Document quality evaluation method, device, equipment, storage medium and program product

Abstract

The present application relates to the technical field of artificial intelligence, and provides a manuscript quality evaluation method, device, equipment, storage medium and program product, the method comprises: carrying out semantic segmentation on a source document to obtain a multi-ary semantic tree, each node in the multi-ary semantic tree contains the semantic vector of the corresponding text segment in the source document; the page sequence of the presentation corresponding to the source document is analyzed in multiple modes to obtain a page vector sequence; based on the similarity between each page vector in the page vector sequence and the semantic vector of each node in the multi-ary semantic tree, a page segment association matrix is constructed, realizing accurate semantic alignment from the document paragraph to the PPT page; based on the page segment association matrix, consistency quality evaluation is carried out, which can not only effectively identify whether the content is accurate, but also locate the specific deviation source, thereby significantly improving the accuracy and reliability of PPT quality evaluation, and providing strong technical support for the consistency check of PPT and source document in the office automation scene.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus, device, storage medium, and program product for evaluating document quality. Background Technology

[0002] With the development of artificial intelligence technology, automatically generated PPTs have greatly improved office efficiency. However, in scenarios such as finance and education where the accuracy of content is extremely important, automatically generated PPTs have problems such as semantic drift and omission of core data.

[0003] Currently, PPT evaluation methods mainly include similarity-based evaluation, which calculates the overall similarity between the PPT and the document, but this method has too coarse an evaluation granularity; rule-based hard checks, which can only check the compliance of attributes such as font size and word count limits; and manual sampling, which relies on manual verification of the correspondence between the document and the slides page by page, and is inefficient. Summary of the Invention

[0004] This invention provides a method, apparatus, device, storage medium, and program product for document quality assessment, which solves the problem in the prior art that the presentation document cannot be deeply verified for content consistency with the source document, resulting in poor accuracy and interpretability of the assessment results.

[0005] This invention provides a method for evaluating manuscript quality, comprising:

[0006] Obtain the presentation to be evaluated, as well as the source document corresponding to the presentation;

[0007] The source document is semantically segmented to obtain a multi-branch semantic tree, and each node in the multi-branch semantic tree contains the semantic vector of the corresponding text segment in the source document;

[0008] Multimodal parsing is performed on the page sequence of the presentation to obtain a page vector sequence;

[0009] Based on the similarity between each page vector in the page vector sequence and the semantic vector of each node in the multi-branch semantic tree, a page segment association matrix is ​​constructed;

[0010] Based on the page segment association matrix, a consistency quality assessment is performed on the presentation and the source document to obtain the quality assessment results.

[0011] According to a manuscript quality assessment method provided by the present invention, each node in the multi-branch semantic tree further includes a node weight, wherein the node weight represents the content importance of the corresponding text segment;

[0012] The consistency quality assessment of the presentation and the source document based on the page segment association matrix yields the quality assessment results, including:

[0013] Based on the page segment association matrix, each page in the page sequence of the presentation is associated with and located with each text segment in the source document to obtain the content location result;

[0014] Based on the node weights corresponding to each node in the multi-branch semantic tree, core nodes are determined, and based on the similarity of the core nodes in the page segment association matrix, important content coverage detection is performed on the presentation and the source document to obtain key coverage results.

[0015] Based on the page segment association matrix, conflict detection is performed on the text content of each page of the presentation and each text segment in the source document to obtain the conflict detection result.

[0016] Based on the content location results, the key coverage results, and the conflict detection results, the quality assessment results are determined.

[0017] According to a manuscript quality assessment method provided by the present invention, determining the quality assessment result based on the content positioning result, the key coverage result, and the conflict detection result includes:

[0018] Based on the content positioning results, the key coverage results, and the conflict detection results, the content evaluation results are determined.

[0019] Extract the semantic evolution path of the page vector sequence of the presentation and perform a depth-first search on the multi-branch semantic tree to obtain the narrative logic path of the source document;

[0020] Based on the topological similarity between the semantic evolution path and the narrative logic path, the logical consistency between the presentation and the source document is evaluated to obtain the logical evaluation result.

[0021] The quality assessment result is determined based on the content assessment result and the logical assessment result.

[0022] According to a document quality assessment method provided by the present invention, the step of evaluating the logical consistency between the presentation document and the source document based on the topological similarity between the semantic evolution path and the narrative logic path to obtain a logical assessment result includes:

[0023] Based on the topological similarity between the semantic evolution path and the narrative logic path, the logical consistency between the presentation and the source document is evaluated to obtain the consistency evaluation result.

[0024] Based on the semantic transfer distance between the page vectors of adjacent pages in the page vector sequence, a logical coherence analysis is performed on the presentation and the source document to obtain the coherence analysis results.

