Deep learning-based commercial experiment report automatic generation system and method

Abstract

The invention discloses a business experiment report automatic generation system and method based on deep learning, and the method comprises the following steps: obtaining an experiment data set, executing the structured processing, and generating an experiment fact input structure; constructing a hierarchical semantic reasoning graph structure, and establishing a cross-layer semantic association relationship; constructing an improved NBFNet reasoning model, executing cross-layer path relaxation propagation processing, and generating a candidate reasoning path set; constructing a path-level semantic memory structure and executing cumulative updating processing to generate a path semantic representation set; in the improved NBFNet reasoning model, multi-path competition type conclusion forming processing is executed, and a conclusion support distribution result is generated; executing conclusion type judgment processing to generate a conclusion type set; and carrying out commercial experiment report generation processing to generate a commercial experiment report. Structured reasoning and traceable expression of commercial experiment conclusion generation are realized, and report reliability and consistency are improved.

Description

Technical Field

[0001] This invention relates to the field of business data statistical analysis and empirical research, and in particular to a system and method for automatically generating business lab reports based on deep learning. Background Technology

[0002] Experimental research and empirical analysis in business studies typically rely on large amounts of experimental data. Through statistical testing, effect size analysis, and condition control, experimental conclusions are drawn and experimental reports are written. These reports are widely used in economics, management, finance, and market research, and their quality directly impacts the credibility and reusability of the research conclusions. However, the current process of generating business experimental reports primarily relies on manual labor. Researchers need to repeatedly compare and integrate various statistical results, experimental conditions, and analytical conclusions, making the entire process time-consuming and highly dependent on personal experience.

[0003] In the existing technology, some analytical tools can perform statistical calculations on experimental data or output basic analysis results. However, these tools usually only stay at the numerical level and lack systematic modeling of the relationship between experimental facts, analytical reasoning process and conclusion expression. They are difficult to comprehensively evaluate multiple analytical paths and cannot effectively characterize the evidence conflicts and conditional differences in the conclusion formation process. As a result, the generated analytical conclusions are not interpretable and have weak traceability.

[0004] Furthermore, existing automated report generation solutions mostly use general text generation or template filling methods, failing to fully integrate information on statistical significance, sample size, condition constraints, and consistency of evidence in the experimental analysis process. The generated report content often remains at the descriptive level, lacking quantitative support for the reliability of the conclusions, and is difficult to meet the requirements of rigor and auditability for business experiments.

[0005] Therefore, how to provide a system and method for automatically generating business lab reports based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose an automatic business experiment report generation system and method based on deep learning. This invention constructs a hierarchical semantic reasoning graph structure based on the experimental fact input structure, introduces an improved NBFNet reasoning model, and performs cross-layer path relaxation propagation processing on the cross-layer semantic relationships between the experimental fact node set, the analysis and reasoning node set, and the conclusion expression node set to generate a candidate reasoning path set. Furthermore, it performs cumulative updates on the node contributions, edge constraints, and evidence trajectories in the reasoning paths through a path-level semantic memory structure, forming a path semantic representation set. Based on this, this invention generates a conclusion support distribution result through multi-path competitive conclusion formation processing, and completes the conclusion type determination processing based on the conclusion support distribution result. Finally, it combines the candidate reasoning path set, the path semantic representation set, and the conclusion type set to generate a business experiment report. This invention achieves the structured generation of experimental analysis process, conclusion derivation basis, and conclusion reliability information, possessing the advantages of traceable conclusion formation process, characterizable evidence conflicts, and clear conclusion reliability, effectively improving the standardization and credibility of business experiment report generation.

[0007] The method for automatically generating business lab reports based on deep learning according to embodiments of the present invention includes the following steps: Obtain the experimental data set, perform structured processing, and generate the experimental fact input structure; Based on the experimental fact input structure, a hierarchical semantic reasoning graph structure is constructed, including the experimental fact layer, the analysis and reasoning layer, and the conclusion expression layer, and cross-layer semantic associations are established. An improved NBFNet inference model is constructed based on a hierarchical semantic inference graph structure. The experimental facts are input into the model, and cross-layer path relaxation propagation processing is performed based on cross-layer semantic relationships to generate a set of candidate inference paths. For each candidate reasoning path in the candidate reasoning path set, construct a path-level semantic memory structure, perform cumulative update processing, and generate a path semantic representation set; The path semantic representation set is input into the improved NBFNet inference model, and multi-path competitive conclusion formation processing is performed to generate conclusion support distribution results. Based on the conclusion support distribution results, perform conclusion type determination processing on each conclusion expression node to generate a conclusion type set; Based on the candidate reasoning path set, path semantic representation set, and conclusion type set, a business lab report generation process is performed to generate a business lab report.

[0008] Optionally, the generation of the experimental fact input structure includes: Obtain the raw experimental data related to business experiments, divide it into experimental condition data, experimental variable data and raw statistical result data, perform integrity verification processing, and obtain a valid set of experimental data. Perform data consistency processing on the valid experimental data set to generate a standardized experimental data set; Based on the standardized experimental dataset, condition encoding is performed on the experimental condition data to generate experimental condition information; variable mapping is performed on the experimental variable data to generate experimental variable information; and statistical result processing is performed on the raw statistical result data to generate raw statistical result information. The experimental conditions, experimental variables, and raw statistical results are organized in a structured manner to generate an experimental fact input structure.

