies all constraints. For cases containing citation anchor pairs, the argument nodes corresponding to the citation anchor pairs are forcibly added to the clique, and the search space is pruned accordingly. Specifically, the following three-layer search space pruning strategy is adopted to ensure computational efficiency:

[0144] Pruning Strategy 1 (Fixed Anchor Pruning): Citation anchor pairs are forcibly included in the matching scheme. Therefore, in the node compatibility graph, candidate node pairs that conflict with citation anchor pairs (i.e., other candidate pairs that share the same source or target node with the citation anchor) are directly removed. Let the number of citation anchor pairs be... This strategy can reduce the size of the node compatibility graph by approximately [amount missing] on average. A number of candidate node pairs.

[0145] Pruning Strategy Two (Type-Level Pruning): Since Constraint Two (Structure Preservation Constraint) requires consistent reasoning directions, and the type-level constraint of the argument nodes limits the reasoning edges to only point from lower levels to higher levels, we can solve the problem in segments according to the level—first align the first-level nodes (P-type / D-type / A-type), then align the second-level nodes (N-type) under the existing alignment constraints, and finally align the third-level nodes (C-type). The solution space of each level is much smaller than the global solution.

[0146] Pruning Strategy 3 (Neighborhood Constraint Pruning): For candidate node pairs that are not citation anchors, if none of their neighboring nodes in either proof tree satisfy the matching threshold, the candidate node pair can be safely removed because it cannot satisfy the structure preservation constraint together with other paired proof nodes.

[0147] After the above three-layer pruning, for a typical think tank report proof tree (the number of nodes is usually in the range of 15 to 50), the actual effective size of the node compatibility graph is usually compressed to the order of ten nodes, and the solution of the maximum weight clique can be completed in seconds.

[0148] In this embodiment, the solution to the optimal argument node pairing scheme (maximum common reasoning subgraph) in step S105 divides the argument nodes of the two proof trees into three disjoint sets:

[0149] (1) Shared set of argument nodes : Appears in the best matching scheme The argument nodes in the argument represent the reasoning basis shared by the two arguments.

[0150] (2) The unique set of argument nodes : Unmatched argument nodes, i.e. , representing only existing The reasoning content in the argument.

[0151] (3) Unique set of argument nodes : Unmatched argument nodes, expressed in the same formula as above, represent those that only exist in [the following context]. The reasoning content in the argument.

[0152] In this embodiment, the branching node in step S106 is the precise location where the two proof trees diverge. Before the branching node (including the branching node itself), the two proof trees follow the same reasoning path. Starting from the next step after the branching node, the two proof trees introduce different analyses or reasoning, which may ultimately lead to different conclusions.

[0153] The divergence node is expressed as: (Assume it is in) Chinese correspondence ,exist Chinese correspondence The following conditions must be met: ,in, In the proof tree Middle argumentation node The set of direct successor argument nodes (i.e. (A proof node directly supported by a directed reasoning edge). The set of all branching nodes is: .

[0154] In this embodiment, in step S107, for each branch node, the branch type is determined based on its structural characteristics in the subsequent branch regions of the two proof trees, combined with the following branch types:

[0155] Type 1 Disagreement: Factual Divergence

[0156] Judgment criteria: branching nodes It is an N-type (analysis) node, and in and In this context, the subsequent argument nodes of this divergence node are mainly different D-type (data citation) nodes or P-type (premise) nodes, expressed as follows:

[0157] ,

[0158] in, Indicates in From the point of divergence Starting from and reachable along the reverse reasoning edge belongs to The set of argument nodes (i.e.) Unique, for (Provide supporting argumentation nodes) The threshold for determining factual disagreement can be 0.6, and N, D, and P represent the types of argument nodes.

[0159] The meaning of Type 1 Disagreement: Two proof trees reason based on different factual basis or data sources. For example, expert A (one proof tree) cites domestic market data, while expert B (another proof tree) cites international market data, leading to differences in subsequent arguments.

[0160] Type 2 of Disagreement: Reasoning Divergence

[0161] Judgment criteria: branching nodes For P-type (prerequisite), D-type (data citation), or A-type (authoritative citation) nodes, and in and In this context, the subsequent argument nodes of this divergence node are mainly different N-type (analysis) nodes, expressed as:

[0162] ,

[0163] in, Indicates the divergence node type Indicates in Middle Divergence Node The subsequent argument nodes belong to The set of argument nodes, This represents the threshold for determining reasoning divergence, and can take a value of 0.5.

[0164] The second type of disagreement means that two proof trees are based on the same (or highly similar) factual foundation, but disagreements arise in the interpretation and analysis. For example, two experts (two proof trees) cite the same GDP growth data, but one believes "this indicates a healthy economic fundamentals," while the other believes "this masks structural contradictions." Reasoning disagreements are the most informationally valuable type of disagreement in think tank analysis because they reveal fundamental differences in experts' cognitive frameworks and analytical methodologies.

[0165] Type 3 of Disagreement: Value Divergence

[0166] Judgment criterion: Shared set of argument nodes The coverage is high (i.e., the reasoning processes of the two proof trees are mostly consistent), but the positions of the conclusion nodes are opposite, and the expression is:

[0167] ,

[0168] in, This represents the high coverage threshold, and can take a value of 0.7. This represents the position similarity between the sets of conclusion nodes of two proof trees (the range of values ​​is...). -1 indicates complete opposition, +1 indicates complete agreement. This represents the negative position threshold, which can take a value of -0.3.

[0169] Similarity of the above positions The calculation method is as follows:

[0170] (1) Construction of the conclusion text:

[0171] Set the conclusion nodes of the two proof trees respectively. and The text content is pieced together to form a complete conclusion statement. and .

[0172] (2) NLI model inference:

[0173] Will and Input the above Natural Language Inference (NLI) model, and the model outputs predicted probabilities in two dimensions: The probability implied by the two parties having the same stance, with a value range of [value missing]. , The probability of opposing / contradictory positions between the two, with a range of values. .

[0174] (3) Calculation of position similarity:

[0175] use and The difference will map the result to The interval is calculated using the following formula: .

[0176] Value description:

[0177] Approaching +1: Explanation Extremely high Extremely low indicates that the two conclusions are completely consistent in policy direction or core propositions.