[0025] Based on the consistency assessment results and the coherence analysis results, the logical assessment results are determined.

[0026] According to a document quality assessment method provided by the present invention, the step of performing logical coherence analysis on the presentation document and the source document based on the semantic transfer distance between page vectors of adjacent pages in the page vector sequence to obtain coherence analysis results includes:

[0027] If the semantic transfer distance exceeds the fault threshold, then the logical breakpoint between the corresponding adjacent pages is determined.

[0028] Using the preceding and following pages of the logical breakpoint as search criteria, a local search is performed in the multi-branch semantic tree;

[0029] If a node with semantic association with both the preceding and following pages is found in the multi-branch semantic tree, the coherence analysis result is determined to be a logical jump in the presentation; otherwise, the coherence analysis result is determined to be a logical missing element in the source document.

[0030] According to a document quality assessment method provided by the present invention, the step of performing multimodal parsing on the page sequence of the presentation document to obtain a page vector sequence includes:

[0031] The page sequence of the presentation is structured and parsed to obtain the text content of each page in the page sequence, and the text content is mapped into a text semantic vector;

[0032] Extract the visual information of each page and project the visual information onto the vector space where the text semantic vector is located to obtain the visual semantic vector of each page;

[0033] By fusing the text semantic vector and visual semantic vector of each page, a page vector is obtained, and the page vector sequence is determined based on the page vector of each page.

[0034] According to a document quality assessment method provided by the present invention, the step of semantically segmenting the source document to obtain a multi-branch semantic tree includes:

[0035] The source document is segmented into paragraphs and parsed hierarchically to obtain a multi-branch text tree. The nodes in the multi-branch text tree correspond to text segments in the source document, and there is a hierarchical relationship between the nodes.

[0036] Encode each node in the multi-branch text tree to obtain the semantic vector corresponding to each node;

[0037] Based on the multi-branch text tree and the semantic vectors corresponding to each node, a multi-branch semantic tree is constructed.

[0038] According to a document quality assessment method provided by the present invention, the method further includes, based on the page segment association matrix, performing a consistency quality assessment on the presentation document and the source document to obtain a quality assessment result, and then further comprising:

[0039] Construct a reward function that includes accuracy, coverage, logicality, and redundancy;

[0040] Based on the quality assessment results and the reward function, the quality score of the presentation is determined;

[0041] Obtain feedback scores for the quality assessment results;

[0042] Based on the quality score and the feedback score, the reward function is updated so that the quality score determined based on the updated reward function and the corresponding feedback score meet the consistency condition.

[0043] The present invention also provides a manuscript quality assessment device, comprising:

[0044] The acquisition unit is used to acquire the presentation to be evaluated and the source document corresponding to the presentation;

[0045] The segmentation unit is used to perform semantic segmentation on the source document to obtain a multi-branch semantic tree, wherein each node in the multi-branch semantic tree contains the semantic vector of the corresponding text fragment in the source document;

[0046] The parsing unit is used to perform multimodal parsing on the page sequence of the presentation to obtain a page vector sequence;

[0047] The association unit is used to construct a page segment association matrix based on the similarity between each page vector in the page vector sequence and the semantic vector of each node in the multi-branch semantic tree;

[0048] The evaluation unit is used to perform a consistency quality assessment on the presentation and the source document based on the page segment association matrix, and obtain the quality assessment result.

[0049] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the document quality assessment method as described above.

[0050] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the manuscript quality assessment method as described above.

[0051] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the manuscript quality assessment method as described above.

[0052] The document quality assessment method, apparatus, device, storage medium, and program products provided by this invention achieve a deep understanding of the logical structure of the source document and the multimodal content of the PPT by structuring the source document into a multi-branch semantic tree and parsing the PPT into a sequence of page vectors. Furthermore, based on the multi-branch semantic tree and the page vector sequence, a page segment association matrix is ​​constructed at a fine-grained semantic level, achieving precise semantic alignment from document paragraphs to PPT pages. Evaluation based on this page segment association matrix can not only effectively identify whether the content is accurate, but also locate the specific source of deviation, significantly improving the accuracy and reliability of PPT quality assessment and providing strong technical support for consistency verification between PPT and source documents in office automation scenarios. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0054] Figure 1 This is a flowchart illustrating the manuscript quality assessment method provided by the present invention;

[0055] Figure 2 This is the overall flowchart of the manuscript quality assessment method provided by the present invention;

[0056] Figure 3 This is a schematic diagram of the manuscript quality assessment device provided by the present invention;

[0057] Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0059] With the development of AIGC (Artificial Intelligence Generated Content) technology, automatically generating PowerPoint presentations has become a reality, greatly improving office efficiency. However, in scenarios such as finance and education, where content accuracy is paramount, AI-generated PowerPoint presentations often suffer from serious problems such as semantic drift and omission of core data. Currently, most PowerPoint generation solutions on the market focus on the aesthetics of layout and formatting, lacking an assessment of the accuracy of the generated content. Therefore, ensuring the accuracy of PowerPoint content and avoiding content distortion remains a bottleneck in the field of office automation.