[0009] Optionally, the construction of the hierarchical semantic reasoning graph structure includes: Based on the experimental fact input structure, an experimental fact layer and a set of experimental fact nodes are constructed in the hierarchical semantic reasoning graph structure, an analytical reasoning layer and a set of analytical reasoning nodes are constructed, and a conclusion expression layer and a set of conclusion expression nodes are constructed. Establish cross-layer semantic associations from facts to reasoning based on the experimental fact node set and the analytical reasoning node set; A cross-layer semantic relationship between reasoning and conclusion is established based on the set of analytical reasoning nodes and the set of conclusion expression nodes; Invalid edge removal is performed on cross-layer connections that do not satisfy the cross-layer semantic association constraints from facts to reasoning and from reasoning to conclusion, thereby generating a hierarchical semantic reasoning graph structure.

[0010] Optionally, the generation of the candidate inference path set includes: Based on the hierarchical semantic reasoning graph structure, an improved NBFNet reasoning model is constructed, and path order constraint rules are set to limit the propagation order of reasoning paths and the upper limit of the path length of reasoning paths. The experimental facts are input into the structure and then into the improved NBFNet inference model. During the cross-layer path relaxation propagation process, the edge weights of the cross-layer semantic associations are updated to generate updated edge weights. During the cross-layer path relaxation propagation process, a feasibility determination process is performed on each candidate inference path; After completing the cross-layer path relaxation propagation process, for each conclusion expression node, the inference path score pointing to the conclusion expression node is calculated based on the updated edge weight. The inference path scores are sorted and the Top-K inference paths are retained to generate a candidate inference path set.

[0011] Optionally, the generation of the path semantic representation set includes: During the cross-layer path relaxation propagation process in the improved NBFNet inference model, a path-level semantic memory structure is constructed for each candidate inference path, including node contribution cache, edge constraint cache, and evidence trajectory cache. The path-level semantic memory structure is cumulatively updated based on the candidate reasoning paths; After completing the cumulative update processing of the path-level semantic memory structure, the node contribution cache, edge constraint cache, and evidence trajectory cache are concatenated to generate a path semantic representation set.

[0012] Optionally, the conclusion supports the generation of distribution results including: Input the path semantic representation set into the improved NBFNet inference model, and for each conclusion expression node, obtain the path semantic representation subset pointing to that conclusion expression node; Based on a subset of path semantic representations, a support strength evaluation process is performed on each path semantic representation to calculate the single-path support strength value of the corresponding path. Based on the single path support strength value, conflict relationship modeling is performed to identify whether there are directional conflicts and conditional conflicts in the path semantic representation subset corresponding to the same conclusion expression node. When the proportion of directional conflicts exceeds the preset conflict threshold, the support ratio of the high support interval in the conclusion support distribution results of the corresponding conclusion expression node decreases, while the support ratio of the medium support interval and the low support interval increases. When a condition conflict is determined, condition cluster partitioning is performed to generate the conclusion support distribution results.

[0013] Optionally, the generation of the conclusion type set includes: Based on the conclusion support distribution results, the conclusion support distribution index is calculated for each conclusion expression node, including total support strength, conflict degree and effective evidence coverage. Based on the conclusion support distribution index, a conclusion type determination process is performed on each conclusion expression node. A conclusion type is assigned to the corresponding conclusion expression node according to a preset threshold condition. The conclusion types include strongly supported conclusion types, conditionally supported conclusion types, and insufficient evidence conclusion types. The conclusion types are summarized to generate a conclusion type set.

[0014] Optionally, the generation of the business school lab report includes: Based on the candidate reasoning path set and the path semantic representation set, an experimental analysis process text segment is generated, which is constructed by referencing the node sequence corresponding to a preset number of high-scoring candidate reasoning paths and the evidence summary vector in the path semantic representation; Based on the conclusion support distribution results and conclusion type set, a conclusion derivation explanation text segment is generated. It is constructed by referencing the conclusion support distribution results of the corresponding conclusion expression node and the path identification information of the main supporting candidate reasoning paths. Based on the conclusion type set and conclusion support distribution results, a conclusion reliability description text segment is generated, which is constructed by referencing the conclusion type, conflict degree and effective evidence coverage of the corresponding conclusion expression node; In the process of generating the experimental analysis process text segment, the conclusion derivation explanation text segment, and the conclusion reliability description text segment, each sentence in each text segment is bound to the corresponding candidate reasoning path identifier and evidence summary identifier to generate traceable annotation results, which are written into the business experiment report together with the corresponding text segment to generate the business experiment report.

[0015] An automatic business lab report generation system based on deep learning according to an embodiment of the present invention includes: The experimental fact input structure generation module is used to acquire experimental datasets, merge them, perform structured processing, and generate experimental fact input structures. The hierarchical semantic reasoning graph structure construction module is used to construct a hierarchical semantic reasoning graph structure based on the experimental fact input structure. The hierarchical semantic reasoning graph structure includes an experimental fact layer, an analysis and reasoning layer, and a conclusion expression layer, and establishes cross-layer semantic relationships. An improved NBFNet inference model and candidate inference path set generation module is used to construct an improved NBFNet inference model based on a hierarchical semantic inference graph structure. The model takes experimental facts as input, performs cross-layer path relaxation propagation processing based on cross-layer semantic relationships, and generates a candidate inference path set. The path-level semantic memory structure and path semantic representation set generation module is used to construct a path-level semantic memory structure for each candidate reasoning path in the candidate reasoning path set, perform cumulative update processing, and generate a path semantic representation set. The conclusion support distribution generation module is used to input the path semantic representation set into the improved NBFNet inference model, perform multi-path competitive conclusion formation processing, and generate conclusion support distribution results. The conclusion type set generation module is used to perform conclusion type determination processing on each conclusion expression node in the conclusion expression node set based on the conclusion support distribution results, and generate a conclusion type set. The Business Experiment Report Generation Module is used to generate business experiment reports based on a set of candidate reasoning paths, a set of path semantic representations, and a set of conclusion types.