[0178] Approaching -1: Explanation Extremely low, "Extremely high" indicates that the two conclusions are in direct conflict or completely contradictory.

[0179] Approaching 0: Explanation and "Close to each other" means that the two conclusions do not support or refute each other.

[0180] The third type of disagreement means that the two proof trees are highly consistent in their factual basis and logical reasoning, but the final conclusions are completely opposite. This usually stems from fundamental differences in the two parties' value judgment standards or risk preferences. For example, two experts may have completely identical analyses of the economic situation, but one, based on a value orientation of prioritizing growth, recommends an easing policy, while the other, based on a value orientation of risk prevention and control, recommends a prudent policy.

[0181] Type 4 of the divergence: Scope Divergence

[0182] Criterion: The set of argument nodes of a proof tree is approximately a subset of the set of arguments of another proof tree, meaning the larger tree contains all the reasoning content of the smaller tree and introduces additional reasoning branches. The expression is:

[0183] ,

[0184] in, This represents the subset coverage threshold, which can take a value of 0.8. This represents the threshold for the difference in the number of unique argument nodes, and can take the value 3.

[0185] The meaning of Disagreement Type Four: In two arguments (two proof trees), one side considers factors or dimensions not addressed by the other. For example, expert A's analysis focuses only on domestic economic data, while expert B, based on expert A's complete analysis, additionally incorporates international trade friction factors, thus arriving at different conclusions.

[0186] In this embodiment, after obtaining the shared set of argument nodes and the unique set of argument nodes in step S108, the relationship type of each pair of paired proof trees is determined by combining the following relationship types:

[0187] Relationship 1: Echoing Relationship

[0188] Judgment criterion: Shared set of argument nodes With high coverage and consistent stance among the conclusion nodes, the expression is:

[0189] ,

[0190] in, This represents the first positive stance threshold, which can be 0.5.

[0191] The meaning of relation one: The two proof trees are highly consistent in reasoning structure and final conclusion, and the viewpoint published later echoes the viewpoint published earlier.

[0192] Relationship Two: Continuing and Deepening the Relationship

[0193] Decision criteria: The smaller tree is approximately a subgraph of the larger tree, and the conclusion nodes have the same position. The expression is:

[0194] ,

[0195] Among them, subscript (After) and (Prior) determined based on the publication date of the opinion document, This represents the second positive stance threshold, which can take the value 0.3.

[0196] The meaning of relation two: The later-published viewpoint is a further development of the reasoning of the earlier-published viewpoint, adding new arguments or branches of reasoning.

[0197] Relationship 3: Supplementary Relationship

[0198] Judgment criterion: The conclusion nodes have the same stance, but share the same set of argument nodes. The coverage is moderate, and each of the two proof trees has its own unique inference branch. The expression is:

[0199] ,

[0200] in, This represents the low coverage threshold, and can take a value of 0.3.

[0201] The meaning of relation three: Two proof trees (two viewpoints) arrive at similar conclusions, but each is argued from different angles or based on different evidence, thus complementing each other.

[0202] Relationship 4: Structural Opposition Relationship with Positioning

[0203] Decision criterion: The positions of the conclusion nodes are opposite, expressed as:

[0204] ,

[0205] Combined with the set of divergence nodes And the type of disagreement at each point of disagreement (factual disagreement / reasoning disagreement / value disagreement / scope disagreement), forming a structured description of the opposing relationships.

[0206] The meaning of relation four: The two proof trees (two viewpoints) have opposing conclusions and can accurately point out the specific location (divergence node and divergence type) where the opposition occurs in the reasoning.

[0207] Relationship 5: Evolution of Viewpoints

[0208] Judgment criteria: Two viewpoints originate from the same entity (same person or institution), are published at different times, and it is proven that the tree structure has undergone significant changes. The expression is:

[0209] ,

[0210] in, Proof Tree The entity to which it belongs Proof Tree The corresponding opinion document publication time, Indicates the time difference threshold. This represents the threshold for the rate of structural change.

[0211] The meaning of relation five: The argumentative structure of the same entity on the same issue has changed at different times, reflecting the evolution of its viewpoint.

[0212] In this embodiment, in step S108, for every two opinion documents, the optimal argument node pairing scheme, the shared argument node set, the unique argument node set, the divergence node set, the divergence type, and the relationship type are combined to obtain the relationship reasoning result, which is specifically structured as follows:

[0213] ,

[0214] in, Indicates the result of relational reasoning. Indicates the relation type, This represents a list of branching types for each branching node. The overall confidence score reflects the degree of certainty in relational reasoning.

[0215] Taking into account the structural stability of the two proof trees, the rigor of cross-document alignment, and the clarity of the stance of the conclusion nodes, The calculation method is as follows:

[0216] ,

[0217] in, , and It is an adjustable weight parameter, and In this embodiment, it can be set to , , .

[0218] and Represent the proof tree respectively and The structural stability score reflects the anti-illusion ability of a single document's argumentation logic in self-consistency checks, and is calculated using the following formulas:

[0219] ,

[0220] ,

[0221] Take the smaller of the two values. This is to reflect the "barrel effect," which states that the overall credibility of the relationship between two proof trees is limited by the proof tree that generates the least stable (i.e., has a higher risk of illusion).

[0222] The average matching quality of the optimal argument node pairing scheme (maximum common reasoning subgraph) is represented by the following formula: .

[0223] If the alignment quality of two viewpoints is high and the determination of their differences is clear, the overall credibility of the relationship between the viewpoints is high. If the alignment of two viewpoints is close to 1, the overall credibility of the viewpoint relationship will be weakened if the alignment is too tight (e.g., barely exceeding the threshold), thus preventing low-quality matching from polluting the whole.

[0224] In one specific embodiment, three viewpoint documents from a think tank report in the field of macroeconomic policy are obtained as follows:

[0225] Opinion Document 1: A research report published in June 2025 by a researcher surnamed Zhang from an economic research institute, titled "Analysis of the Current Macroeconomic Situation and Recommendations for Monetary Policy," contains the following:

[0226] China's GDP grew by 6.2% year-on-year in the first quarter of 2025, indicating that the economy is operating smoothly overall.

[0227] Meanwhile, the Consumer Price Index (CPI) has remained below 1.5% for three consecutive months, reflecting insufficient momentum in the recovery of domestic demand.

[0228] Based on the above data, the main problem in the current economy lies in insufficient effective demand rather than inflationary pressure.