[0060] Currently, PPT evaluation methods mainly include similarity-based evaluation, such as using language models like BERT (Bidirectional Encoder Representations from Transformers) to calculate the similarity between PPT text and the document, and giving an overall similarity score; rule-based hard checks, which mainly check whether attributes such as font size and word count limits are compliant; and manual sampling, which relies entirely on manual verification of the correspondence between the document and the slides page by page.

[0061] However, the above-mentioned solutions all have obvious limitations in practical applications. Specifically, on the one hand, simple overall similarity calculation cannot delve into the internal structure of the document, making it difficult to identify whether information in the document has been effectively extracted by the PPT, resulting in overly coarse evaluation granularity. On the other hand, existing solutions treat the document as a flat text stream, failing to determine the correspondence between pages in the PPT and chapters, paragraphs, etc., in the document. This lack of mapping relationships prevents the system from performing fine-grained illusion detection; that is, when the PPT contains content that does not exist in the document, the system cannot disprove it through precise location, leading to poor accuracy and interpretability of the evaluation results.

[0062] In response, this invention provides a document quality assessment method, which aims to solve the problems of coarse evaluation granularity and inability to establish precise mapping relationships between PPT pages and document paragraphs in the prior art by converting the source document into a multi-branch semantic tree and combining it with multimodal parsing technology, thereby achieving in-depth consistency quality verification between the presentation and the source document from macro structure to micro knowledge points.

[0063] Figure 1 This is a flowchart illustrating the manuscript quality assessment method provided by the present invention, which is applied to a manuscript quality assessment system (hereinafter referred to as the system). Figure 1 As shown, the method includes:

[0064] Step 110: Obtain the presentation to be evaluated, and the source document corresponding to the presentation;

[0065] Step 120: Semantically segment the source document to obtain a multi-branch semantic tree. Each node in the multi-branch semantic tree contains the semantic vector of the corresponding text segment in the source document.

[0066] Step 130: Perform multimodal parsing on the page sequence of the presentation to obtain the page vector sequence;

[0067] Step 140: Construct a page segment association matrix based on the similarity between each page vector in the page vector sequence and the semantic vector of each node in the multi-branch semantic tree;

[0068] Step 150: Based on the page segment association matrix, perform a consistency quality assessment on the presentation and the source document to obtain the quality assessment results.

[0069] Specifically, in office automation or intelligent content generation scenarios, users often need to verify whether the generated PPT is faithful to the original text. Therefore, when evaluating PPT quality, the system first needs to obtain the PPT to be evaluated and its corresponding source document. Here, the PPT to be evaluated can be a slide file containing highly condensed information and visual layout elements, such as a .ppt or .pptx file. It can be AI-generated or manually created; this embodiment of the invention does not specifically limit this. The source document is the basic file used as the basis for PPT generation or comparison benchmark, such as a .doc, .pdf, or .txt file. This file usually contains detailed text descriptions, data support, and a complete logical deduction framework. After obtaining the source document, the system usually preprocesses it, such as format conversion and decryption.

[0070] Furthermore, to understand the logical structure of the source document, in this embodiment of the invention, after obtaining the source document, it does not simply perform sentence segmentation, but rather deep semantic segmentation. Specifically, the system utilizes natural language processing techniques, such as a pre-trained document segmentation model with a small number of parameters based on semantic paragraph segmentation theory, to identify the logical hierarchy within the source document and accurately divide it into multiple text segments, such as chapters, sections, subsections, and specific paragraph text blocks.

[0071] Based on this granular division, the system can reconstruct a flat source document into a multi-branch semantic tree. This is a data structure that can represent the logical structure of the source document. The root node in the tree represents the document topic, the branch nodes represent chapters at various levels, and the leaf nodes correspond to specific paragraphs. This structure fully preserves the hierarchical relationships between text content, such as which chapter, section, or subsection a paragraph belongs to, as well as the sequential relationships, such as the logical order between paragraphs.

[0072] Building upon this foundation, to make these nodes computable, the system utilizes conventional feature extraction models and large language models to encode each node in the multi-branch semantic tree, i.e., the corresponding text fragment, transforming it into a high-dimensional dense vector, i.e., a semantic vector. Through this process, the originally unstructured source document is transformed into a knowledge base with rigorous hierarchical relationships, composed of a large number of semantic vectors, thus providing accurate benchmark coordinates for subsequent comparisons.