[0016] The beneficial effects of this invention are: First, this invention constructs a hierarchical semantic reasoning graph structure, which hierarchically organizes the experimental condition information, experimental variable information, and raw statistical result information in the experimental fact input structure. It also establishes constrained cross-layer semantic associations between the experimental fact layer, the analysis and reasoning layer, and the conclusion expression layer. This transforms the experimental analysis process from discrete data processing into a structured and computable reasoning process, effectively avoiding the problems of relying on human experience and unclear reasoning links in the generation of experimental conclusions in existing technologies. This significantly improves the logical consistency and controllability of the business experimental analysis process.

[0017] Secondly, this invention introduces an improved NBFNet inference model and performs cross-layer path relaxation propagation processing. Through path order constraints, inference path feasibility determination, and path score ranking mechanisms, a set of candidate inference paths is generated. Combined with a path-level semantic memory structure, node contributions, edge constraints, and evidence trajectories are cumulatively updated to form a set of path semantic representations. This enables a comprehensive characterization of the evidence strength and constraint consistency of multiple inference paths, overcoming the problems of scattered statistical results and difficulty in quantifying evidence conflicts in existing technologies.

[0018] Furthermore, this invention uses multi-path competitive conclusion formation processing and conclusion support distribution modeling to uniformly model the support strength, directional conflict, and conditional conflict of different reasoning paths, and further completes the conclusion type determination processing, enabling the conclusion expression node to distinguish between strongly supported conclusion types, conditionally supported conclusion types, and insufficient evidence conclusion types. This mechanism reduces the risk of erroneous conclusions being directly output and enhances the objectivity and stability of conclusion reliability judgment. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0020] Figure 1 This is an overall flowchart of a deep learning-based method for automatically generating business lab reports proposed in this invention. Figure 2 This is a schematic diagram of cross-layer path relaxation propagation and path-level semantic memory construction based on the hierarchical semantic reasoning graph structure and the improved NBFNet reasoning model in this invention; Figure 3 This is a schematic diagram of the multi-path competitive conclusion formation and conclusion type determination process in this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0022] refer to Figure 1-3 A deep learning-based method for automatically generating business lab reports includes the following steps: Obtain a set of experimental data related to business experiments, perform structured processing on the experimental data set, and generate an experimental fact input structure. The experimental fact input structure includes experimental condition information, experimental variable information, and raw statistical result information. Based on the experimental fact input structure, a hierarchical semantic reasoning graph structure is constructed. In the hierarchical semantic reasoning graph structure, an experimental fact layer, an analysis and reasoning layer, and a conclusion expression layer are constructed. An experimental fact node set is constructed in the experimental fact layer, an analysis and reasoning node set is constructed in the analysis and reasoning layer, and a conclusion expression node set is constructed in the conclusion expression layer. Cross-layer semantic associations are established, which connect the experimental fact node set and the analysis and reasoning node set, and also connect the analysis and reasoning node set and the conclusion expression node set. An improved NBFNet inference model is constructed based on a hierarchical semantic inference graph structure. The experimental facts are input into the improved NBFNet inference model. Cross-layer path relaxation propagation is performed based on cross-layer semantic relationships to generate a set of candidate inference paths. Each candidate inference path in the set of candidate inference paths starts from the starting node in the set of experimental facts nodes, passes through the set of analysis and inference nodes, and reaches the set of conclusion expression nodes. During the cross-layer path relaxation propagation process in the improved NBFNet inference model, a path-level semantic memory structure is constructed for each candidate inference path in the candidate inference path set. The path-level semantic memory structure is cumulatively updated based on the node sequence and edge sequence of the candidate inference path to generate a path semantic representation set. Each path semantic representation in the path semantic representation set corresponds to a candidate inference path. The path semantic representation set is input into the improved NBFNet inference model, and a multi-path competitive conclusion formation process is performed. For each conclusion expression node in the conclusion expression node set, a subset of path semantic representations pointing to that conclusion expression node is obtained. Based on the subset of path semantic representations, support strength evaluation and conflict relationship modeling are performed to generate conclusion support distribution results. Based on the conclusion support distribution results, a conclusion type determination process is performed on each conclusion expression node in the conclusion expression node set to generate a conclusion type set. The conclusion type set includes strongly supported conclusion types, conditionally supported conclusion types, and insufficient evidence conclusion types. The conclusion type determination process assigns a conclusion type from the conclusion type set to each conclusion expression node. Based on the candidate reasoning path set, path semantic representation set, and conclusion type set, a business experiment report generation process is performed to generate a business experiment report. The business experiment report includes a text segment of the experiment analysis process, a text segment of the conclusion derivation explanation, and a text segment of the conclusion reliability description.

[0023] In this embodiment, the generation of the experimental fact input structure includes: Obtain the raw experimental data related to business experiments. The raw experimental data is divided into experimental condition data, experimental variable data and raw statistical result data according to the data source and data attributes. The raw experimental data is then subjected to integrity verification to obtain a valid set of experimental data. Perform data consistency processing on the valid experimental dataset to unify the representation dimensions, measurement scales, and data formats of experimental condition data, experimental variable data, and raw statistical result data, and generate a standardized experimental dataset. Based on the standardized experimental dataset, condition encoding is performed on the experimental condition data to generate experimental condition information; variable mapping is performed on the experimental variable data to generate experimental variable information; and statistical result processing is performed on the raw statistical result data to generate raw statistical result information. The experimental condition information, experimental variable information, and raw statistical result information are organized in a structured manner to generate an experimental fact input structure, which serves as the input basis for constructing a hierarchical semantic reasoning graph structure.