[0229] Therefore, it is recommended to moderately lower policy interest rates to stimulate credit expansion and promote a recovery in consumption and investment.

[0230] Opinion Document 2: A policy commentary published in July 2025 by a researcher surnamed Li from a financial research institute, titled "A Reassessment of the Current Monetary Policy Space," the specific content of which is as follows:

[0231] According to data released by the National Bureau of Statistics, GDP growth in the first quarter of 2025 was 6.2%, continuing the previous moderate recovery trend.

[0232] However, it is worth noting that during the same period, the corporate leverage ratio climbed to a historical high of 172%, and the non-performing loan ratio of commercial banks showed a marginal increase.

[0233] Against this backdrop, while further interest rate cuts may boost credit demand in the short term, they will exacerbate the accumulation of corporate debt risks in the medium to long term, which is detrimental to the stability of the financial system.

[0234] Therefore, at the current stage, policy interest rates should be kept unchanged, and structural tools should be used to provide targeted support to weak links.

[0235] Opinion Document 3: A follow-up report published in September 2025 by a researcher surnamed Zhang from the Economic Research Institute, titled "Tracking the Economic Situation in the Second Half of the Year and Suggestions for Policy Adjustment," the specific content of which is as follows:

[0236] China's GDP is projected to grow by 6.1% in the first half of 2025, a slight slowdown compared to the first quarter.

[0237] Export growth slowed significantly in the second quarter, increasing by only 2.3% year-on-year, far below market expectations.

[0238] Meanwhile, the consumer price index fell to 1.2%, indicating further weakening of demand.

[0239] It is worth noting that financing costs for some small and medium-sized enterprises have recently seen a structural increase.

[0240] Based on comprehensive assessment, the previously proposed plan for a comprehensive interest rate cut needs to be revised. A more prudent approach would be to alleviate the financing pressure on small and medium-sized enterprises through targeted reserve requirement ratio cuts rather than a comprehensive interest rate cut, thereby avoiding excessive stimulus that could trigger systemic risks.

[0241] For the three opinion documents mentioned above, perform steps S101-S102 to convert each opinion document into a proof tree. Taking opinion document one as an example, the process of constructing the proof tree is explained below:

[0242] (1) Identify the structure of document 1 and divide it into three paragraph types:

[0243] Factual data paragraphs (GDP growth rate and CPI data section);

[0244] Analytical and inferential paragraphs ("Based on the above data, the main contradiction in the current economy lies in...");

[0245] Policy recommendation paragraphs ("Therefore, it is recommended to appropriately lower the policy interest rate...").

[0246] (2) Extract and classify the argument nodes (as shown in Table 1):

[0247] Table 1. Argumentation Nodes and Types in Opinion Document 1

[0248]

[0249] (3) Determine the supporting relationships in the reasoning and construct directed reasoning edges:

[0250] Identify the following reasoning support relationships and establish directed reasoning edges:

[0251] (Analysis and judgment that the GDP growth rate of 6.2% supports the "stable economic operation")

[0252] (Analysis and judgment that low CPI supports "insufficient domestic demand")

[0253] ("Stable economy" supports the view that "the main contradiction is insufficient demand")

[0254] ("Insufficient momentum for domestic demand recovery" supports the view that "the main contradiction is insufficient demand")

[0255] (The conclusion that "the main problem is insufficient demand" supports the suggestion to "cut interest rates".)

[0256] Based on the set of argument nodes and the set of directed reasoning edges, the opinion document is ultimately transformed into a proof tree. Its structural diagram is as follows Figure 2 As shown.

[0257] (4) Self-consistency check

[0258] Perform three independent proof tree constructions on the opinion document, resulting in three candidate proof trees. The three candidate proof trees contain: argument nodes... to The consistency score is 1.0 for all three constructions (i.e., all three constructions contain these argument nodes); directed reasoning edges , , The consistency scores are all 1.0; directed inference edges and The consistency score of the directed inference edges is 0.67 (appeared in 2 constructions, and was merged in 1 construction). and direct connection All argument nodes and directed reasoning edges have consistency scores higher than the threshold of 0.6, therefore all are retained.

[0259] Using the same processing flow, proof trees are constructed for both opinion document 2 and opinion document 3. and .

[0260] Proof tree of viewpoint document 2 The argument nodes and directed reasoning edges are as follows:

[0261] Table 2. Argumentation Nodes and Types in Opinion Document Two

[0262]

[0263] Proof tree of document 3 The argument nodes and directed reasoning edges are as follows:

[0264] Table 3. Argumentation Nodes and Types in Opinion Document 3

[0265]

[0266] Next, step S103 is executed, in which issue fingerprint vectors are extracted from the three proof trees respectively, and paired according to the similarity of the issue fingerprint vectors. The specific process is as follows:

[0267] (1) Using fusion weights Issue fingerprint vector extraction

[0268] Issue fingerprint vector From the conclusion node ("Recommendation of interest rate cut") and prerequisites , The semantic embedding vectors of (GDP and CPI data) are weighted and fused together.

[0269] Issue fingerprint vector From the conclusion node ("Maintaining interest rates unchanged") and preconditions , , The semantic embedding vectors of (GDP, leverage ratio, and non-performing loan ratio data) are weighted and fused together.

[0270] Issue fingerprint vector From the conclusion node (Targeted RRR cut) and prerequisite nodes , , , The semantic embedding vectors are obtained by weighted fusion.

[0271] (2) Topic matching (pairing)

[0272] Calculate the cosine similarity of the issue fingerprint vectors of any two proof trees out of the three proof trees:

[0273] (Both discuss the issue of monetary policy interest rate adjustments.)

[0274] (Both discuss the macroeconomic situation and policy responses.)

[0275] (Both involve the selection of monetary policy tools);

[0276] All three similarities exceeded the threshold. Therefore, all of them are added to the set of paired proof trees. .

[0277] Next, steps S104-S109 are executed to align the pairing proof tree and locate the branch nodes, thereby obtaining the relation reasoning result. The specific process is as follows:

[0278] With pairwise proof trees For example:

[0279] (1) Construct the node semantic matching matrix

[0280] There are 6 points of argumentation. There are 7 argumentation nodes, constructing Semantic matching matrix of order nodes The following lists the calculation results for the key elements (only scores above the pairing threshold are shown). (argument node pair)

[0281] Table 4. Argument Node Pairing Table

[0282]

[0283] Notice: and Although both are conclusion nodes and involve the same topic (interest rate policy), their semantic similarity is only 0.31 due to their opposing positions, which is lower than the matching threshold of 0.55. Therefore, they are not included in the matching scheme. (CPI data) in There are no corresponding data nodes in it, so it is not included in the pairing scheme.

