Process result deduction method and system based on large model and causal analysis

Abstract

The application relates to the technical field of large models, in particular to a process result deduction method and system based on a large model and causal analysis, which first segments a to-be-measured file, constructs a causal diagram according to a segmentation result, optimizes the causal diagram through a graph optimization rule, obtains a structured causal diagram, then compressively encodes the structured causal diagram, stores a structured text file, finally inputs the structured text file into a preset review report generation model, and generates a final project review report; the application provides a process result deduction system based on a large model and causal analysis, which comprises a causal diagram construction module, a project summary generation module, a score calculation and evaluation generation module and a report output module, realizes intelligent deduction of a whole process of a project, and improves the accuracy, objectivity and intelligent level of project review.

Description

Technical Field

[0001] This invention relates to the field of large model technology, and in particular to a method and system for extrapolating process results based on large models and causal analysis. Background Technology

[0002] Project review is a core decision-making process that ensures project investment security, controls implementation quality, and mitigates potential risks. It directly impacts resource allocation efficiency and the project's lifecycle value. This is especially true in sectors like power and infrastructure, where large-scale investments and high technological complexity necessitate strong review conclusions that determine project approval and optimal solution design. However, traditional review methods generally rely on manual analysis, evaluation, and result extrapolation of project materials. Reviewers must read extensive project documents and rely on experience to assess feasibility, risks, and potential outcomes. The review process is highly dependent on the reviewers' personal experience and cognitive abilities, lacking standardized objective criteria. Furthermore, subjective human perspectives can easily overlook hidden influencing factors, further reducing the reliability of review conclusions. To address the subjectivity limitations of human review, large-scale model technology has been introduced into the field of intelligent project review. Leveraging its powerful semantic understanding and rule learning capabilities, it avoids subjective differences caused by human experience. However, large-scale models are limited by the fixed size of the context window and cannot directly handle long text materials such as reports and cost lists that are often hundreds of pages long in projects. Information is often forcibly truncated due to the text length exceeding the capacity limit. This lack of long text processing capabilities means that the cross-chapter related data and complete technical logic chains required for review cannot be perceived by the model as a whole, directly affecting the completeness of the review analysis. Summary of the Invention

[0003] The technical problem this invention aims to solve is to address the shortcomings of existing technologies by providing a method for inferring process results based on a large model and causal analysis. First, the test file is segmented, and a causal graph is constructed based on the segmentation results. Second, the causal graph is optimized using graph optimization rules to obtain a structured causal graph. Then, causal chains are extracted from the structured causal graph and represented as alternating sequences of node and relation symbol units. Quaternary combination encoding is performed on the node symbol units, and symbol pair encoding is performed on the relation symbol units to obtain a set of causal strings. Next, the causal strings are split into fixed-length segment units, clustered according to semantic similarity and frequency of occurrence to form a pattern dictionary. Dictionary symbols are used to replace repeated segments in the causal strings. The frequency and average weight of the pattern dictionary in the causal string set are calculated to construct a causal spectrum matrix. Finally, the causal spectrum matrix is ​​converted into a token sequence adapted to the input of a large model, completing the structured causal graph compression encoding and storing it as a structured text file. Finally, the structured text file is input into a preset review report generation model to generate the final project review report. This invention provides a process result extrapolation system based on large models and causal analysis, including a causal graph construction module, a project overview generation module, a score calculation and evaluation generation module, and a report output module, which realizes intelligent extrapolation of the entire project process and improves the accuracy, objectivity, and intelligence level of project review.

[0004] To achieve the above objectives, the present invention adopts the following technical solution:

[0005] The method for extrapolating process results based on large models and causal analysis includes:

[0006] Obtain the file to be tested;

[0007] The file to be tested is segmented, and a causal graph is constructed based on the segmentation results. The nodes of the causal graph are the text content corresponding to the segmentation and standardization of the file to be tested, and the edges of the causal graph are the causal logic between the segments.

[0008] The causal graph is optimized using graph optimization rules to obtain a structured causal graph;

[0009] The structured cause-effect graph is compressed and encoded, and then stored to obtain a structured text file;

[0010] The structured text file is input into the preset review report generation model, which then outputs a project overview, a project evaluation report, and a project overall score to obtain the final review report.

[0011] Construct a cause-effect graph based on the segmentation results, including:

[0012] Initialize the cause-effect graph by inserting the overall project score as a graph node into the cause-effect graph.

[0013] After the file to be tested is segmented, the corresponding text content is cleaned and the segments are standardized to obtain standardized segments.

[0014] The standardized segments are input into the semantic analysis model. By calculating and ranking the semantic weights of the words in the standardized segments, the core topic words of the standardized segments are selected.

[0015] Based on the core keywords, determine whether there is a causal relationship between the standardized segment and existing nodes in the causal graph. If there is, insert the text content of the standardized segment as a new node into the causal graph and establish a connection with the existing nodes in the causal graph that have a causal relationship with the standardized segment. If there is no causal relationship, insert the text content of the standardized segment as a new node into the causal graph and do not establish a connection with the existing nodes in the causal graph.

[0016] Iterate through all segments to form a cause-effect graph.

[0017] The causal graph is optimized using graph optimization rules to obtain a structured causal graph, including:

[0018] Traverse all nodes in order and record the core keywords of each node;

[0019] Calculate the semantic similarity of core keywords between nodes, filter out nodes whose semantic similarity exceeds a preset similarity threshold, and form a set of nodes to be optimized;

[0020] The text content of the node to be optimized is split into the smallest semantic units. The semantic similarity between the smallest semantic units is calculated. The smallest semantic units with semantic similarity exceeding the preset similarity threshold are merged and deleted. The smallest semantic units are integrated to obtain new nodes.

[0021] Connect all edges that connect the corresponding node to be optimized to the new node;

[0022] The set of nodes to be optimized is iteratively optimized to obtain a structured cause-effect graph.

[0023] The structured causal graph is compressed and encoded, including:

[0024] Causal chains are extracted based on the causal type, weight range, time delay parameter, and node level of the edges in the structured causal graph, where each causal chain is represented as a symbol sequence of alternating node symbol units and relation symbol units;

[0025] The node symbol unit is encoded using quaternary combination encoding, and the relation symbol unit is encoded using symbol pairs according to the causal relationship category to obtain a set of causal strings, wherein the causal relationship category includes strong cause, weak cause, parallel causality and reverse causality;

[0026] Each causal string is decomposed into fixed-length segment units. The segment units are clustered according to their semantic similarity and frequency of occurrence to obtain a pattern dictionary. Duplicate segments in the causal string set are replaced according to the dictionary symbols of the pattern dictionary.