[0024] In this embodiment, the construction of the hierarchical semantic reasoning graph structure includes: Based on the experimental fact input structure, an experimental fact layer is constructed in the hierarchical semantic reasoning graph structure. A set of experimental fact nodes is constructed in the experimental fact layer. Each experimental fact node contains experimental condition field, experimental variable field, statistical field, confidence interval field, significance marker field and sample size field. An analytical reasoning layer is constructed in the hierarchical semantic reasoning graph structure. An analytical reasoning node set is constructed in the analytical reasoning layer. Each analytical reasoning node contains a test type field, an effect size type field, a control variable set field, a hypothesis direction field, and an inference rule identifier field. In the hierarchical semantic reasoning graph structure, a conclusion expression layer is constructed, and a set of conclusion expression nodes is constructed in the conclusion expression layer. Each conclusion expression node contains a conclusion template ID field, an applicable condition field, a reliability threshold parameter field, and a list of cited evidence field. A cross-layer semantic association between facts and reasoning is established based on the set of experimental fact nodes and the set of analytical reasoning nodes. The cross-layer semantic association between facts and reasoning is established only when the determination satisfies the correspondence between the test type field and the experimental variable field and the variable role of the experimental variable field is matched. Based on the set of analytical reasoning nodes and the set of conclusion expression nodes, a cross-layer semantic relationship from reasoning to conclusion is established. The cross-layer semantic relationship from reasoning to conclusion is established only when the determination satisfies the consistency of the inference rule identifier field and the conclusion template corresponding to the conclusion template ID field is available. For cross-layer connections that do not satisfy the cross-layer semantic association constraints from fact to reasoning and cross-layer semantic association constraints from reasoning to conclusion, invalid edge removal is performed. The corresponding cross-layer connections are marked as unavailable and the reason code is recorded, generating a hierarchical semantic reasoning graph structure with consistent constraints.

[0025] In this embodiment, the generation of the candidate inference path set includes: Based on the hierarchical semantic reasoning graph structure, an improved NBFNet reasoning model is constructed. In the improved NBFNet reasoning model, a path order constraint rule is set. The path order constraint rule limits the propagation order of the reasoning path to the experimental fact layer, the analysis reasoning layer, and then the conclusion expression layer. It also limits the upper limit of the path length of the reasoning path, so that the reasoning path is not allowed to jump back across layers during the propagation process. The experimental facts are input into the improved NBFNet inference model. During the cross-layer path relaxation propagation, the cross-layer semantic associations in the hierarchical semantic inference graph structure are updated by edge weights. The edge weight update is based on the statistical field, significance marker field, and sample size field in the experimental fact node set, and combined with the effect size type field in the analysis and inference node set to generate updated edge weights for path relaxation propagation. During the cross-layer path relaxation propagation process, a feasibility determination process is performed on each candidate inference path. The feasibility determination process is based on the missing label status of the sample size field and the experimental variable field in the experimental fact node set. When the sample size field does not meet the preset validity requirements or the experimental variable field has a missing label status, the corresponding inference path is determined to be infeasible and marked as unavailable. After completing the cross-layer path relaxation propagation process, for each conclusion expression node in the conclusion expression node set, the inference path score pointing to the conclusion expression node is calculated based on the updated edge weight. The inference path scores are sorted and the Top-K inference paths are retained to generate a candidate inference path set. Each candidate inference path in the candidate inference path set contains the corresponding path score.

[0026] In this embodiment, the generation of the path semantic representation set includes: During the cross-layer path relaxation propagation process in the improved NBFNet inference model, a path-level semantic memory structure is constructed for each candidate inference path in the candidate inference path set. The path-level semantic memory structure includes a node contribution cache, an edge constraint cache, and an evidence trajectory cache. The node contribution cache is used to store the contribution vector of each node in the path to the corresponding conclusion expression node. The edge constraint cache is used to store the semantic type, constraint satisfaction flag, and conflict flag of each edge in the path. The evidence trajectory cache is used to store the statistical test result summary, effect size summary, and condition trigger summary corresponding to the path. Based on the node arrangement order and connection relationship between nodes in the candidate inference path, the path-level semantic memory structure is subjected to cumulative update processing. The cumulative update processing writes the contribution vector corresponding to each node in the path into the node contribution cache in the order of appearance of nodes and edges in the candidate inference path, and writes the semantic type, constraint satisfaction flag and conflict flag corresponding to each edge in the path into the edge constraint cache. When it is determined that the conflict flag of the corresponding edge in the edge constraint cache is valid, the conflict flag is immediately recorded and the cumulative influence of the corresponding contribution vector in the node contribution cache after the conflicting edge is reduced. After completing the cumulative update of the path-level semantic memory structure, the node contribution cache, edge constraint cache, and evidence trajectory cache are concatenated to generate a path semantic representation set. Each path semantic representation in the path semantic representation set consists of a contribution vector, a constraint vector, and an evidence summary vector, and corresponds one-to-one with the candidate reasoning paths in the candidate reasoning path set.