[0073] Unlike the source document, a presentation contains rich visual information. To accurately capture this information, the system performs multimodal parsing on the PPT pages arranged in playback order to obtain a sequence of page vectors. Specifically, for each page in the PPT, the system uses both XML (Extensible Markup Language) structure parsing and OCR (Optical Character Recognition) visual recognition technologies to extract text information such as titles, body text, and notes. Simultaneously, it utilizes image processing techniques, such as the Vision-Language Model (VLM), to extract visual information such as page screenshots, layout structure, and font size. After parsing, the system merges all the text and visual information of the page, mapping it into a unified page vector. This vector contains both the literal meaning of the page and the implied visual emphasis conveyed by the layout. As each page is parsed sequentially, the system ultimately generates page vectors corresponding one-to-one with the PPT page numbers. These page vectors, arranged in the playback order, constitute the page vector sequence.

[0074] After obtaining the page vector sequence of the PPT and the corresponding multi-branch semantic tree of the source document, the system needs to establish a mapping relationship between the PPT and the source document. Specifically, the system traverses each page vector in the page vector sequence, compares it with the semantic vector of each node in the multi-branch semantic tree, and calculates the similarity between them, such as cosine similarity, Euclidean distance, etc., to measure the closeness of a PPT page to a text segment in the source document in terms of text and semantic theme. Through this calculation, the system constructs a two-dimensional page segment association matrix. In this matrix, rows can represent PPT page indices, columns can represent source document node indices, and each element value in the matrix reflects the matching score (similarity) between the corresponding page (page vector) and the corresponding text segment (semantic vector). This matrix intuitively shows the content corresponding to each page of the PPT in the source document, as well as the strength of this correspondence (similarity level).

[0075] Following this, the system uses the page segment association matrix as the core basis to perform a consistency quality assessment on the PPT and the source document, obtaining the quality assessment result. This process aims to determine whether the PPT content faithfully reproduces the knowledge structure of the source document. That is, the system will deeply analyze the numerical distribution in the page segment association matrix. If all similarities in a certain row of the page segment association matrix (representing a certain PPT page) are lower than a preset threshold, it indicates that the content of that page cannot find a basis for high similarity in the source document, and the system judges that there may be illusions or fabricated content. Conversely, if the similarity of columns corresponding to key information in the source document or core nodes in the multi-branch semantic tree in the page segment association matrix is ​​generally low, it suggests that the key information in the source document may not be effectively reflected in the PPT, that is, there is a lack of core knowledge points. Here, the quality assessment result can take various forms, including quantitative scores, specific diagnostic reports, such as indicating which specific page in the PPT has a problem and what the problem is, or a pass / fail label. This embodiment of the invention does not specifically limit this.

[0076] The document quality assessment method provided by this invention achieves a deep understanding of the logical structure of the source document and the multimodal content of the PPT by structuring the source document into a multi-branch semantic tree and parsing the PPT into a sequence of page vectors. Furthermore, based on the multi-branch semantic tree and the page vector sequence, a page segment association matrix is ​​constructed at a fine-grained semantic level, achieving precise semantic alignment from document paragraphs to PPT pages. Evaluation based on this page segment association matrix can not only effectively identify whether the content is accurate, but also locate the specific source of deviation, thereby significantly improving the accuracy and reliability of PPT quality assessment and providing strong technical support for consistency verification between PPT and source documents in office automation scenarios.

[0077] Based on the above embodiments, each node in the multi-branch semantic tree also contains a node weight, which represents the content importance of the corresponding text segment; step 150 includes:

[0078] Based on the page segment association matrix, each page in the page sequence of the presentation is associated with and located with each text segment in the source document to obtain the content location result;

[0079] Based on the node weights corresponding to each node in the multi-branch semantic tree, the core nodes are determined, and based on the similarity of the core nodes in the page segment association matrix, important content coverage detection is performed on the presentation and the source document to obtain key coverage results.

[0080] Based on the page segment association matrix, conflict detection is performed on the text content of each page of the presentation and the text segments in the source document to obtain the conflict detection results.

[0081] The quality assessment results are determined based on the content positioning results, key coverage results, and conflict detection results.

[0082] Specifically, in this embodiment of the invention, each node in the multi-branch semantic tree not only contains a semantic vector representing the content, but also a node weight representing the importance of the content, that is, the importance of the corresponding text segment within the entire source document. That is, when constructing the multi-branch semantic tree, the system assigns an importance score to each text segment in the source document. For example, for first-level headings or paragraphs containing core viewpoints or core data points, the system assigns a higher node weight, such as 0.9; while for secondary paragraphs such as notes or transitional statements, a lower node weight, such as 0.1, is assigned. Thus, the multi-branch semantic tree is not merely an index structure of content, but also a knowledge value graph marked with "priority" indicators.