[0284] (2) Solve for the optimal pairing scheme of argument nodes (maximum common reasoning subgraph)

[0285] Citation anchor pair Forced inclusion, and based on this, considering structural preservation constraints: exist It was established in China, and exist The establishment of China, at the same time and The matching score is 0.87, which satisfies the threshold and structure preservation constraints, therefore... Included in the matching scheme.

[0286] There are no more matching pairs that satisfy all the constraints.

[0287] Optimal argument node pairing scheme: .

[0288] Therefore, the division is as follows:

[0289] Shared set of argument nodes ( side): , ( side): ;

[0290] Unique set of argument nodes ;

[0291] Unique set of argument nodes ;

[0292] Shared set of argument nodes Coverage = .

[0293] (3) Locating the points of divergence

[0294] Check whether the successor of each argument node in the shared argument node set belongs to the unique argument node set:

[0295] (correspond In ):exist Its successor node is , The divergence node condition is met, and at the same time... middle No successor node belongs to ( exist (The middle node is a leaf node with no successor), but exist The subsequent conditions in the middle have been met, therefore / It is marked as a branch node.

[0296] (correspond In ):exist Its successors are only It does not belong to the set of argument nodes. middle The successor to It also does not belong to the set of argument nodes. Therefore / It is not a point of divergence.

[0297] Set of branching nodes: That is, the shared analytical judgment of "stable / moderate economic recovery" is where the two arguments begin to diverge.

[0298] (4) Determine the type of disagreement

[0299] Divergence point / For N-type (analysis) nodes, examine their subsequent branching regions:

[0300] exist Side: from The subsequent entry ,Include (Type N: Insufficient domestic demand) (Type D: CPI data) (Type N: "The main contradiction is insufficient demand") (Type C: "Recommendation of interest rate cut") The proportion of D-type and P-type nodes in the middle is .

[0301] exist side: The unique set of argument nodes Include (Type D: Leverage ratio 172%) (Type D: Non-performing loan ratio) (Type N: "Interest rate cuts exacerbate debt risks") (Type N: "Financial stability is threatened") (Type C: "Keep interest rates unchanged") The proportion of D-type nodes is .

[0302] Since the branching node is an N-type node, and Introduced in The lack of new data points (leverage ratio and non-performing loan ratio) in the data sets indicates a discrepancy in factual data: both arguments, following the consensus of "overall economic stability," introduce different dimensions of factual data. Focus on demand-side indicators (CPI). The introduction of financial risk indicators (leverage ratio, non-performing loan ratio) has led to the opposite policy recommendations.

[0303] (5) Determine the type of relationship

[0304] Coverage of the shared argument node set S Furthermore, the conclusions of the two proof trees ("recommend cutting interest rates" and "maintain interest rates unchanged") are contradictory. It meets the criteria for determining a structural opposition relationship.

[0305] (6) Final output of relational reasoning results:

[0306] The corresponding relational reasoning result diagram is as follows: Figure 3 As shown.

[0307] This can be understood as follows: Researcher Zhang and Researcher Li hold structurally opposing views on monetary policy. Both acknowledge that the economy is generally stable, but the disagreement stems from a further analysis of the stable situation—Researcher Zhang mainly focuses on insufficient demand (CPI data), while Researcher Li additionally introduces the dimension of financial risk (leverage ratio, non-performing loan ratio data), leading to opposing policy recommendations. The type of disagreement is a factual disagreement.

[0308] Similarly, pairing proof trees The results of the relational reasoning are as follows:

[0309] and From the same entity (a researcher surnamed Zhang), with a time interval of 3 months. After alignment, the following was found: Retained The CPI-related inference branch was revised, but export data and SME financing cost data were added, and the conclusion was changed from "across-the-board interest rate cuts" to "targeted reserve requirement ratio cuts." The coverage rate of the shared argument node set is... The similarity of the positions of the conclusion nodes is (All are in a relaxed direction but with different degrees), which is determined to be a relationship of evolving viewpoints, that is, the same researcher revised the policy recommendations after obtaining new data.

[0310] Similarly, pairing proof trees The results of the relational reasoning are as follows: and Although there are no citation anchor pairs, semantic matching reveals: In (“A comprehensive interest rate cut carries the risk of excessive stimulus”) and In The semantic matching score for "(interest rate cuts exacerbate debt risks)" is: ,and The conclusion ("targeted RRR cut") and The similarity of the conclusions ("Structural Tools-Oriented Support") is: And the coverage rate of the shared argument node set is The relationship was determined to be complementary, meaning the two researchers analyzed data from different perspectives. Pay attention to leverage ratio and non-performing loans. They focused on exports and the financing costs of SMEs, and reached a consensus on policy recommendations (both opposed a comprehensive interest rate cut and supported structural tools).

[0311] Example 2

[0312] Based on the same inventive concept, embodiments of the present invention also provide a method for constructing a cross-document viewpoint relationship graph, referring to... Figure 4 As shown, the method includes the following steps:

[0313] S201: Obtain the relationship reasoning result between every two opinion documents obtained according to the method described in Embodiment 1 above.

[0314] S202: Based on the aforementioned relational reasoning results, construct a three-layer nested viewpoint relational graph, wherein:

[0315] The first layer is the entity relationship layer. Nodes are the personnel or organizational entities corresponding to the opinion document, and edges are the overall stance relationships between entities. Based on the relationship reasoning results of all involved entities, the overall stance relationship is scored. When the score of the overall stance relationship is greater than zero, it is marked as overall consistency; otherwise, it is marked as overall opposition.

[0316] The second layer is the viewpoint relationship layer, where nodes are the proof trees corresponding to each entity, and edges represent the relationship types between the proof trees.

[0317] The third layer is the argument structure layer, where nodes are the argument nodes in the proof tree, and edges include directed reasoning edges of the same proof tree and paired edges of argument nodes of different proof trees obtained based on the shared set of argument nodes.