[0027] In this embodiment, the conclusion supporting the generation of distribution results includes: Input the path semantic representation set into the improved NBFNet inference model, and for each conclusion expression node in the conclusion expression node set, obtain the path semantic representation subset pointing to that conclusion expression node; Based on a subset of path semantic representations, a support strength evaluation process is performed on each path semantic representation. The support strength evaluation process takes the path score, the evidence summary vector in the path semantic representation, and the proportion of corresponding edge constraint satisfaction in the path semantic representation as inputs to calculate the single path support strength value of the corresponding path. Based on the single-path support strength value in the path semantic representation subset, conflict relationship modeling is performed. The conflict relationship modeling is used to identify whether there are directional conflicts and conditional conflicts in the path semantic representation subset corresponding to the same conclusion expression node. Directional conflicts are when there are path semantic representations with opposite support directions in the same path semantic representation subset corresponding to the same conclusion expression node. Conditional conflicts are when the experimental condition information of different path dependencies in the path semantic representation subset is inconsistent. When the proportion of directional conflicts exceeds the preset conflict threshold, the support ratio of high support intervals in the conclusion support distribution results of the corresponding conclusion expression node decreases, while the support ratio of medium and low support intervals increases. When a condition conflict is determined, condition cluster partitioning is performed based on the experimental condition information of path dependence in the path semantic representation subset to generate the condition set corresponding to the condition support conclusion and generate the conclusion support distribution results.

[0028] In this embodiment, the generation of the conclusion type set includes: Based on the conclusion support distribution results, a conclusion support distribution index is calculated for each conclusion expression node in the conclusion expression node set. The conclusion support distribution index includes total support strength, conflict degree, and effective evidence coverage. The total support strength is determined by the cumulative support strength of the corresponding conclusion expression node in the conclusion support distribution results. The conflict degree is determined by the support ratio corresponding to different support directions in the conclusion support distribution results. The effective evidence coverage is determined by the ratio of the number of inference paths judged as feasible in the conclusion support distribution results to the total number of inference paths pointing to the conclusion expression node. Based on the conclusion support distribution index, a conclusion type determination process is performed on each conclusion expression node in the conclusion expression node set. A conclusion type is assigned to the corresponding conclusion expression node according to a preset threshold condition. The conclusion types include strongly supported conclusion types, conditionally supported conclusion types, and insufficient evidence conclusion types. When the total support strength is not less than the preset support threshold, the conflict degree is not greater than the preset conflict threshold, and the effective evidence coverage is not less than the preset coverage threshold, the corresponding conclusion expression node is determined to be a strongly supported conclusion type. When the total support strength is not less than the preset support threshold, the conflict degree is greater than the preset conflict threshold, or the effective evidence coverage is less than the preset coverage threshold, the corresponding conclusion expression node is determined to be a conditionally supported conclusion type. When the total support strength is less than the preset support threshold, or the number of feasible reasoning paths pointing to the conclusion expression node is less than the preset minimum number of paths, the corresponding conclusion expression node is determined to be a conclusion type with insufficient evidence. The conclusion types corresponding to each conclusion expression node in the conclusion expression node set are summarized to generate a conclusion type set, which includes strongly supported conclusion types, conditionally supported conclusion types, and insufficient evidence conclusion types.

[0029] In this embodiment, the generation of the business school lab report includes: Based on the candidate inference path set and the path semantic representation set, an experimental analysis process text segment is generated. The experimental analysis process text segment is constructed by referencing the node sequence corresponding to a preset number of high-scoring candidate inference paths and the evidence summary vector in the path semantic representation. The evidence summary vector is used to describe the relationship between experimental variables, statistical significance results, and sample size information. Based on the conclusion support distribution results and the conclusion type set, a conclusion derivation explanation text segment is generated. The conclusion derivation explanation text segment is constructed by referencing the conclusion support distribution results of the corresponding conclusion expression node and the path identification information of the main supporting candidate reasoning paths. It is used to describe the reasoning basis and main sources of evidence for the formation of the conclusion. Based on the conclusion type set and conclusion support distribution results, a conclusion reliability description text segment is generated. The conclusion reliability description text segment is constructed by referencing the conclusion type, conflict degree and effective evidence coverage of the corresponding conclusion expression node, and is used to describe the reliability of the conclusion and the consistency of the evidence. In the process of generating the experimental analysis process text segment, the conclusion derivation explanation text segment, and the conclusion reliability description text segment, each sentence in each text segment is bound to the corresponding candidate reasoning path identifier and evidence summary identifier to generate traceable annotation results. The traceable annotation results and the corresponding text segments are written into the business experiment report to generate the business experiment report.