[0083] Based on this, the process of assessing the consistency quality of the presentation and the source document according to the page segment association matrix can specifically include the following steps:

[0084] First, the system can use a page segment association matrix to associate and locate each page in the presentation's page sequence with each text segment in the source document, thereby determining the corresponding text segment in the source document for each page in the PPT and obtaining the content location result.

[0085] Specifically, the system scans the page segment association matrix row by row. For each row representing the i-th page of the PPT, the system searches for one or more column indices with the highest similarity. The node corresponding to this column index is the text segment that the page of the PPT matches in the source document. For example, if the system finds that the similarity between page 3 of the PPT and a paragraph in Chapter 2, Section 1 of the source document is as high as 0.85, while the similarity with other text segments is all below 0.3, then it determines that page 3 of the PPT is a display of the paragraph in Chapter 2, Section 1 of the source document. If it finds that all similarities in a row of a certain page of the PPT are below a preset threshold, such as 0.3, it indicates that no high similarity evidence can be found for that page of the PPT in the source document. The resulting set containing each page of the PPT and its corresponding paragraph index in the source document is the content location result. This result helps users quickly trace the source of the PPT content and verify its authenticity.

[0086] Meanwhile, the system can identify core nodes from all nodes based on the node weights of each node in the multi-branch semantic tree, and use the similarity of the core nodes in the page segment association matrix to perform important content coverage detection on the presentation and source document, so as to detect whether the PPT covers important content, core knowledge points, etc., and thus obtain the key coverage results.

[0087] Specifically, the system first filters out core nodes from all nodes in the multi-branch semantic tree based on a preset weight threshold, such as 0.8. These nodes are those with a weight greater than the threshold. These nodes represent key information in the source document, such as core viewpoints, core data points, and important conclusions. Next, the system examines the columns corresponding to these core nodes in the page segment association matrix. That is, for each core node, the system checks the maximum similarity of its column. If the similarity exceeds a set value, such as 0.7, it means the core node is reflected on a specific page in the PPT. Conversely, if the similarity is below the set value, it means the core node has not found a corresponding high-similarity page in the entire PPT, indicating that the core node is not effectively reflected in the PPT, thus confirming the omission of core knowledge points. The system then summarizes the coverage of all core nodes in the PPT to obtain the key coverage results. This result helps users quickly determine the key coverage of the PPT, avoiding problems such as inappropriate detail or unclear focus in AI-generated PPTs.

[0088] In addition, the system can also perform conflict detection on the text content of each page in the presentation and the text fragments in the source document based on the page segment association matrix. This detects whether the content or key information of the two conflict, or can be understood as verifying whether the facts are consistent, thus obtaining a conflict detection result (or factual consistency verification result). Specifically, this could mean that although some PPT pages and document paragraphs are highly similar, such as both discussing revenue growth, the specific values ​​may be incorrect. Therefore, the system can perform fine-grained entity or numerical comparisons between the text content in the PPT and the text fragments in the source document. For example, if the page vector of a certain PPT page is highly similar to the semantic vector of the corresponding text fragment, but there are significant differences in key entities (such as names of people and places) and values ​​(such as growth rate, sales revenue, etc.) between the two (the corresponding text content and text fragment), it can be determined that the facts are inconsistent, i.e., there is a hard conflict. This detection record for specific factual errors is the conflict detection result.

[0089] Finally, the system integrates the results from the above three dimensions to obtain the final quality assessment result. This result can be a comprehensive score, or a diagnostic report containing a detailed list of issues. For example, the report may include "The source of the PPT on page 5 is unclear (determined based on content positioning results analysis), the core knowledge point coverage rate of the entire text is only 60%, the chapter on market risk is missing (determined based on key coverage results analysis), and the sales figures shown on page 8 are inconsistent with the original text (determined based on conflict detection results analysis)." It can also be a pass / fail label, and this embodiment of the invention does not specifically limit this.

[0090] In this embodiment of the invention, the introduction of node weights and the implementation of a multi-dimensional evaluation strategy significantly improve the depth and practicality of document quality assessment. Specifically, positive content positioning solves the problem of verifying illusions, ensuring that each PPT slide has a traceable source. Simultaneously, the key coverage detection using node weights intelligently identifies and warns of missing key information, ensuring that the PPT not only has correct content but also highlights key points. Furthermore, conflict detection can capture hidden errors such as semantically similar but numerically contradictory statements, thereby greatly improving the rigor and security of PPT content in financial, legal, and other scenarios.

[0091] Based on the above embodiments, the quality assessment results are determined based on the content positioning results, key coverage results, and conflict detection results, including:

[0092] Based on the content positioning results, key coverage results, and conflict detection results, the content evaluation results are determined.