[0318] In this embodiment, the relational reasoning results in step S201 include the optimal argument node pairing scheme (maximum common reasoning subgraph), the set of shared argument nodes, the set of unique argument nodes, the set of divergent nodes, the divergence type, the relation type, and the comprehensive confidence score.

[0319] In this embodiment, in step S202, for the first layer (entity relationship layer), the overall position relationship is scored using the following formula:

[0320] ,

[0321] In the formula, , This represents two different entities (e.g., a researcher surnamed Zhang and a researcher surnamed Li). This indicates the overall positional relationship between two different entities. Indicates all entities involved and A collection of relation types Representing relation type The positive and negative weighting function assigns positive values ​​(+1) to relationships that echo, continue, deepen, or supplement each other, and negative values ​​(-1) to structural opposition relationships. Representing relation type The corresponding overall confidence score.

[0322] when At that time, entity and The space is marked as "overall consistent"; when At that time, it was marked as "overall opposition".

[0323] For the second layer, the proof tree corresponding to each entity is converted into a proof tree according to the method described in Example 1. The relationship type between the proof trees is the echo, continuation and deepening, supplement, structural opposition or view evolution relationship between any two proof trees. The attributes of the edges can also include the set of divergent nodes and the divergent type.

[0324] For the third layer, various argument nodes include premise, data citation, authoritative citation, analysis, and conclusion nodes. The edges include two types: one is directed reasoning edges within the same proof tree, and the other is pairing edges between argument nodes across documents (two different proof trees) (i.e., paired nodes in the best argument node pairing scheme).

[0325] The three layers are connected by affiliation relationships: the entity relationship layer nodes are connected to the opinion relationship layer nodes through the "publish" relationship, and the opinion relationship layer nodes are connected to the argument structure layer nodes through the "include" relationship.

[0326] In one specific embodiment, for the three viewpoint documents involving the aforementioned researcher surnamed Zhang and researcher surnamed Li (two entities), steps S201-S202 are executed to obtain a viewpoint relationship graph with a three-level nested structure (e.g., Figure 5 (as shown)

[0327] The first layer (entity relationship layer) contains two entity nodes: Researcher Zhang and Researcher Li. The overall position relationship between Researcher Zhang and Researcher Li is rated as follows:

[0328] ,

[0329] A negative score, marked as "overall opposition," indicates that the Zhang researcher's position gradually shifted from opposition to agreement with the Li researcher's position (weak opposition) during the evolution process.

[0330] The second layer (viewpoint relationship layer) contains three viewpoint nodes. , , And three relation edges: , , .

[0331] The third layer (argument structure layer) contains all nodes of the three proof trees, directed reasoning edges of the same proof tree, and paired edges of argument nodes of different proof trees (across documents).

[0332] Furthermore, an evolutionary diagram of opinion camps and tracing chains can be constructed through the following process:

[0333] (1) Divide into opinion camps

[0334] Based on the entity relationships at the first level, opinion camps are divided. The relationships between entities are scored. As edge weights (which can be positive or negative) of the signature graph, entities are divided into several camps using a spectral clustering algorithm (an unsupervised clustering algorithm that uses the Laplacian matrix and eigenvalue decomposition of a graph to solve the graph partitioning problem). Entities within the same camp hold consistent positions on most issues, while entities in different camps have significantly opposing viewpoints. For each camp, a camp profile is generated, including representative conclusion nodes within the camp, shared core premises and data citations, and core entities within the camp (sorted by influence).

[0335] (2) Constructing a chain of source tracing of viewpoints

[0336] The purpose of the opinion tracing chain is to track the dissemination path of a specific argument or point of view across different entities and points in time, thereby achieving version control and management at the argument level.

[0337] The construction method is as follows: For all proof trees under the same topic, sort them according to the publication time of their corresponding documents. ( For each pair of temporally adjacent proof trees Using the relational reasoning results obtained according to the method described in Example 1, the following three types of tracing events are identified:

[0338] (a) The event of succession of arguments: The shared set of argument nodes contains A certain premise or data citation node indicates that the subsequent argument adopts a certain factual basis of the previous argument.

[0339] (b) New events added to the argument: The unique set of argument nodes It appeared in The absence of a premise or data citation node in the argument indicates that the subsequent argument introduces entirely new factual evidence.

[0340] (c) Counter-argument: There is an analysis node whose content is related to... The fact that a certain shared data citation node in the middle forms a negation of the implication relationship (determined as a "contradictory" relationship by the natural language inference model) indicates that the subsequent argument questions a certain factual basis of the previous argument.

[0341] By linking the aforementioned source events in chronological order, a source chain of viewpoints can be formed. This source chain can fully present who first introduced a certain argument and when, which entities subsequently adopted or cited it, when it was questioned or refuted by a certain entity, and the subsequent chain reaction triggered by the refutation.

[0342] In one specific embodiment, regarding the aforementioned researcher surnamed Zhang and researcher surnamed Li, there are only two entities, and the overall positional relationship score is close to zero. The camp division result is: initially, there are two opposing camps (the "lenient faction" of researcher Zhang versus the "prudent faction" of researcher Li), but... After publication, the stance of the researcher surnamed Zhang shifted towards that of the researcher surnamed Li, creating a convergence. For example... Figure 6 As shown, sorted by time (June) → (July) → (September) Identify the following source events:

[0343] Added event to the argument — Introduced Data on leverage ratios and non-performing loan ratios that are not available in China.

[0344] Argument inheritance incident— Inherited CPI data cited; new events added to the argument— Export data and financing cost data for SMEs have been added.

[0345] Argument inheritance incident— Analysis of the risks of a comprehensive interest rate cut The semantic relationship between "interest rate cuts exacerbating debt risks" and "the risk warnings from researchers surnamed Li" indicates that the researcher surnamed Zhang incorporated the risk warnings from researcher surnamed Li into his subsequent report.

[0346] The aforementioned chain of origin reveals a key trajectory of viewpoint evolution: Researcher surnamed Li... The argument for "debt risk" put forward in the paper was conveyed to the subsequent research of the researcher surnamed Zhang through academic discussions. This prompted the latter to revise its original recommendation for a comprehensive interest rate cut.