[0027] Using the pattern dictionary as the row index and the edge weight intensity distribution, node hierarchy distribution, and document location distribution as the column index, the occurrence frequency and average weight of the pattern dictionary in the causal string set are calculated to obtain the causal spectrum matrix, wherein the edge weight intensity distribution, node hierarchy distribution, and document location distribution are calculated for the causal string according to a preset bucketing rule;

[0028] The causal spectrum matrix is ​​transformed into a token sequence that conforms to the input format of a large model.

[0029] The node symbol units are quaternary combined encodings, and the relation symbol units are symbol-pair encodings based on causal relationship categories to obtain a set of causal strings, including:

[0030] The objective cost function is determined based on the statistical distribution of node attributes. The minimum cost corresponding to each node symbol unit is calculated, and a corresponding quaternary prefix code dictionary is assigned based on the minimum cost. The node attributes include node category, saliency level, and time attribute. The objective cost function is... Where 'a' represents a combination of node attributes. This represents the probability distribution of node attribute combinations in a causal chain. This represents the encoded string assigned to the combination of node attributes. This indicates the encoding length of the node attribute combination. This represents the expected average code length of the encoding scheme;

[0031] Based on the relation symbol unit, the corresponding causal relationship category is quantized into discrete buckets and mapped into binary relation symbol pairs, wherein the binary relation symbol pairs include causal relationship categories and time delay parameters corresponding to the causal relationship categories;

[0032] Traversing along the causal chain, the symbol sequence is alternately output according to the order of the quaternary prefix code dictionary and the binary relation symbol pairs through a finite state transformer to obtain a set of causal strings.

[0033] Clustering the fragment units based on their semantic similarity and frequency of occurrence yields a pattern dictionary. Repeated fragments in the causal string set are then replaced using dictionary symbols from the pattern dictionary, including:

[0034] Based on the semantic similarity and frequency of occurrence, the fragment homology vector is calculated by combining the node symbol sequence, relation symbol sequence, and hierarchical position symbol of the causal string;

[0035] The fragment units are aggregated according to the fragment homology vector to obtain at least one homology fragment cluster, wherein the fragment units in each homology fragment cluster are causally compatible.

[0036] Sequence alignment is performed on the homologous fragment clusters to generate a consensus skeleton, and the differing sites are marked in the consensus skeleton;

[0037] Encoding is performed on the differential sites to obtain dictionary symbols and corresponding variant parameter vectors, wherein the encoding includes mutation codes and cross codes. The mutation codes are used for substitution, insertion, and deletion of differential sites, and the cross codes are used for combining commutative subpaths of differential sites.

[0038] Replace fragment units in the same fragment cluster according to the dictionary symbols and the variant parameter vectors corresponding to the dictionary symbols.

[0039] The structured text file includes a standardized segmented content text file and a causal relationship information text file. The standardized segmented content text file is used to store the standardized segments obtained after data cleaning and text standardization of the segments. The causal relationship information text file is used to store the compressed encoding result obtained by compressing and encoding the structured causal graph.

[0040] The structured text file is input into a preset review report generation model, which then outputs a project overview, a project evaluation report, and an overall project score to obtain the final review report, including:

[0041] Obtain standardized segmented text files, extract the core content of all segments through a semantic analysis model, and logically integrate the core content through a large text generation model to obtain an overall project overview.

[0042] Extract the influencing factors that affect the overall project score from the causal relationship information text file, set the score and weight of the influencing factors, compare the influencing factors with the preset project evaluation criteria, calculate the overall project score and generate a project evaluation report;

[0043] The final review report is obtained by integrating the overall project overview, project evaluation report, and overall project score.

[0044] The influencing factors are compared with the preset project evaluation criteria, the overall project score is calculated, and a project evaluation report is generated, including:

[0045] Compare the influencing factors with the preset project evaluation criteria to determine the advantageous and disadvantageous influencing factors;

[0046] Based on the scores and weights of the advantageous factors, the bonus points for the project are calculated, and based on the scores and weights of the disadvantageous factors, the deduction points for the project are calculated, and finally the overall score of the project is obtained.

[0047] Based on the factors influencing strengths and weaknesses, a large model is generated from text to output a project evaluation report.

[0048] A process outcome extrapolation system based on large-scale models and causal analysis, the system comprising a causal graph construction module, a project overview generation module, a score calculation and evaluation generation module, and a report output module:

[0049] The causal graph construction module is used to identify causal relationships between segments, construct an initial causal graph, and optimize the initial causal graph to obtain a structured causal graph.

[0050] The project overview generation module is used to extract the core content of each standardized segment and logically integrate it to obtain the project overview.

[0051] The score calculation and evaluation generation module is used to quantitatively analyze the key factors affecting the project, calculate the overall project score based on the evaluation criteria, and generate a project evaluation report.

[0052] The report output module is used to integrate the overall project overview, overall project score, and project evaluation report to form the final review report.

[0053] Compared with existing technologies, this application fully preserves implicit long causal chains through a causal graph storage structure and further introduces a causal graph compression encoding mechanism, which not only improves the information carrying capacity of a single token and avoids context window explosion, but also provides a traceable and interpretable causal logic sequence when inputting large models. This achieves an upgrade from "human experience review" to "data-driven, logically traceable, and precise decision-making", enabling intelligent deduction of the entire project process and improving the accuracy, objectivity, and intelligence level of project review. Attached Figure Description

[0054] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0055] Figure 1 This is a flowchart illustrating the process result deduction method based on large model and causal analysis in Embodiment 1 of the present invention;

[0056] Figure 2 This is a schematic diagram illustrating the principle of merging similar nodes in a cause-effect graph according to Embodiment 1 of the present invention.

[0057] Figure 3 This is a schematic diagram of the principle of quaternary combination encoding of node symbol units in Embodiment 1 of the present invention;

[0058] Figure 4 This is a flowchart illustrating the conversion of a causal chain into a cluster of homologous fragments in Embodiment 1 of the present invention.

[0059] Figure 5 This is a schematic diagram illustrating the consensus skeleton generation principle of Embodiment 1 of the present invention;

[0060] Figure 6 This is a system module diagram of the process result inference system based on large model and causal analysis in Embodiment 2 of the present invention. Detailed Implementation

[0061] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0062] Example 1

[0063] Please see Figure 1 The present invention provides an embodiment of a process result deduction method based on a large model and causal analysis, the specific steps of which are as follows:

[0064] S1: Obtain the file to be tested;

[0065] In this embodiment, the project business plan is used as the test document. Its content is built around the key dimensions of project review to construct a complete information system, including executive summary, project background and market analysis, product / service introduction, business model introduction, operation and team plan, financial planning and risk assessment and response. The test documents are generally quite long, from market demand mining to product and service solutions, to operation execution path, financial profit model and risk response plan. Each module forms a closely related logical chain, and all conclusions are supported by authoritative industry reports, historical operation data, patent certificates and other real materials, presenting rich content and logical closed loop.