[0030] A deep learning-based automatic business lab report generation system, including: The experimental fact input structure generation module is used to acquire experimental data sets related to business experiments, perform structured processing on the experimental data sets, and generate experimental fact input structures. The experimental fact input structures include experimental condition information, experimental variable information, and raw statistical result information. The hierarchical semantic reasoning graph structure construction module is used to construct a hierarchical semantic reasoning graph structure based on the experimental fact input structure. In the hierarchical semantic reasoning graph structure, an experimental fact layer, an analysis and reasoning layer, and a conclusion expression layer are constructed. An experimental fact node set is constructed in the experimental fact layer, an analysis and reasoning node set is constructed in the analysis and reasoning layer, and a conclusion expression node set is constructed in the conclusion expression layer. Cross-layer semantic associations are established, which connect the experimental fact node set and the analysis and reasoning node set, and also connect the analysis and reasoning node set and the conclusion expression node set. The improved NBFNet inference model and candidate inference path set generation module is used to construct an improved NBFNet inference model based on a hierarchical semantic inference graph structure. The experimental facts input structure is input into the improved NBFNet inference model, and cross-layer path relaxation propagation processing is performed according to cross-layer semantic association to generate a candidate inference path set. Each candidate inference path in the candidate inference path set starts from the starting node in the experimental facts node set, passes through the analysis and inference node set, and reaches the conclusion expression node set. The path-level semantic memory structure and path semantic representation set generation module is used to construct a path-level semantic memory structure for each candidate inference path in the candidate inference path set during the cross-layer path relaxation propagation process of the improved NBFNet inference model. The module performs cumulative update processing on the path-level semantic memory structure based on the node sequence and edge sequence of the candidate inference path to generate a path semantic representation set. Each path semantic representation in the path semantic representation set corresponds to a candidate inference path. The conclusion support distribution generation module is used to input the path semantic representation set into the improved NBFNet inference model, perform multi-path competitive conclusion formation processing, obtain a subset of path semantic representations pointing to the conclusion expression node for each conclusion expression node in the conclusion expression node set, perform support strength evaluation processing and conflict relationship modeling processing based on the subset of path semantic representations, and generate conclusion support distribution results. The conclusion type set generation module is used to perform conclusion type determination processing on each conclusion expression node in the conclusion expression node set based on the conclusion support distribution results, and generate a conclusion type set. The conclusion type set includes strongly supported conclusion types, conditionally supported conclusion types, and insufficient evidence conclusion types. The conclusion type determination processing assigns one conclusion type from the conclusion type set to each conclusion expression node. The Business Experiment Report Generation Module is used to generate business experiment reports based on a set of candidate reasoning paths, a set of path semantic representations, and a set of conclusion types. The business experiment report includes an experiment analysis process text segment, a conclusion derivation explanation text segment, and a conclusion reliability description text segment. The module generates the experiment analysis process text segment based on the candidate reasoning path set and the path semantic representation set, referencing the node sequences corresponding to a preset number of high-scoring candidate reasoning paths and the evidence summary vectors in the path semantic representations. It generates the conclusion derivation explanation text segment based on the conclusion support distribution results and the conclusion type set, referencing the conclusion support distribution results of the corresponding conclusion expression nodes and the path identifier information of the main supporting candidate reasoning paths. It generates the conclusion reliability description text segment based on the conclusion type set and the conclusion support distribution results, referencing the conclusion type, conflict degree, and effective evidence coverage of the corresponding conclusion expression nodes. During the generation of the experiment analysis process text segment, the conclusion derivation explanation text segment, and the conclusion reliability description text segment, each sentence in each text segment is bound to a corresponding candidate reasoning path identifier and evidence summary identifier, generating traceable annotation results, which are then written into the business experiment report along with the corresponding text segment.

[0031] Example 1: To verify the feasibility of this invention in practice, it was applied to a teaching and research scenario combining course experiments and scientific research training in a business school. In this scenario, the research object was a large amount of experimental data generated from multiple business courses within the same semester. This experimental data originated from course experiments, classroom surveys, behavioral experiments, and statistical analysis assignments, and included experimental condition information, experimental variable information, and raw statistical results. In practical application, it was found that the generation of existing business school experiment reports heavily relies on manual compilation and writing. The logical connection between experimental facts, analysis processes, and conclusions is unstable; different researchers give significantly different conclusions on the same experimental results; and the credibility of conclusions in reports lacks quantitative evidence, making it difficult to support teaching evaluation and scientific research review. This problem is particularly prominent when multiple experiments are conducted in parallel and the data scale continues to grow.

[0032] In this application scenario, the experimental data set generated from course experiments is input into the method of this invention. The experimental data set undergoes structured processing to generate an experimental fact input structure. This structure clearly distinguishes between experimental condition information, experimental variable information, and raw statistical results, organizing data from different sources within a unified representation framework. Based on this experimental fact input structure, a hierarchical semantic reasoning graph structure is constructed. This structure simultaneously contains an experimental fact layer, an analytical reasoning layer, and a conclusion expression layer. Constrained connections are established between experimental facts and analytical reasoning, and between analytical reasoning and conclusion expression, through cross-layer semantic relationships, enabling the experimental data to possess reasonability at the structural level.

[0033] In actual operation, the experimental facts are input into the improved NBFNet inference model. Under the constraints of the hierarchical semantic inference graph structure, cross-layer path relaxation propagation is performed to obtain a set of candidate inference paths. Each candidate inference path starts from the experimental fact node, passes through the analysis and inference node, and reaches the conclusion expression node, fully reflecting the logical chain from experimental facts to conclusion. During path propagation, a path-level semantic memory structure is simultaneously constructed for each candidate inference path, and cumulative update processing is performed based on the node sequence and edge sequence of the candidate inference path to obtain a set of path semantic representations, so that each inference path corresponds to a quantifiable and comparable semantic representation.

[0034] Building upon this foundation, the path semantic representation set is input again into the improved NBFNet inference model to perform multi-path competitive conclusion formation processing. For the same conclusion expression node, the system can simultaneously consider multiple candidate inference paths, comprehensively model the path support strength and conflict relationships, and generate conclusion support distribution results. Subsequently, based on the conclusion support distribution results, conclusion type determination processing is performed on each conclusion expression node to generate a conclusion type set, classifying conclusions into strongly supported conclusion types, conditionally supported conclusion types, and insufficient evidence conclusion types. Finally, combining the candidate inference path set, the path semantic representation set, and the conclusion type set, a complete and logically clear business experiment report is automatically generated. The report content includes text segments of the experimental analysis process, text segments explaining the conclusion derivation, and text segments describing the reliability of the conclusion.