[0093] Extract the semantic evolution path of the page vector sequence of the presentation and perform a depth-first search on the multi-branch semantic tree to obtain the narrative logic path of the source document;

[0094] Based on the topological similarity between the semantic evolution path and the narrative logic path, the logical consistency between the presentation and the source document is evaluated, and the logical evaluation results are obtained.

[0095] The quality assessment results are determined based on the content assessment results and the logical assessment results.

[0096] Specifically, in order to ensure that the generated PPT can faithfully reproduce the narrative logic of the source document, this embodiment of the invention introduces a dynamic evaluation mechanism for the narrative logic.

[0097] Accordingly, the process of determining the quality assessment results based on the content positioning results, key coverage results, and conflict detection results may specifically include:

[0098] First, the system can aggregate the above content location results, key coverage results, and conflict detection results. For example, it can calculate the content accuracy score corresponding to the content location results and the core information recall rate corresponding to the key coverage results, and organize the specific error items in the conflict detection results into a list. These data sets reflecting the static content quality are collectively referred to as content evaluation results, which represent the quality of the PPT in terms of points and surfaces, but have not yet involved the quality of lines (logic flow).

[0099] Simultaneously, the system can extract the semantic evolution path of the page vector sequence of the presentation and perform a depth-first search on the multi-branch semantic tree to obtain the narrative logic path of the source document. This process aims to abstract two different modalities of documents into logical lines that can be compared with each other. Specifically, for PowerPoint presentations, which inherently have a linear playback order, the system can directly connect the page vectors in the page vector sequence according to the page number order to form a semantic evolution path that reflects the presentation order of the PowerPoint presentation. This path represents the order in which the PowerPoint presentation conveys information to the audience. Meanwhile, for the source document, its logical structure is implicit in the multi-branch semantic tree. In order to transform the tree structure into a linear logical flow, the system can perform a depth-first search (DFS) on the multi-branch semantic tree to search the branches of the tree as deeply as possible. It should be noted that, in the context of the document, the DFS search precisely conforms to the natural order in which humans read documents (i.e., first read the title of the first chapter, then the first section of the first chapter, then the paragraphs under the first section, and after reading the first section, return to read the second section of the first chapter, and so on). The sequence of nodes generated by DFS traversal constitutes the narrative logic path representing the standard logical flow of the source document.

[0100] Next, the system maps the semantic evolution path of the PPT to the narrative logic path of the source document. That is, it observes whether the jump order of the PPT pages follows the DFS traversal order of the document nodes. This comparison process is called topological similarity analysis.

[0101] For example, if the source document's logic is to raise problem A - analyze the cause B - provide solution C, the corresponding narrative logic path is ABC; if the PPT's page order is also to first present the pages of content type A, then type B, and finally type C, then the topological structures of the two paths are highly similar, and the logic is considered to be coherent; however, if the PPT's page order is to first present solution C and then go back to present background A, although the content is correct, the path is reversed or out of order, resulting in a decrease in topological similarity.

[0102] Based on this path matching degree, such as calculating the length of the longest common subsequence and edit distance, the system generates a logical evaluation result. This result can identify whether the PPT has deep structural problems such as logical inversion, narrative jumps, and causal inversion.

[0103] Finally, the system integrates or comprehensively judges the static content evaluation results with the dynamic logical evaluation results to arrive at the final quality evaluation result. This result can not only tell users where there are errors in the PPT, but also point out where the presentation is disjointed or logically flawed.

[0104] In this embodiment of the invention, by extracting semantic evolution paths and narrative logic paths, the linear structure of the PPT and the tree structure of the source document are unified and compared along the path dimension. Furthermore, by using topological similarity for evaluation, it can not only check the correctness of words, but also check the logical coherence like a human. It can effectively identify advanced errors where each PPT page looks correct on its own, but the logic is chaotic when viewed together, thus significantly improving the intelligence level of the evaluation. It is suitable for scenarios such as consulting reports and academic papers converted into PPTs that have extremely high requirements for logical rigor.

[0105] Based on the above embodiments, and based on the topological similarity between the semantic evolution path and the narrative logic path, a logical consistency assessment is performed on the presentation and the source document to obtain the logical assessment results, including:

[0106] Based on the topological similarity between the semantic evolution path and the narrative logic path, the logical consistency between the presentation and the source document is evaluated, and the consistency evaluation results are obtained.

[0107] Based on the semantic transfer distance between page vectors of adjacent pages in the page vector sequence, logical coherence analysis is performed on the presentation and the source document to obtain coherence analysis results.

[0108] Based on the consistency assessment results and the coherence analysis results, the logical assessment results are determined.

[0109] Specifically, in order to make the logical evaluation more comprehensive and three-dimensional, this embodiment of the invention not only focuses on whether the PPT is in reverse or out of order (macro consistency), but also on whether the connection between PPT pages is abrupt or discontinuous (micro continuity).