[0066] S2: The file to be tested is segmented, and a causal graph is constructed based on the segmentation results. The nodes of the causal graph are the text content corresponding to the segmentation and standardization of the file to be tested, and the edges of the causal graph are the causal logic between the segments.

[0067] In this embodiment, the core transformation from unstructured text to a structured logical framework is achieved. On the one hand, the file to be tested is divided into standardized segments according to semantic integrity, and then the segmented text is formatted to ensure that each causal graph node can accurately correspond to the core information unit in the file, avoiding ambiguity caused by text format chaos in subsequent analysis. On the other hand, by identifying the causal logic between segments and constructing causal edges, the originally scattered text fragments are transformed into a causal graph with clear logical links. This not only clearly presents the inherent relationship of the file content, but also transforms the abstract text information into a graphical carrier that can be quantified and analyzed later.

[0068] S3: Optimize the causal graph using graph optimization rules to obtain a structured causal graph;

[0069] In this embodiment, the implementation of the causal graph is optimized according to the graph optimization rules, which realizes the upgrade from the initial graphical carrier to the standardized and highly reliable analysis carrier. In the optimization process, the redundant or duplicate nodes that may exist in the initial causal graph are effectively eliminated by the preset optimization rules, and a structured causal graph is obtained. This step optimizes the causal graph from "only presenting logic" to "quantifiable and traceable structured data", ensuring that the subsequent encoding and compression process is efficient and does not lose key information.

[0070] S4: Compress and encode the structured cause-effect graph to obtain a structured text file;

[0071] In this embodiment, the nodes and edges in the structured causal graph are first converted into machine-recognizable symbol sequences by quaternary combination encoding and binary relation symbol pairing technology, which solves the problem that graph structure data is difficult to directly input into the model. Then, through compression encoding, data redundancy is greatly reduced, storage volume and subsequent transmission costs are reduced while the core structure of the causal graph is fully preserved, and a structured text file with readability and adaptability is generated.

[0072] S5: Input the structured text file into the preset review report generation model, and output the overall project overview, project evaluation report and overall project score through the review report generation model to obtain the final review report.

[0073] In this embodiment, the review report generation model performs multi-dimensional analysis based on the causal logic in a structured text file, extracts information such as the core background and overall situation of the project to form an overall project overview, analyzes the advantages and disadvantages in different dimensions to form a project evaluation report, and calculates the overall project score by combining the causal edge weights and node levels. Finally, the overall project overview, project evaluation report and overall project score are integrated to generate a review report. This not only provides accurate data support for project decision-making, but also clarifies the core direction for subsequent project optimization through clear causal logic and score explanations, thereby improving the accuracy, objectivity and intelligence of project review.

[0074] The specific steps for S2 are as follows:

[0075] S2.1: Initialize the cause-effect graph by inserting the overall project score as a graph node into the cause-effect graph;

[0076] Specifically, causal graphs are a special data structure based on association graphs. They possess the core characteristics of graph structures: nodes carry information and edges represent associations. However, due to their "causal orientation," they also have unique attributes that distinguish them from ordinary association graphs. Causal graphs are directed weighted graphs, where all edges are directed and strictly follow the logical direction of "cause to effect." By quantifying the direction and weight of edges, discrete information units are transformed into traceable and computable causal links, perfectly meeting the review requirements of "logical closure and data-driven" business plans.

[0077] Furthermore, to ensure that the causal graph has logical integrity and adaptability to project review from the beginning of its construction, the final node needs to be determined in conjunction with the core objectives of the project review during the initialization phase. In this embodiment, the overall project score needs to serve as the core conclusion carrier of the project review report. Therefore, the overall project score is inserted into the causal graph as the core result node (Z) during the initialization phase of the causal graph. It is the endpoint of the causal relationship of all intermediate cause nodes, ultimately forming a "multi-cause-single-result" aggregated causal structure, ensuring that the overall score can accurately reflect the comprehensive impact of each module indicator.

[0078] S2.2: Perform data cleaning and text standardization on the text content corresponding to the segmented file to be tested to obtain standardized segments;

[0079] Specifically, in the text content preprocessing stage, character-level cleaning is first performed, using regular expressions to remove garbled characters, meaningless special symbols, and invisible characters, while compressing consecutive spaces and removing leading and trailing spaces to ensure that the node text does not contain invalid characters. Second, format-level cleaning is performed, removing headers, footers, page numbers, and annotation marks, and simultaneously correcting sentence breaks caused by segmentation errors to avoid incomplete node information or the inclusion of irrelevant format content. Finally, format standardization is performed, unifying character encoding and synonyms to ensure that the same node concept is expressed consistently across different segments.

[0080] Furthermore, the consistency of the node structure in the causal graph is a core prerequisite for ensuring the clarity of causal logic. It provides high-quality input for the standardized definition of nodes. If there are differences in the node structure, it will not only directly cause ambiguity in the meaning of the nodes, but also destroy the accuracy of the causal chain, thus avoiding logical breaks in the nodes caused by textual descriptions or structural confusion.

[0081] S2.3: Input the standardized segments into the semantic analysis model, and filter the core topic words of the standardized segments by calculating and sorting the semantic weights of the words in the standardized segments;

[0082] Specifically, the standardized segmented text is first split into independent words, and a customized stop word list is loaded in combination with the project review scenario. At the same time, entity recognition is performed on compound words to ensure that professional terms are not split, so as to retain complete semantic units for subsequent weight calculation. Secondly, based on the standardized word sequence, semantic features are extracted to transform the discrete text into numerical features that the model can calculate.

[0083] Furthermore, the weights are calculated using an appropriate method based on semantic features. If word distinguishability is emphasized, a word frequency and inverse document frequency (IVF) algorithm is used. First, the word frequency in the current segment and the IVF across all segments are calculated. The basic weight is obtained by multiplying the word frequency and IVF, highlighting words that are high-frequency in the current segment but low-frequency in other segments. If contextual semantic association is emphasized, a pre-trained language model is used to calculate the cosine similarity between the word embedding vector and the overall semantic vector of the segment, assigning higher weights to words that align with the segment's theme. If the inter-word association network is emphasized, the TextRank algorithm is used, based on word co-occurrence relationships. A graph structure is constructed, and the importance scores of nodes are calculated iteratively to extract the weights of core related terms such as "progress management - key tasks - delay". Then, multi-dimensional weight integration is performed. If multiple weight calculation methods are used, a weighted average is used to obtain the comprehensive weight of the words to balance the semantic contributions of different dimensions. Finally, the comprehensive weights of the words are sorted and selectively filtered. The words are sorted from high to low according to their comprehensive weights, and the words with comprehensive weights reaching a specific threshold are retained. At the same time, in combination with the causal graph review scenario, stop words and high-frequency words that are irrelevant to the review are removed, and the focus is on meaningful words. These words are used as the core topic words of the standardized segmentation.