[0035] To quantitatively evaluate the practical effectiveness of this invention in this scenario, the same batch of experimental data was selected, and a comparative analysis was conducted using the method of this invention and the traditional, manually-led business experimental report writing method. The traditional method involves researchers with teaching experience manually writing reports based on experimental statistical results, without incorporating a hierarchical semantic reasoning graph structure and path-level semantic memory mechanism. The comparison metrics included conclusion consistency, conclusion verifiability, report generation time, and coverage of supporting evidence for the conclusions. The statistical results are shown below.

[0036] Table 1. Comparison of Business Experiment Report Generation Results

[0037] As shown in Table 1, regarding the consistency ratio of conclusions, the traditional method achieved a consistency of 0.72 for the same experimental conclusions when different researchers wrote their reports. However, the method of this invention, after introducing a hierarchical semantic reasoning graph structure and multi-path competitive conclusion formation processing, increased this ratio to 0.83. This change indicates that the present invention, through unified modeling of candidate reasoning paths, effectively reduces the influence of subjective human judgment on the conclusions, making the conclusion formation more stable.

[0038] Regarding the verifiability of conclusions, traditional methods, lacking explicit records of reasoning paths, often only allow for retrospective analysis of statistical results during the verification of experimental conclusions, resulting in a verifiability ratio of 0.68. The method of this invention accumulates and records the reasoning process through a path-level semantic memory structure, enabling each conclusion to be traced back to its corresponding candidate reasoning path and semantic representation. This increases the verifiability ratio to 0.81, demonstrating stronger auditability.

[0039] Regarding the time required to generate a report, traditional methods typically take an average of 48.5 minutes to complete a full experimental report. However, the method of this invention, after automatically generating the experimental analysis process text segment, the conclusion derivation explanation text segment, and the conclusion reliability description text segment, reduces the average time to 36.2 minutes. The reduction in time is within a reasonable range, demonstrating the advantages of automation without causing quality imbalance due to excessive compression.

[0040] Regarding the coverage of supporting evidence for the conclusion, traditional methods often rely on a small number of significant statistical results, resulting in a coverage rate of 0.65. The method of this invention comprehensively utilizes information from multiple candidate reasoning paths in the multi-path competitive conclusion formation process, increasing the coverage of supporting evidence for the conclusion to 0.79, which helps to improve the completeness of the conclusion's reliability expression.

[0041] Furthermore, the method of this invention is superior to the traditional method in both the identification ratio of conditionally supported conclusion types and the identification ratio of insufficient evidence conclusion types. This indicates that the present invention can more meticulously distinguish the conclusion situations under different support strengths and conflict states in the conclusion type determination process, and avoid directly classifying the results of insufficient evidence as definitive conclusions.

[0042] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for automatically generating business lab reports based on deep learning, characterized in that, Includes the following steps: Obtain the experimental data set, perform structured processing, and generate the experimental fact input structure; Based on the experimental fact input structure, a hierarchical semantic reasoning graph structure is constructed, including the experimental fact layer, the analysis and reasoning layer, and the conclusion expression layer, and cross-layer semantic associations are established. An improved NBFNet inference model is constructed based on a hierarchical semantic inference graph structure. The experimental facts are input into the model, and cross-layer path relaxation propagation processing is performed based on cross-layer semantic relationships to generate a set of candidate inference paths. For each candidate reasoning path in the candidate reasoning path set, construct a path-level semantic memory structure, perform cumulative update processing, and generate a path semantic representation set; The path semantic representation set is input into the improved NBFNet inference model, and multi-path competitive conclusion formation processing is performed to generate conclusion support distribution results. Based on the conclusion support distribution results, perform conclusion type determination processing on each conclusion expression node to generate a conclusion type set; Based on the candidate reasoning path set, path semantic representation set, and conclusion type set, a business lab report generation process is performed to generate a business lab report.

2. The method for automatically generating business lab reports based on deep learning according to claim 1, characterized in that, The generation of the experimental fact input structure includes: Obtain the raw experimental data related to business experiments, divide it into experimental condition data, experimental variable data and raw statistical result data, perform integrity verification processing, and obtain a valid set of experimental data. Perform data consistency processing on the valid experimental data set to generate a standardized experimental data set; Based on the standardized experimental dataset, condition encoding is performed on the experimental condition data to generate experimental condition information; variable mapping is performed on the experimental variable data to generate experimental variable information; and statistical result processing is performed on the raw statistical result data to generate raw statistical result information. The experimental conditions, experimental variables, and raw statistical results are organized in a structured manner to generate an experimental fact input structure.

3. The method for automatically generating business lab reports based on deep learning according to claim 1, characterized in that, The construction of the hierarchical semantic reasoning graph structure includes: Based on the experimental fact input structure, an experimental fact layer and a set of experimental fact nodes are constructed in the hierarchical semantic reasoning graph structure, an analytical reasoning layer and a set of analytical reasoning nodes are constructed, and a conclusion expression layer and a set of conclusion expression nodes are constructed. Establish cross-layer semantic associations from facts to reasoning based on the experimental fact node set and the analytical reasoning node set; A cross-layer semantic relationship between reasoning and conclusion is established based on the set of analytical reasoning nodes and the set of conclusion expression nodes; Invalid edge removal is performed on cross-layer connections that do not satisfy the cross-layer semantic association constraints from facts to reasoning and from reasoning to conclusion, thereby generating a hierarchical semantic reasoning graph structure.