[0110] Accordingly, the process of assessing the logical consistency between the presentation and the source document based on the topological similarity between the semantic evolution path and the narrative logic path specifically includes:

[0111] First, the system can compare the semantic evolution path of the PPT with the narrative logic path of the source document to assess the consistency of the narrative logic between the PPT and the source document. This logical consistency assessment primarily determines whether the overall narrative framework of the PPT deviates from the baseline of the source document. For example, if the source document has a general-specific-general structure, and the PPT becomes a specific-specific-specific structure, losing the summary section, or if the PPT treats two parallel sections in the source document as causal, it will lead to a decrease in topological similarity. The consistency assessment result output by the system mainly reflects whether the PPT faithfully preserves the skeletal structure of the source document.

[0112] Meanwhile, the system can use the semantic transfer distance between the page vectors of adjacent pages in the page vector sequence to perform logical coherence analysis on the presentation and the source document, so as to detect whether the text content of adjacent pages in the PPT is coherent, and thus obtain the coherence analysis results.

[0113] Specifically, in this context, PowerPoint, as a presentation medium, typically exhibits a certain degree of semantic continuity between adjacent pages. To quantify this continuity, the system calculates the semantic transfer distance between the page vector corresponding to the i-th page and the page vector corresponding to the (i+1)-th page in the PowerPoint presentation. This distance can be the Euclidean distance between the two vectors, or it can be an indicator such as the reciprocal of the cosine similarity.

[0114] Furthermore, if the semantic transition distance between two adjacent pages is small, it indicates that the content of the two pages is closely connected and the transition is natural. However, if the semantic transition distance suddenly increases, such as exceeding a certain normal range, it indicates that the content of the two pages has undergone a drastic semantic jump. This may mean that there is a logical break in the PPT at this point, for example, jumping abruptly from market analysis to code implementation without any transition pages, or that the PPT has omitted necessary connecting content.

[0115] The system performs statistical analysis on the semantic transition distance between all adjacent pages to identify abnormal abrupt transitions, thus obtaining coherence analysis results. These results can pinpoint where parts of the PowerPoint presentation are presented abruptly, causing confusion for the audience.

[0116] Finally, the system integrates the consistency assessment results with the coherence analysis results to obtain the logical assessment results. For example, if the PPT structure perfectly matches the source document (good consistency), but the content spans between each page is very large and lacks necessary transitions (poor coherence), the final logical assessment result will still indicate a quality risk.

[0117] In this embodiment of the invention, the semantic transfer distance, a micro-level indicator, is introduced to compensate for the shortcomings of simply relying on macro-level topological structure comparison. This enables a three-dimensional evaluation combining macro and micro perspectives, checking both whether the PPT is out of format and whether it is lagging. Furthermore, in office scenarios, AI-generated PPTs often show situations where the previous page is still discussing the background and the next page suddenly presents the conclusion. Such micro-level logical jumps are difficult to detect through macro-level structures, but by calculating the semantic transfer distance between adjacent pages, this unnatural transition can be keenly captured, prompting users to add transition pages or adjust content, thereby significantly improving the readability and presentation smoothness of the PPT.

[0118] Based on the above embodiments, logical coherence analysis is performed on the presentation and source document based on the semantic transfer distance between page vectors of adjacent pages in the page vector sequence, resulting in coherence analysis results, including:

[0119] If the semantic transfer distance exceeds the fault threshold, then the logical breakpoint between the corresponding adjacent pages is determined.

[0120] Using the preceding and following pages of logical breakpoints as search criteria, local searches are performed in the multi-branch semantic tree.

[0121] If a node with semantic association with both the preceding and following pages is found in the multi-branch semantic tree, the coherence analysis result is determined to be a logical jump in the presentation; otherwise, the coherence analysis result is determined to be a logical missing element in the source document.

[0122] Specifically, in practical applications, when a large semantic gap is detected between adjacent pages in a PowerPoint presentation, it does not necessarily mean that the PowerPoint presentation was generated incorrectly; it could also be that the source document was written that way. To avoid false alarms, this embodiment of the invention introduces an attribution mechanism based on local re-retrieval to delve deeper into the causes of logical discontinuities. Figure 2 This is a flowchart of the overall process for the manuscript quality assessment method provided by the present invention, as follows: Figure 2 As shown, the specific process is as follows:

[0123] First, the system can identify semantic breakpoints based on a set discontinuity threshold. Specifically, when the semantic transition distance between the page vector corresponding to the i-th page and the page vector corresponding to the (i+1)-th page in the PPT exceeds the discontinuity threshold, it indicates a significant semantic gap between the two pages, and the system immediately marks a logical breakpoint between them. For example, if page i discusses global warming and page i+1 directly presents a company's first-quarter financial report, this abrupt jump will be detected by the system. The discontinuity threshold can be dynamically calculated based on statistical distributions or a preset empirical value.