[0347] Furthermore, when acquiring newly published opinion documents from think tank reports or authoritative news sources, incremental updates (regarding the relationships between viewpoints) are performed:

[0348] Perform steps S101-S102 above on the new viewpoint document to construct a new proof tree. ;

[0349] For the new proof tree Perform step S103 above to extract The issue fingerprint vector is obtained and paired with an existing proof tree of the same issue to obtain a new paired proof tree;

[0350] Perform the above steps S104-S109 on the new pairing proof tree to obtain the new relational reasoning result;

[0351] The new relational reasoning results are added to the existing viewpoint relational graph, and the camp division and viewpoint tracing chain are updated locally.

[0352] The incremental update mechanism described above aims to maintain the timeliness of the graph while avoiding the high computational overhead of full recalculation. This adapts to scenarios with continuous document stream input and addresses the problem of excessive computational cost in full graph construction when facing continuous massive document input. By designing an incremental update mechanism, new viewpoint documents only need to undergo local alignment calculations with subsets of the same topic and the incremental viewpoint relationships are added to the existing graph. This significantly reduces the computational cost of continuous system operation and improves project deployability and timeliness.

[0353] Example 3

[0354] Based on the same inventive concept, embodiments of the present invention also provide a cross-document viewpoint relationship reasoning device, referring to... Figure 7 As shown, it includes:

[0355] The paragraph type annotation module 101 is used to acquire multiple opinion documents and annotate each paragraph in each opinion document according to the preset paragraph type to obtain the opinion document after annotating the paragraph type.

[0356] The proof tree conversion module 102 is used to extract sentence-level argument nodes for each paragraph of each opinion document after each paragraph is labeled with a paragraph type, using a pre-trained language model according to a preset argument node type, to obtain an argument node set for each opinion document, and to determine the reasoning support relationship between any two argument nodes in the argument node set through a natural language inference model, to obtain a set of directed reasoning edges for each opinion document, and to convert each opinion document into a proof tree based on the argument node set and the set of directed reasoning edges.

[0357] The proof tree pairing module 103 is used to extract the issue fingerprint vector of each proof tree based on the set of argument nodes in each proof tree, calculate the similarity of the issue fingerprint vectors of any two proof trees, and determine each pair of proof trees whose issue fingerprint vector similarity is greater than a set issue matching threshold as a paired proof tree.

[0358] The node semantic matching matrix construction module 104 is used to calculate the matching score between any argument node of one proof tree and any argument node of the other proof tree in each pair of paired proof trees based on preset rules, and to construct the node semantic matching matrix of each pair of paired proof trees.

[0359] The optimal argument node pairing solution module 105 is used to solve the optimal argument node pairing solution for each pair of paired proof trees based on the node semantic matching matrix and the preset node constraints, so as to obtain the shared argument node set of each pair of paired proof trees and the unique argument node set of each proof tree in each pair of paired proof trees.

[0360] The divergence node location module 106 is used to locate divergence nodes and obtain a divergence node set based on the shared argument node set and the unique argument node set.

[0361] The branch type determination module 107 is used to determine the branch type of all branch nodes based on the set of branch nodes and the preset branch types.

[0362] The relationship type determination module 108 is used to determine the relationship type of each pair of paired proof trees based on the shared set of proof nodes, the unique set of proof nodes, and the preset relationship type.

[0363] The relation reasoning result module 109 is used to obtain the relation reasoning result for each pair of opinion documents based on the optimal argument node pairing scheme, the shared argument node set, the unique argument node set, the divergence node set, the divergence type, and the relation type.

[0364] Example 4

[0365] Based on the same inventive concept, embodiments of the present invention also provide a cross-document viewpoint relationship graph construction apparatus, referring to... Figure 8 As shown, it includes:

[0366] The paragraph type annotation module 101 is used to acquire multiple opinion documents and annotate each paragraph in each opinion document according to the preset paragraph type to obtain the opinion document after annotating the paragraph type.

[0367] The proof tree conversion module 102 is used to extract sentence-level argument nodes for each paragraph of each opinion document after each paragraph is labeled with a paragraph type, using a pre-trained language model according to a preset argument node type, to obtain an argument node set for each opinion document, and to determine the reasoning support relationship between any two argument nodes in the argument node set through a natural language inference model, to obtain a set of directed reasoning edges for each opinion document, and to convert each opinion document into a proof tree based on the argument node set and the set of directed reasoning edges.

[0368] The proof tree pairing module 103 is used to extract the issue fingerprint vector of each proof tree based on the set of argument nodes in each proof tree, calculate the similarity of the issue fingerprint vectors of any two proof trees, and determine each pair of proof trees whose issue fingerprint vector similarity is greater than a set issue matching threshold as a paired proof tree.

[0369] The node semantic matching matrix construction module 104 is used to calculate the matching score between any argument node of one proof tree and any argument node of the other proof tree in each pair of paired proof trees based on preset rules, and to construct the node semantic matching matrix of each pair of paired proof trees.

[0370] The optimal argument node pairing solution module 105 is used to solve the optimal argument node pairing solution for each pair of paired proof trees based on the node semantic matching matrix and the preset node constraints, so as to obtain the shared argument node set of each pair of paired proof trees and the unique argument node set of each proof tree in each pair of paired proof trees.

[0371] The divergence node location module 106 is used to locate divergence nodes and obtain a divergence node set based on the shared argument node set and the unique argument node set.

[0372] The branch type determination module 107 is used to determine the branch type of all branch nodes based on the set of branch nodes and the preset branch types.

[0373] The relationship type determination module 108 is used to determine the relationship type of each pair of paired proof trees based on the shared set of proof nodes, the unique set of proof nodes, and the preset relationship type.

[0374] The relation reasoning result module 109 is used to obtain the relation reasoning result for each pair of opinion documents based on the optimal argument node pairing scheme, the shared argument node set, the unique argument node set, the divergence node set, the divergence type, and the relation type.

[0375] The relationship graph construction module 110 is used to construct a three-layer nested viewpoint relationship graph based on the relationship reasoning results, wherein:

[0376] The first layer is the entity relationship layer. Nodes are the personnel or organizational entities corresponding to the opinion document, and edges are the overall stance relationships between entities. Based on the relationship reasoning results of all involved entities, the overall stance relationship is scored. When the score of the overall stance relationship is greater than zero, it is marked as overall consistency; otherwise, it is marked as overall opposition.