4. The method for automatically generating business lab reports based on deep learning according to claim 1, characterized in that, The generation of the candidate reasoning path set includes: Based on the hierarchical semantic reasoning graph structure, an improved NBFNet reasoning model is constructed, and path order constraint rules are set to limit the propagation order of reasoning paths and the upper limit of the path length of reasoning paths. The experimental facts are input into the structure and then into the improved NBFNet inference model. During the cross-layer path relaxation propagation process, the edge weights of the cross-layer semantic associations are updated to generate updated edge weights. During the cross-layer path relaxation propagation process, a feasibility determination process is performed on each candidate inference path; After completing the cross-layer path relaxation propagation process, for each conclusion expression node, the inference path score pointing to the conclusion expression node is calculated based on the updated edge weight. The inference path scores are sorted and the Top-K inference paths are retained to generate a candidate inference path set.

5. The method for automatically generating business lab reports based on deep learning according to claim 1, characterized in that, The generation of the path semantic representation set includes: During the cross-layer path relaxation propagation process in the improved NBFNet inference model, a path-level semantic memory structure is constructed for each candidate inference path, including node contribution cache, edge constraint cache, and evidence trajectory cache. The path-level semantic memory structure is cumulatively updated based on the candidate reasoning paths; After completing the cumulative update processing of the path-level semantic memory structure, the node contribution cache, edge constraint cache, and evidence trajectory cache are concatenated to generate a path semantic representation set.

6. The method for automatically generating business lab reports based on deep learning according to claim 1, characterized in that, The conclusions support the generation of the distribution results in the following ways: Input the path semantic representation set into the improved NBFNet inference model, and for each conclusion expression node, obtain the path semantic representation subset pointing to that conclusion expression node; Based on a subset of path semantic representations, a support strength evaluation process is performed on each path semantic representation to calculate the single-path support strength value of the corresponding path. Based on the single path support strength value, conflict relationship modeling is performed to identify whether there are directional conflicts and conditional conflicts in the path semantic representation subset corresponding to the same conclusion expression node. When the proportion of directional conflicts exceeds the preset conflict threshold, the support ratio of the high support interval in the conclusion support distribution results of the corresponding conclusion expression node decreases, while the support ratio of the medium support interval and the low support interval increases. When a condition conflict is determined, condition cluster partitioning is performed to generate the conclusion support distribution results.

7. The method for automatically generating business lab reports based on deep learning according to claim 1, characterized in that, The generation of the conclusion type set includes: Based on the conclusion support distribution results, the conclusion support distribution index is calculated for each conclusion expression node, including total support strength, conflict degree and effective evidence coverage. Based on the conclusion support distribution index, a conclusion type determination process is performed on each conclusion expression node. A conclusion type is assigned to the corresponding conclusion expression node according to a preset threshold condition. The conclusion types include strongly supported conclusion types, conditionally supported conclusion types, and insufficient evidence conclusion types. The conclusion types are summarized to generate a conclusion type set.

8. The method for automatically generating business lab reports based on deep learning according to claim 1, characterized in that, The generation of the business school lab report includes: Based on the candidate reasoning path set and the path semantic representation set, an experimental analysis process text segment is generated, which is constructed by referencing the node sequence corresponding to a preset number of high-scoring candidate reasoning paths and the evidence summary vector in the path semantic representation; Based on the conclusion support distribution results and conclusion type set, a conclusion derivation explanation text segment is generated. It is constructed by referencing the conclusion support distribution results of the corresponding conclusion expression node and the path identification information of the main supporting candidate reasoning paths. Based on the conclusion type set and conclusion support distribution results, a conclusion reliability description text segment is generated, which is constructed by referencing the conclusion type, conflict degree and effective evidence coverage of the corresponding conclusion expression node; In the process of generating the experimental analysis process text segment, the conclusion derivation explanation text segment, and the conclusion reliability description text segment, each sentence in each text segment is bound to the corresponding candidate reasoning path identifier and evidence summary identifier to generate traceable annotation results, which are written into the business experiment report together with the corresponding text segment to generate the business experiment report.

9. A deep learning-based automatic business lab report generation system, comprising the deep learning-based automatic business lab report generation method according to any one of claims 1 to 8, characterized in that, include: The experimental fact input structure generation module is used to acquire experimental datasets, merge them, perform structured processing, and generate experimental fact input structures. The hierarchical semantic reasoning graph structure construction module is used to construct a hierarchical semantic reasoning graph structure based on the experimental fact input structure. The hierarchical semantic reasoning graph structure includes an experimental fact layer, an analysis and reasoning layer, and a conclusion expression layer, and establishes cross-layer semantic relationships. An improved NBFNet inference model and candidate inference path set generation module is used to construct an improved NBFNet inference model based on a hierarchical semantic inference graph structure. The model takes experimental facts as input, performs cross-layer path relaxation propagation processing based on cross-layer semantic relationships, and generates a candidate inference path set. The path-level semantic memory structure and path semantic representation set generation module is used to construct a path-level semantic memory structure for each candidate reasoning path in the candidate reasoning path set, perform cumulative update processing, and generate a path semantic representation set. The conclusion support distribution generation module is used to input the path semantic representation set into the improved NBFNet inference model, perform multi-path competitive conclusion formation processing, and generate conclusion support distribution results. The conclusion type set generation module is used to perform conclusion type determination processing on each conclusion expression node in the conclusion expression node set based on the conclusion support distribution results, and generate a conclusion type set. The Business Experiment Report Generation Module is used to generate business experiment reports based on a set of candidate reasoning paths, a set of path semantic representations, and a set of conclusion types.

Images