[0084] S2.4: Determine whether there is a causal relationship between the standardized segment and the existing nodes in the causal graph based on the core keywords. If there is, insert the text content of the standardized segment as a new node into the causal graph and establish a connection with the existing nodes in the causal graph that have a causal relationship with the standardized segment. If there is no causal relationship, insert the text content of the standardized segment as a new node into the causal graph and do not establish a connection with the existing nodes in the causal graph.

[0085] Specifically, a topic-node mapping library for existing nodes in the causal graph needs to be constructed first to record the core topic words of each node. Then, the core topic words of the standardized segments are matched with the mapping library. If the topic words are completely consistent or the semantic similarity exceeds the preset similarity threshold, or if there is a verified pair in the causal rule library that can reflect causal logic, then it is included as a candidate causal relationship. If the semantic similarity is lower than the preset similarity threshold and there is no rule association, then it is initially determined that there is no causal relationship. Then, based on the matching results, the causal rule library is used for verification. Candidate associations that do not match the rules are processed through a pre-trained causal inference model, and fuzzy associations are reviewed by experts to finally determine whether a causal relationship exists.

[0086] Furthermore, new nodes are inserted based on the judgment results. When a causal relationship exists, the text content of the standardized segments is used as a new node, and a connection is established according to the causal relationship direction. If no causal relationship exists, the text content of the standardized segments is also used as a new node, and it is inserted into the causal graph as an independent node without establishing a connection with existing nodes. This ensures that the new nodes are accurately integrated into the causal graph logic and that the content and structure of the graph nodes are consistent, supporting the problem identification and root cause analysis in project review.

[0087] S2.5: Iterate through all segments to form a cause-effect graph.

[0088] Specifically, for each segment, the established rules of core keyword extraction, causal relationship determination, node insertion and connection are followed. All segments are inserted into the causal graph as nodes. Since the causal relationship between segments is local in the early stage, such as cost segments being associated only with cost-related nodes and schedule segments being associated only with schedule-related nodes, multiple scattered subgraphs will naturally form during the iteration process, covering different dimensions of project review, including cost, schedule, quality, process, etc.

[0089] Furthermore, with the generation of the causal graph, taking the "overall project score" as the core node, the potential relationship between each subgraph and the core node is analyzed one by one through semantic matching of keywords and causal rules. Causal connections are established between the key nodes of each subgraph and the "overall project score" node, gradually eliminating the dispersion of subgraphs and forming a causal network centered on the core anchor point.

[0090] The specific steps for S3 are as follows:

[0091] S3.1: Traverse all nodes in order and record the core keywords of all nodes;

[0092] Specifically, the core node of the causal graph, "overall project score," is used as the initial node. All nodes in the causal graph are traversed in a depth-first manner. If there are independent subgraphs that are not connected to the core tree, the subgraphs are traversed starting from the core node of each subgraph after the core tree traversal is completed. During the traversal, a stack of nodes to be traversed, a set of traversed nodes, and a keyword record table are constructed. The stack structure is combined to ensure that the traversal process has a clear operation carrier and recording framework, ensuring full coverage of traversed nodes and no omissions or errors in the traversal process.

[0093] S3.2: Calculate the semantic similarity of core keywords between nodes, filter nodes whose semantic similarity exceeds the preset similarity threshold, and form a set of nodes to be optimized;

[0094] Specifically, calculating semantic similarity first converts the text into a quantifiable numerical form, and then uses mathematical methods to measure the "closeness" between these values. The closer the values ​​are, the higher the semantic similarity of the text. The specific method needs to be selected based on the text granularity and scenario requirements. Commonly used methods include string-based methods, TF-IDF and cosine similarity-based methods, word embedding-based methods, and pre-trained language model-based methods. In practical applications, the method selection should be based on the scenario requirements. If efficiency, short text, and high overlap are desired, the TF-IDF and cosine similarity-based method should be chosen. If word-level semantic associations need to be captured, the word embedding-based method should be chosen. If long texts, polysemy, or complex causal sentences need to be processed, the pre-trained model-based method should be chosen. In the scenario of causal graphs for project review, the word embedding-based method is used to quickly filter candidate nodes with high topic word similarity.

[0095] Specifically, based on the semantic analysis results, nodes with semantic similarity exceeding a preset similarity threshold are extracted, such as "cost overrun" and "budget shortage", "schedule delay" and "critical task delay", "insufficient test coverage" and "high delivery defect rate", forming a candidate pool of high similarity nodes. Then, hierarchical clustering is used to cluster the nodes in the candidate pool into groups. For example, "cost overrun", "budget shortage" and "excessive use of funds" are merged into a temporary "cost association group", ultimately forming multiple groups of semantically focused and dimensionally clear semantic similarity node groups.

[0096] Furthermore, by integrating all semantically similar node groups into a set of nodes to be optimized, and then proceeding with subsequent optimization on a group-by-group basis, this approach ensures the accuracy and completeness of the optimization from the outset. On the one hand, nodes within a group share the same semantic dimension, allowing for the unified splitting of the smallest semantic units during optimization, and then merging similar units based on the group's similarity matrix, thus avoiding omissions or erroneous mergings caused by semantic fragmentation during the optimization of individual nodes. On the other hand, when migrating associated edges on a group-by-group basis, the causal edges of all original nodes within the group can be located at once and uniformly migrated to the optimized new nodes, eliminating the need to check each node individually and preventing edge breakage issues caused by the omission of a single node.

[0097] S3.3: The text content of the node to be optimized is split into the smallest semantic units, the semantic similarity between the smallest semantic units is calculated, the smallest semantic units with semantic similarity exceeding the preset similarity threshold are merged and deleted, and the smallest semantic units are integrated to obtain new nodes;

[0098] Specifically, the breakdown of the smallest semantic units must adhere to the principle of semantic integrity and indivisibility. It requires precise decomposition based on the logical characteristics of the project review node text. For complex expressions containing relationships, specific issues, or impacts on results, connectors must be removed during decomposition to extract phrase units containing independent core semantics. For expressions containing modifiers, the core limiting information must be retained to ensure that the units are concise yet fully cover the key review information, without omitting details that affect semantic understanding. Simultaneously, when decomposing the smallest semantic units, the logical connections between the original node texts can be preserved by marking the unit order. A combination of "serial number-logical relation symbol" can be used to ensure that the logical chain of the original text is not disrupted during subsequent unit merging, avoiding causal reversal or logical breaks after semantic recombination.