[0377] The second layer is the viewpoint relationship layer, where nodes are the proof trees corresponding to each entity, and edges represent the relationship types between the proof trees.

[0378] The third layer is the argument structure layer. The nodes are the argument nodes in the proof tree, and the edges include directed reasoning edges of the same proof tree and paired edges of argument nodes of different proof trees obtained based on the shared set of argument nodes.

[0379] Example 5

[0380] Based on the same inventive concept, embodiments of the present invention provide a computer-readable storage medium storing a computer program / instruction thereon, which, when executed by a processor, implements the cross-document viewpoint relationship reasoning method as described in Embodiment 1 above, and / or the cross-document viewpoint relationship graph construction method as described in Embodiment 2 above.

[0381] Example 6

[0382] Based on the same inventive concept, embodiments of the present invention provide a computer device, including a memory, a processor, and a computer program stored in the memory. The processor executes the computer program to implement the cross-document viewpoint relationship reasoning method described in Embodiment 1 above, and / or the cross-document viewpoint relationship graph construction method described in Embodiment 2 above.

[0383] In this embodiment, when describing representative embodiments, the specification may have presented the method and / or process as a specific sequence of steps. However, the method or process should not be limited to the specific order of steps described herein, to the extent that the method or process does not depend on the specific order of steps described herein. Other sequences of steps are also possible, as understood by those skilled in the art. Therefore, the specific order of steps set forth in the specification should not be construed as a limitation of the claims. Furthermore, the claims for the method and / or process should not be limited to the steps performed in the written order, and those skilled in the art will readily understand that these orders can be varied and still remain within the spirit and scope of the embodiments of the invention.

[0384] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0385] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus, and computer program products according to embodiments of the invention. Those skilled in the art will understand that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0386] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0387] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

Claims

1. A method for cross-document viewpoint relationship reasoning, characterized in that, The method includes: Obtain multiple opinion documents and annotate each paragraph in each opinion document according to preset paragraph types to obtain opinion documents with annotated paragraph types; For each paragraph of an opinion document after labeling the paragraph type, a pre-trained language model is used to extract sentence-level argument nodes according to the preset argument node type to obtain an argument node set for each opinion document. Then, a natural language inference model is used to determine the reasoning support relationship between any two argument nodes in the argument node set to obtain a set of directed reasoning edges for each opinion document. Based on the argument node set and the set of directed reasoning edges, each opinion document is converted into a proof tree. Based on the set of argument nodes in each proof tree, an issue fingerprint vector is extracted from each proof tree, and the similarity between the issue fingerprint vectors of any two proof trees is calculated. Each pair of proof trees whose issue fingerprint vector similarity is greater than the set issue matching threshold is determined as a paired proof tree. Based on preset rules, the matching score between any argument node of one proof tree and any argument node of the other proof tree in each pair of paired proof trees is calculated, and the node semantic matching matrix of each pair of paired proof trees is constructed. Based on the node semantic matching matrix and the preset node constraints, the optimal argument node pairing scheme for each pair of paired proof trees is solved, resulting in the shared argument node set for each pair of paired proof trees and the unique argument node set for each proof tree in each pair of paired proof trees. Based on the shared set of argument nodes and the unique set of argument nodes, the divergence nodes are located, and the set of divergence nodes is obtained. Based on the set of branch nodes and the preset branch types, determine the branch type of all branch nodes; Based on the shared set of argument nodes, the unique set of argument nodes, and the preset relationship type, determine the relationship type of each pair of paired proof trees; Based on the optimal argument node pairing scheme, the shared argument node set, the unique argument node set, the divergence node set, the divergence type, and the relationship type, the relationship reasoning result for each pair of opinion documents is obtained.

2. The cross-document viewpoint relationship reasoning method as described in claim 1, characterized in that, Also includes: For the pre-trained language model, argument nodes are extracted independently multiple times from the same opinion document by using different random seeds or prompt templates, resulting in multiple argument node sets. The reasoning support relationship between any two argument nodes in each argument node set is determined by a natural language inference model, resulting in multiple directed reasoning edge sets. Based on the multiple argument node sets and the corresponding multiple directed reasoning edge sets, each opinion document is converted into multiple candidate proof trees. For each argument node in each candidate proof tree, the consistency score of the argument node is calculated using the following formula: ， In the formula, This indicates the consistency score of the argument nodes. , These represent the argument nodes in two different candidate proof trees. , Representing the argument nodes respectively , semantic embedding vector, Represents the cosine similarity function. This represents the total number of candidate proof trees. This represents the consensus threshold for the argument nodes; For each directed inference edge in each candidate proof tree, the consistency score of the directed inference edge is calculated using the following formula: ， In the formula, This represents the consistency score of directed inference edges. , Let represent the directed inference edges in two different candidate proof trees. , , , Representing the argument nodes respectively , , , semantic embedding vector, This represents the consistency threshold for directed inference edges; Based on the set of argument nodes whose consistency scores reach the first threshold and the set of directed inference edges whose consistency scores reach the second threshold, each opinion document is transformed into a proof tree.

3. The cross-document viewpoint relationship reasoning method as described in claim 1 or 2, characterized in that, The types of argumentation nodes include premise nodes, data citation nodes, authoritative citation nodes, analysis nodes, and conclusion nodes; Based on the set of argument nodes in each proof tree, an issue fingerprint vector is extracted from each proof tree, and the similarity between the issue fingerprint vectors of any two proof trees is calculated. Each pair of proof trees whose issue fingerprint vector similarity is greater than a set issue matching threshold is identified as a paired proof tree, including: Based on the set of argument nodes in each proof tree, the issue fingerprint vector is extracted for each proof tree using the following formula: ， In the formula, This represents the issue fingerprint vector of any proof tree. This represents the set of conclusion nodes within the set of argument nodes. This represents the set of argument nodes in the argument node, which includes premise nodes, data citation nodes, and authoritative citation nodes. Indicates the argument node semantic embedding vector, Indicates the fusion weights; The similarity of the issue fingerprint vectors of any two proof trees is calculated using the following formula. Each pair of proof trees whose issue fingerprint vector similarity is greater than a set issue matching threshold is identified as a paired proof tree: ， In the formula, This represents the similarity of the issue fingerprint vectors of any two proof trees. , This represents the issue fingerprint vector of any two proof trees.