[0099] Furthermore, the pairwise similarity of all units is calculated using a word embedding model. Units exceeding a preset threshold are identified as duplicate associations. A merging strategy is implemented for duplicate associations, removing redundant and repetitive expressions from the merged units and retaining only the merged units, ensuring that the unit set is deduplicated and without semantic loss. Finally, the units are strictly reorganized according to the order marked during splitting, while incorporating the merged units, ensuring logical coherence and semantic focus. New nodes must meet the requirements of concise length, unambiguous semantics, and logical closed loop, laying a precise semantic foundation for subsequent migration of associated edges.

[0100] S3.4: Connect all edges that connect to the corresponding node to be optimized to the new node;

[0101] Specifically, for semantically similar node groups, the edge information connecting all nodes in the group is first recorded to form an edge information list, ensuring that no edge related to any node in the group is missed. Then, a complete migration of the edges is performed based on this information list, replacing all positions involving nodes in the group with new nodes. After the migration is completed, the integrity is verified by comparing the total number of edges in the edge information list with the number of new nodes and the number of deleted edges, ensuring that the information of all edges is not lost, the causal chain is continuous and complete, and the logic is consistent with the original association.

[0102] Please see Figure 2This invention provides a schematic diagram of the principle of merging similar nodes in a causal graph. In the diagram, similar node 1 and similar node 2 belong to a semantically similar node group. First, the edge information connecting similar node 1 and similar node 2 is recorded. The edge connected to similar node 1 is related to node N1 and node N3, and the edge connected to similar node 2 is related to node N2 and node N3. The edge information is stored to form an edge information list. Then, a complete migration of the edges is performed. All positions in all edges involving similar node 1 and similar node 2 are uniformly replaced with "merged node". The original edge from node N1 to similar node 1 and then to node N3 becomes from node N1 to merged node and then to node N3. The original edge from node N2 to similar node 2 and then to node N3 becomes from node N2 to merged node and then to node N3.

[0103] S3.5: Iteratively optimize the set of nodes to be optimized to obtain a structured causal graph.

[0104] Specifically, repeat the above operations until all nodes to be optimized are optimized, forming a structured causal graph with complete causal chains and no redundant nodes or invalid edges, providing effective and clear support for project review.

[0105] The specific steps for S4 are as follows:

[0106] S4.1: Compress and encode the structured cause-effect graph;

[0107] S4.1.1: Extract causal chains based on the causal type, weight range, time delay parameter, and node level of the edges in the structured causal graph, wherein each causal chain is represented as a symbol sequence of alternating node symbol units and relation symbol units;

[0108] Specifically, when extracting causal chains, the core attributes of the causal chain should be retained, including the causal type of the edge, the weight range, the time delay parameter, and the node level information. Edges with clear causal types, weights in the high correlation range, and quantifiable time delay parameters should be selected. Then, combined with the subordinate logic of the node level, they should be linked together to form a complete influence path. Each causal chain is ultimately presented as a symbol sequence in which node symbol units and relation symbol units are arranged alternately, ensuring that the association logic between nodes and edges in the chain is complete and traceable.

[0109] Furthermore, causal type refers to the relational nature represented by the edge, used to distinguish the directness of the influence path and provide direction for root cause identification; weight interval is the range of causal correlation strength of the edge, a quantitative range reflecting the degree of influence of the preceding node on the subsequent node, used to screen key causal chains; time lag parameter refers to the time difference between the action of the preceding node and the result produced by the subsequent node, which is the time span of causal action, and aims to provide a time basis for project risk warning and intervention timing selection; node level refers to the hierarchical position of the node in the causal graph, such as root node, first-level node, second-level node, which is the node's subordination and scope of influence, which can clarify the influence boundary of the causal chain and ensure that the scope and priority of the problem can be accurately grasped during the review.

[0110] S4.1.2: Perform quaternary combination encoding on the node symbol unit, and perform symbol pair encoding on the relation symbol unit according to the causal relationship category to obtain a set of causal strings, wherein the causal relationship category includes strong cause, weak cause, parallel causality and reverse causality;

[0111] The target cost function is determined based on the statistical distribution of node attributes, the minimum cost corresponding to the node symbol unit is calculated, and the corresponding quaternary prefix code dictionary is allocated based on the minimum cost, wherein the node attributes include node category, saliency level and time attribute;

[0112] Specifically, the statistical distribution of node attributes involves first extracting complete attribute information from all nodes to be encoded in batches to form a "node-attribute" mapping dataset, ensuring that no valid nodes are missed. Second, the statistics are split and analyzed by attribute dimension: in the node category dimension, the proportion of nodes in each category is counted; in the salience level dimension, the frequency of nodes at high, medium, and low levels is counted; in the time attribute dimension, intervals are divided according to the time delay parameter, and the distribution of the number of nodes in each interval is counted. Finally, the statistical results are presented quantitatively through a frequency distribution table or probability density map, providing objective data support for the subsequent parameter setting of the objective cost function.

[0113] Furthermore, the objective cost function is determined as follows: Where 'a' represents a combination of node attributes. This represents the probability distribution of node attribute combinations in a causal chain. This represents the encoded string assigned to the combination of node attributes. This indicates the encoding length of the node attribute combination. This represents the expected average code length of the encoding scheme;

[0114] Please see Figure 3This invention relates to a schematic diagram of quaternary combination encoding of node symbol units. Based on the statistical distribution of node attributes, the basic cost of each node symbol unit is calculated. Then, a quaternary encoding iterative process is adopted to sort all node symbol units in ascending order of basic cost. Each time, the four nodes with the lowest cost are selected and merged into one "virtual node". The sum of the costs of the four nodes is calculated as the cost of the virtual node, and quaternary branch codes of 0, 1, 2, and 3 are assigned to these four nodes respectively. The above merging and allocation operations are repeated until all nodes are finally merged into one root node. At this time, the "branch code sequence" from the root node to itself of each node symbol unit is its corresponding quaternary prefix code. Finally, the "node symbol unit-quaternary prefix code" are matched one by one to form a quaternary prefix code dictionary, ensuring that the encoding length of high-frequency and high-importance nodes is minimized.

[0115] Based on the relation symbol unit, the corresponding causal relationship category is quantized into discrete buckets and mapped into binary relation symbol pairs, wherein the binary relation symbol pairs include causal relationship categories and time delay parameters corresponding to the causal relationship categories;

[0116] Specifically, the node attribute dimensions are first broken down to identify the key attributes that need to be quantified, ensuring coverage of the essential relationship and quantifiable features. Then, discrete bucket boundaries are defined, and the number and boundaries of each bucket are set according to business rules. Next, instance classification is implemented, matching the attribute values ​​of specific causal relationships to the corresponding buckets. Finally, bucket identifiers are assigned, and a unique symbol is assigned to each bucket, completing the transformation from raw data to discrete categories.

[0117] Furthermore, binary relation symbol pairs can accurately represent the symbol combination of the relationship between two nodes in a causal graph. They not only clarify the essential type of the relationship, but also carry the key quantitative information of causal association. The specific format can be broken down into two parts: the first part is the relation identification symbol, which is usually a combination of uppercase letters and numbers or semantic abbreviations to directly distinguish the core type of the relationship; the second part is the attribute parameter symbol, which corresponds to the time delay parameter of the causal type. This format ensures the simplicity of the symbol pair, while also fully carrying the core attributes of the relationship.

[0118] Furthermore, traversing along the causal chain, the symbol sequence is alternately output according to the order of the quaternary prefix code dictionary and the binary relation symbol pairs through a finite state transformer, resulting in a set of causal strings. The finite state transformer is a core computational model in automata theory. Essentially, it is a finite state system that maps input strings to output strings. Unlike finite state automata, it focuses more on the correspondence transformation function of "input-output". Starting from the initial state, each time an input symbol is read, the system switches to the corresponding next state according to the transition function and outputs a symbol simultaneously. When all input strings have been processed, if the system finally stops at the terminating state, the entire transformation process is valid, and the corresponding causal string is output.

[0119] S4.1.3: Decompose each causal string into fixed-length segment units, cluster the segment units according to their semantic similarity and frequency of occurrence to obtain a pattern dictionary, and replace the repeated segments in the causal string set according to the dictionary symbols of the pattern dictionary.

[0120] Please see Figure 4 The flowchart of converting a causal chain into a cluster of homologous fragments in this embodiment of the invention is as follows: Specifically, the length of a fragment unit is set according to the combination rules of nodes and relation symbols in the causal chain. Based on the semantic similarity and frequency of occurrence of the fragment units, the fragment homology vector is calculated by combining the node symbol sequence, relation symbol sequence and hierarchical position symbol of the causal string. The fragment homology vector is a multi-dimensional numerical carrier used to quantitatively describe the degree of correlation between fragment units obtained after the decomposition of the causal string in terms of causal relationship, semantic features and hierarchical positioning.

[0121] Furthermore, the fragment units are aggregated according to the fragment homology vector to obtain at least one homology fragment cluster, wherein the fragment units in each homology fragment cluster are compatible in causal structure. When aggregating fragment units according to the fragment homology vector, a vector aggregation threshold is set in combination with the attributes of the causal string, and a density clustering algorithm is used to group the fragment units.

[0122] Please see Figure 5 A schematic diagram illustrating the consensus skeleton generation principle of this invention. Figure 5 Four fragment units form a homologous fragment cluster, where N1, N2, N3, and N4 are node symbols, and R0, R1, and R2 are relation symbols, following an alternating combination pattern of node and relation symbols. The homologous fragment clusters are sequence aligned and differing sites are labeled to generate a consensus skeleton. During sequence alignment, the fragment with the most complete structure within the cluster is selected as the reference. A global sequence alignment algorithm is then used to compare other fragments within the cluster with the reference sequence symbol by symbol. Gaps (-) are inserted to compensate for differences in fragment length. For example… Figure 5The fourth segment unit is the most structurally complete segment unit. Using this fourth segment unit as the baseline, the other three segment units are globally sequence aligned with it, and gap markers (-) are inserted to compensate for segment length differences. Subsequently, the common structural sequence of all segments is extracted as the consensus skeleton. For specific value differences of variable symbols in the consensus skeleton, "replacement sites (△)" are marked; for symbols present in the reference sequence but missing in other segments, "deletion sites (□)" are marked; and for symbols absent in the reference sequence but added in other segments, "insertion sites (○)" are marked. All differing sites are highlighted using special markers, for example... Figure 5 In the middle, “N1-○(R0)-R1-N △(N2 / N4)-R □(R2)-N3” represents the common structure, or consensus skeleton, shared by all segments. ○ marks the insertion site “R0”, △ marks the replacement site “N2 / N4”, and □ marks the deletion site “R2”. Additionally, “N” represents the node symbol. "" indicates that the position is a variable node placeholder, meaning that there are different specific node symbols at this position in the same fragment cluster. Similarly, "R" represents a relation symbol. "" indicates that the position is a variable relation placeholder, meaning that the specific relation symbol at this position may be different in the same fragment cluster, or there may be cases of deletion.

[0123] Further, the differing sites are encoded to obtain dictionary symbols and corresponding variant parameter vectors. The encoding includes mutation codes and cross codes. Mutation codes are used for the replacement, insertion, and deletion of differing sites. For replacement differences, a "position-skeleton reference symbol-variant symbol set" format is used; for insertion differences, a "insertion position-reference position-insertion symbol" format is used; and for deletion differences, a "deletion position-deletion symbol" format is used. Cross codes are used to combine commutative subpaths of differing sites. For cross code encoding, functionally equivalent commutative subpaths within the cluster are first identified, and the start-end positions of the subpaths in the skeleton are marked. Then, they are encoded using a "position interval-commutative subpath set" format. Finally, all mutation codes are sorted according to the order of replacement, insertion, and deletion. Cross codes are separately categorized and integrated into variant parameter vectors, which are then bound to the dictionary symbols corresponding to the consensus skeleton, forming a dictionary symbol-variant parameter vector mapping pair. This dictionary symbol-variant parameter vector mapping pair is used to replace fragment units in the same fragment cluster.

[0124] S4.1.4: Using the pattern dictionary as the row index and the edge weight strength distribution, node hierarchy distribution, and document position distribution as the column index, calculate the occurrence frequency and average weight of the pattern dictionary in the causal string set to obtain the causal spectrum matrix, wherein the edge weight strength distribution, node hierarchy distribution, and document position distribution are calculated on the causal string according to the preset bucketing rules;

[0125] Specifically, all causal strings are traversed first, and the specific weight value of the relation symbol of each segment unit, the hierarchical identifier of the first and last nodes of the segment unit, and the position of the segment unit in its respective causal string are extracted as raw data. The raw data are then classified according to the preset bucketing rules, and the number of corresponding segments in each bucket is counted to obtain the edge weight strength distribution, node hierarchy distribution, and document position distribution.