4. The cross-document viewpoint relationship reasoning method as described in claim 3, characterized in that, Based on preset rules, the matching score between any argument node of one proof tree and any argument node of the other proof tree in each pair of paired proof trees is calculated, and a node semantic matching matrix for each pair of paired proof trees is constructed, including: For any pair of proof trees ,in , , , , , Represent the proof tree respectively , The set of argument nodes, , Represent the proof tree respectively , Given a set of directed inference edges, calculate based on the following preset rules. Any argument node and Construct a pairwise proof tree based on the matching scores between any argument nodes. of Semantic matching matrix of order nodes : Rule 1: If The i-th argument node and The Middle One argument node If all of them are data citation nodes or authoritative citation nodes, and cite the same data source or authoritative source, then... , express The i-th argument node and The Middle One argument node Match score between them; Rule 2: If The i-th argument node and The Middle One argument node If the argument nodes are of the same type but do not satisfy the conditions of Rule 1, then... ,in, , Representing the argument nodes respectively , semantic embedding vector; Rule 3: If The i-th argument node and The Middle One argument node The types of argument nodes are different, then ,in, Indicates the type attenuation factor.

5. The cross-document viewpoint relationship reasoning method as described in claim 4, characterized in that, Based on the node semantic matching matrix and preset node constraints, the optimal proof node pairing scheme for each pair of proof trees is solved, including: Define any pair of paired proof trees The set of argument node pairings is: ,in and It can appear in at most one argument node pairing; Solve any pair of proof trees based on the following formula and node constraints. The optimal pairing scheme for argument nodes is obtained for each pair of paired proof trees: ， Constraint 1: When At that time, the argument node pairing Included in the matching scheme, among which The pairing threshold; Constraint 2: If there are two argument nodes paired in the pairing scheme. and And in There exists a directed reasoning edge. Then in There is also a source from arrive Directed reasoning edges; Constraint 3: Argument node pairings that conform to Rule 1 are included in the pairing scheme.

6. A method for constructing a cross-document viewpoint relationship graph, characterized in that, include: Obtain the relational reasoning result between every two opinion documents obtained by the method according to any one of claims 1-5; Based on the aforementioned relational reasoning results, a three-layer nested viewpoint relational graph is constructed, wherein: The first layer is the entity relationship layer. Nodes are the personnel or organizational entities corresponding to the opinion document, and edges are the overall stance relationships between entities. Based on the relationship reasoning results of all involved entities, the overall stance relationship is scored. When the score of the overall stance relationship is greater than zero, it is marked as overall consistency; otherwise, it is marked as overall opposition. The second layer is the viewpoint relationship layer, where nodes are the proof trees corresponding to each entity, and edges represent the relationship types between the proof trees. The third layer is the argument structure layer. The nodes are the argument nodes in the proof tree, and the edges include directed reasoning edges of the same proof tree and paired edges of argument nodes of different proof trees obtained based on the shared set of argument nodes.

7. A cross-document viewpoint relationship reasoning device, characterized in that, include: The paragraph type annotation module is used to acquire multiple opinion documents and annotate each paragraph in each opinion document according to preset paragraph types, so as to obtain opinion documents with annotated paragraph types. The proof tree conversion module is used to extract sentence-level argument nodes for each paragraph of each opinion document after each paragraph is labeled with a paragraph type, using a pre-trained language model according to the preset argument node type, to obtain an argument node set for each opinion document. Then, it uses a natural language inference model to determine the reasoning support relationship between any two argument nodes in the argument node set, to obtain a set of directed reasoning edges for each opinion document. Based on the argument node set and the set of directed reasoning edges, each opinion document is converted into a proof tree. The proof tree pairing module is used to extract the issue fingerprint vector of each proof tree based on the set of argument nodes in each proof tree, calculate the similarity of the issue fingerprint vectors of any two proof trees, and determine each pair of proof trees whose issue fingerprint vector similarity is greater than a set issue matching threshold as a paired proof tree. The node semantic matching matrix construction module is used to calculate the matching score between any argument node of one proof tree and any argument node of the other proof tree in each pair of paired proof trees based on preset rules, and to construct the node semantic matching matrix for each pair of paired proof trees. The optimal argument node pairing solution module is used to solve the optimal argument node pairing solution for each pair of paired proof trees based on the node semantic matching matrix and preset node constraints, so as to obtain the shared argument node set of each pair of paired proof trees and the unique argument node set of each proof tree in each pair of paired proof trees. The divergence node location module is used to locate divergence nodes and obtain a set of divergence nodes based on the shared set of argument nodes and the unique set of argument nodes. The branching type determination module is used to determine the branching type of all branching nodes based on the set of branching nodes and a preset branching type. The relationship type determination module is used to determine the relationship type of each pair of paired proof trees based on the shared set of proof nodes, the unique set of proof nodes, and the preset relationship type. The relational reasoning result module is used to obtain the relational reasoning result for each pair of opinion documents based on the optimal argument node pairing scheme, the shared argument node set, the unique argument node set, the divergence node set, the divergence type, and the relation type.

8. A cross-document viewpoint relationship graph construction device, characterized in that, include: A relational reasoning result acquisition module is used to acquire the relational reasoning result between every two opinion documents obtained by the method according to any one of claims 1-5; The relationship graph construction module is used to construct a three-level nested viewpoint relationship graph based on the relationship reasoning results, wherein: The first layer is the entity relationship layer. Nodes are the personnel or organizational entities corresponding to the opinion document, and edges are the overall stance relationships between entities. Based on the relationship reasoning results of all involved entities, the overall stance relationship is scored. When the score of the overall stance relationship is greater than zero, it is marked as overall consistency; otherwise, it is marked as overall opposition. The second layer is the viewpoint relationship layer, where nodes are the proof trees corresponding to each entity, and edges represent the relationship types between the proof trees. The third layer is the argument structure layer. The nodes are the argument nodes in the proof tree, and the edges include directed reasoning edges of the same proof tree and paired edges of argument nodes of different proof trees obtained based on the shared set of argument nodes.

9. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instruction is executed by the processor, it implements the cross-document viewpoint relationship reasoning method according to any one of claims 1-5, and / or the cross-document viewpoint relationship graph construction method according to claim 6.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the cross-document viewpoint relationship reasoning method according to any one of claims 1-5, and / or the cross-document viewpoint relationship graph construction method according to claim 6.

Images