[0126] Furthermore, using the pattern dictionary as the row index and the edge weight strength distribution, node hierarchy distribution, and document position distribution as the column index, a corresponding dictionary symbol and variant parameter vector mapping pair is matched for each fragment unit. Simultaneously, based on the specific weight value of the relation symbol of the fragment unit, the hierarchy identifier of the first and last nodes of the fragment unit, and the position of the fragment unit in its causal string, the corresponding bucket is matched. Then, the number of occurrences of the mapping pair fragment under the cross dimension of each row index and column index bucket is counted in the current bucket. Then, all the weight values ​​of the mapping pair fragment in the bucket are summed and divided by the number of occurrences to obtain the average weight of the bucket. Finally, the number of occurrences and the average weight are used as cell data to fill the corresponding positions in the causal spectrum matrix.

[0127] S4.1.5: Convert the causal spectrum matrix into a token sequence that conforms to the input format of the large model.

[0128] Specifically, based on the large model's requirements for input format and length, data is concatenated row by row in the order of [row index: column index: cell data]. Information from different columns within the same row is separated by "||", and rows are separated by a recognizable special separator "".<ROW_SPLIT> "Separator"

[0129] S4.2: The structured text file includes a standardized segmented content text file and a causal relationship information text file. The standardized segmented content text file is used to store the standardized segments obtained after data cleaning and text standardization of the segments. The causal relationship information text file is used to store the compression encoding result obtained by compressing and encoding the structured causal graph.

[0130] The specific steps for S5 are as follows:

[0131] S5.1: Obtain standardized segmented content text files, extract the core content of all segments through a semantic analysis model, and logically integrate the core content through a large text generation model to obtain an overall project overview;

[0132] Specifically, the process begins by reading standardized segmented text files and then using semantic analysis models, such as a combination of BERT and TextRank, to perform deep analysis of the segmented text. This identifies the core subjects, key actions, and impact outcomes within each segment, clarifies the logical relationships between these elements, and selects the core keywords for each segment. The extracted results are then organized according to the structure of "segment theme - core keywords - logical relationships," forming a standardized keyword set. This set provides a precise foundation for the subsequent text generation model. The standardized keyword set is then input into the large-scale text generation model. Based on the input keywords and logical prompts, the model first constructs a "general-specific-general" overview framework using contextual modeling capabilities. The framework begins by outlining the project's purpose, then elaborates on key issues and their connections across different dimensions, and concludes by summarizing the core impacts and preliminary recommendations. Finally, through manual verification or secondary model optimization, the framework ensures that the overview is coherent, logically closed, and fully presents the core aspects of the project.

[0133] S5.2: Extract the influencing factors that affect the overall project score from the causal relationship information text file, set the score and weight of the influencing factors, compare the influencing factors with the preset project evaluation criteria, calculate the overall project score, and generate a project evaluation report;

[0134] Specifically, the text file containing causal relationship information is read, and based on the causal spectrum matrix information recorded in the file, the scores and weights of each influencing factor are identified and set.

[0135] Furthermore, the identified influencing factors are compared with the preset project evaluation criteria to determine the dominant and disadvantageous influencing factors: if an influencing factor is consistent with the preset project evaluation criteria, it is determined to be a dominant influencing factor; if an influencing factor contradicts the preset project evaluation criteria, it is determined to be a disadvantageous influencing factor; if none of the identified influencing factors fully cover the preset project evaluation criteria, the uncovered preset project evaluation criteria are determined to be disadvantageous influencing factors; if an influencing factor is not mentioned in the preset project evaluation criteria, it is determined to be an invalid influencing factor and does not affect the overall project score.

[0136] Furthermore, based on the scores and weights of the advantageous influencing factors, bonus points are calculated for the project; based on the scores and weights of the disadvantageous influencing factors, deduction points are calculated for the project. The overall project score is obtained by the difference between the total bonus points and the total deduction points. Based on the advantageous and disadvantageous influencing factors, the semantic integration and structured output capabilities of the text generation model are used to elaborate on the positive value and shortcomings of the project around the core evaluation dimensions, ultimately generating a project evaluation report that is structurally complete, coherent in content, and clearly targeted.

[0137] S5.3: Integrate the overall project overview, project evaluation report, and overall project score to obtain the final review report.

[0138] Specifically, the report begins with an overall project overview, briefly presenting the project background and overall performance outline to quickly establish project understanding. Then, the project evaluation report serves as the core body, systematically elaborating on the specific value of the strengths and the profound impact of the weaknesses across different dimensions, while simultaneously listing improvement directions for each dimension. Finally, the report concludes with the overall project score and calculation logic, using quantitative data to intuitively anchor the project's performance level. This results in a complete review report that combines a macro perspective, quantitative support, and in-depth analysis, providing a comprehensive and accurate basis for subsequent project optimization and decision-making.

[0139] Example 2

[0140] Please see Figure 6 A process outcome extrapolation system based on large-scale models and causal analysis, the system comprising a causal graph construction module, a project overview generation module, a score calculation and evaluation generation module, and a report output module:

[0141] The causal graph construction module is used to identify causal relationships between segments, construct an initial causal graph, and optimize the initial causal graph to obtain a structured causal graph.

[0142] The project overview generation module is used to extract the core content of each standardized segment and logically integrate it to obtain the project overview.

[0143] The score calculation and evaluation generation module is used to quantitatively analyze the key factors affecting the project, calculate the overall project score based on the evaluation criteria, and generate a project evaluation report.

[0144] The report output module is used to integrate the overall project overview, overall project score, and project evaluation report to form the final review report.

Claims

1. A method for extrapolating process results based on large-scale models and causal analysis, characterized in that, The method includes: Obtain the file to be tested; The file to be tested is segmented, and a causal graph is constructed based on the segmentation results. The nodes of the causal graph are the text content corresponding to the segmentation and standardization of the file to be tested, and the edges of the causal graph are the causal logic between the segments. The causal graph is optimized using graph optimization rules to obtain a structured causal graph; The structured cause-effect graph is compressed and encoded, and then stored to obtain a structured text file; The structured text file is input into the preset review report generation model, which then outputs a project overview, a project evaluation report, and a project overall score to obtain the final review report. Among them, the causal graph is optimized using graph optimization rules to obtain a structured causal graph, including: Traverse all nodes in order and record the core keywords of each node; Calculate the semantic similarity of core keywords between nodes, filter out nodes whose semantic similarity exceeds a preset similarity threshold, and form a set of nodes to be optimized; The text content of the node to be optimized is split into the smallest semantic units. The semantic similarity between the smallest semantic units is calculated. The smallest semantic units with semantic similarity exceeding the preset similarity threshold are merged and deleted. The smallest semantic units are integrated to obtain new nodes. Connect all edges that connect the corresponding node to be optimized to the new node; Iterative optimization of the set of nodes to be optimized yields a structured cause-effect graph; The compression encoding of the structured causal graph includes: Causal chains are extracted based on the causal type, weight range, time delay parameter, and node level of the edges in the structured causal graph, where each causal chain is represented as a symbol sequence of alternating node symbol units and relation symbol units; The node symbol unit is encoded using quaternary combination encoding, and the relation symbol unit is encoded using symbol pairs according to the causal relationship category to obtain a set of causal strings, wherein the causal relationship category includes strong cause, weak cause, parallel causality and reverse causality; Each causal string is decomposed into fixed-length segment units. The segment units are clustered according to their semantic similarity and frequency of occurrence to obtain a pattern dictionary. Duplicate segments in the causal string set are replaced according to the dictionary symbols of the pattern dictionary. Using the pattern dictionary as the row index and the edge weight intensity distribution, node hierarchy distribution, and document location distribution as the column index, the occurrence frequency and average weight of the pattern dictionary in the causal string set are calculated to obtain the causal spectrum matrix, wherein the edge weight intensity distribution, node hierarchy distribution, and document location distribution are calculated for the causal string according to a preset bucketing rule; The causal spectrum matrix is ​​transformed into a token sequence that conforms to the input format of a large model.

2. The process result deduction method based on large model and causal analysis according to claim 1, characterized in that, Construct a cause-effect graph based on the segmentation results, including: Initialize the cause-effect graph by inserting the overall project score as a graph node into the cause-effect graph. After the file to be tested is segmented, the corresponding text content is cleaned and the segments are standardized to obtain standardized segments. The standardized segments are input into the semantic analysis model. By calculating and ranking the semantic weights of the words in the standardized segments, the core topic words of the standardized segments are selected. Based on the core keywords, determine whether there is a causal relationship between the standardized segment and existing nodes in the causal graph. If there is, insert the text content of the standardized segment as a new node into the causal graph and establish a connection with the existing nodes in the causal graph that have a causal relationship with the standardized segment. If there is no causal relationship, insert the text content of the standardized segment as a new node into the causal graph and do not establish a connection with the existing nodes in the causal graph. Iterate through all segments to form a cause-effect graph.

3. The process result deduction method based on large model and causal analysis according to claim 2, characterized in that, The node symbol units are quaternary combined encodings, and the relation symbol units are symbol-pair encodings based on causal relationship categories to obtain a set of causal strings, including: The objective cost function is determined based on the statistical distribution of node attributes. The minimum cost corresponding to each node symbol unit is calculated, and a corresponding quaternary prefix code dictionary is assigned based on the minimum cost. The node attributes include node category, saliency level, and time attribute. The objective cost function is... Where 'a' represents a combination of node attributes. This represents the probability distribution of node attribute combinations in a causal chain. This represents the encoded string assigned to the combination of node attributes. This indicates the encoding length of the node attribute combination. This represents the expected average code length of the encoding scheme; Based on the relation symbol unit, the corresponding causal relationship category is quantized into discrete buckets and mapped into binary relation symbol pairs, wherein the binary relation symbol pairs include causal relationship categories and time delay parameters corresponding to the causal relationship categories; Traversing along the causal chain, the symbol sequence is alternately output according to the order of the quaternary prefix code dictionary and the binary relation symbol pairs through a finite state transformer to obtain a set of causal strings.

4. The process result deduction method based on large model and causal analysis according to claim 1, characterized in that, Clustering the fragment units based on their semantic similarity and frequency of occurrence yields a pattern dictionary. Repeated fragments in the causal string set are then replaced using dictionary symbols from the pattern dictionary, including: Based on the semantic similarity and frequency of occurrence, the fragment homology vector is calculated by combining the node symbol sequence, relation symbol sequence, and hierarchical position symbol of the causal string; The fragment units are aggregated according to the fragment homology vector to obtain at least one homology fragment cluster, wherein the fragment units in each homology fragment cluster are causally compatible. Sequence alignment is performed on the homologous fragment clusters to generate a consensus skeleton, and the differing sites are marked in the consensus skeleton; Encoding is performed on the differential sites to obtain dictionary symbols and corresponding variant parameter vectors, wherein the encoding includes mutation codes and cross codes. The mutation codes are used for substitution, insertion, and deletion of differential sites, and the cross codes are used for combining commutative subpaths of differential sites. Replace fragment units in the same fragment cluster according to the dictionary symbols and the variant parameter vectors corresponding to the dictionary symbols.

5. The process result deduction method based on large model and causal analysis according to claim 1, characterized in that, The structured text file includes a standardized segmented content text file and a causal relationship information text file. The standardized segmented content text file is used to store the standardized segments obtained after data cleaning and text standardization of the segments. The causal relationship information text file is used to store the compressed encoding result obtained by compressing and encoding the structured causal graph.

6. The process result deduction method based on large model and causal analysis according to claim 1, characterized in that, The structured text file is input into a preset review report generation model, which then outputs a project overview, a project evaluation report, and an overall project score to obtain the final review report, including: Obtain standardized segmented text files, extract the core content of all segments through a semantic analysis model, and logically integrate the core content through a large text generation model to obtain an overall project overview. Extract the influencing factors that affect the overall project score from the causal relationship information text file, set the score and weight of the influencing factors, compare the influencing factors with the preset project evaluation criteria, calculate the overall project score and generate a project evaluation report; The final review report is obtained by integrating the overall project overview, project evaluation report, and overall project score.

7. The process result deduction method based on large model and causal analysis according to claim 6, characterized in that, The influencing factors are compared with the preset project evaluation criteria, the overall project score is calculated, and a project evaluation report is generated, including: Compare the influencing factors with the preset project evaluation criteria to determine the advantageous and disadvantageous influencing factors; Based on the scores and weights of the advantageous factors, the bonus points for the project are calculated, and based on the scores and weights of the disadvantageous factors, the deduction points for the project are calculated, and finally the overall score of the project is obtained. Based on the factors influencing strengths and weaknesses, a large model is generated from text to output a project evaluation report.

8. A process result deduction system based on large-scale modeling and causal analysis, used to implement the process result deduction method based on large-scale modeling and causal analysis as described in any one of claims 1-7, characterized in that, The system includes a cause-effect graph construction module, a project overview generation module, a score calculation and evaluation generation module, and a report output module. The causal graph construction module is used to identify causal relationships between segments, construct an initial causal graph, and optimize the initial causal graph to obtain a structured causal graph. The project overview generation module is used to extract the core content of each standardized segment and logically integrate it to obtain the project overview. The score calculation and evaluation generation module is used to quantitatively analyze the key factors affecting the project, calculate the overall project score based on the evaluation criteria, and generate a project evaluation report. The report output module is used to integrate the overall project overview, overall project score, and project evaluation report to form the final review report.

Images