[0124] Upon detecting a breakpoint, the system does not immediately report an error. Instead, it returns to the source document to search for clues. Specifically, the system uses the preceding page (e.g., page i) and the following page (e.g., page i+1) of the logical breakpoint as search criteria. More precisely, it uses the page vectors corresponding to the preceding and following pages in the page vector sequence as search criteria to perform a local search in the multi-branch semantic tree. The goal of this search is not to find the text fragments corresponding to these two pages, but rather to find the missing intermediate links. The system attempts to find in the multi-branch semantic tree whether one or more nodes exist that semantically both inherit the content of the preceding page and initiate the topic of the following page.

[0125] Based on the search results, the system will make an attribution determination. That is, if the system finds a node (or a group of nodes) in the multi-branch semantic tree that has a high semantic correlation with both the preceding and following pages, this proves that the source document actually contains transitional content, which was omitted by the PowerPoint presentation during generation. Therefore, the system determines that the PowerPoint presentation has a logical jump (i.e., the PowerPoint presentation omitted transitional information). In this case, the system can suggest that the user add the omitted transitional page.

[0126] Correspondingly, if the system traverses the entire semantic tree and cannot find any intermediate nodes that effectively connect the content of these two pages, it indicates that the source document itself has logical inconsistencies, and the PPT merely faithfully reflects the original document. Therefore, the system determines that the source document has logical gaps. In this case, the system will not deduct points from the PPT, but will instead prompt the user that the source document itself may have logical gaps.

[0127] In this embodiment of the invention, the local re-retrieval and attribution mechanism greatly improves the intelligence and robustness of logical evaluation. Compared with the traditional solution that considers any inconsistency as a problem with the PPT, this embodiment can distinguish whether the PPT is missing something or the original text is missing, avoiding the situation where the PPT is wrongly penalized due to logical jumps in the source document itself, and significantly reducing the false alarm rate. At the same time, when a logical jump is determined, the system has already found the missing intermediate node. This allows the system not only to point out the error, but also to directly recommend the missing original text segment to the user, helping the user quickly generate a transition page. This truly realizes a closed loop from evaluation to optimization, simulating the human cognitive process of going back to reread the original text if they don't understand it, and realizing intelligent logical repair suggestions.

[0128] Based on the above embodiments, step 130 includes:

[0129] The presentation's page sequence is structured and parsed to obtain the text content of each page in the sequence, and the text content is mapped to a text semantic vector.

[0130] Extract the visual information of each page and project the visual information onto the vector space where the text semantic vector is located to obtain the visual semantic vector of each page;

[0131] By fusing the text semantic vector and visual semantic vector of each page, a page vector is obtained, and a page vector sequence is determined based on the page vector of each page.

[0132] Specifically, the process of performing multimodal parsing on the page sequence of the presentation document can include:

[0133] First, the system processes the explicit text information in the PPT. That is, for each page in the PPT's page sequence, the system extracts all text content within the page, such as titles, body text, and notes, using a combination of XML structure parsing and OCR visual recognition technologies. Then, a pre-trained language model, such as BERT, can be used to transform the text content into a high-dimensional vector, namely a text semantic vector. This vector accurately expresses the literal meaning of the text on the page.

[0134] Simultaneously, the system can extract visual information from each page and project this information onto the vector space containing the text semantic vector, obtaining the visual semantic vector for each page. That is, the system not only focuses on the text but also extracts the visual information for each page, including page screenshots, layout structure (such as the position of text boxes), font size, font color, and font bolding status. Furthermore, since the original visual information (such as pixel matrices, layout coordinates, etc.) resides in a completely different space from the text semantic vector, this embodiment of the invention introduces a linear projection layer, such as a linear transformation matrix or a shallow neural network. Through this linear projection layer, the system maps the extracted visual information into the same vector space as the text semantic vector. After this transformation, the visual information is converted into a visual semantic vector.

[0135] Finally, within the same vector space, the system fuses the textual semantic vectors and visual semantic vectors of the same page to obtain the page vectors for each page. This fusion can be achieved through vector addition, concatenation followed by dimensionality reduction, or weighted summation based on an attention mechanism. Through this fusion, the generated page vector becomes a composite of both literal and visual meaning; it is no longer simply text encoding but contains the implicit information that "this sentence is more important because it has been amplified." The page vectors of each page, arranged sequentially, form the page vector sequence.

[0136] In this embodiment of the invention, by projecting visual information into the text space, the problem of traditional methods that only focus on text and not images is solved, and the semanticization of visual information is realized. This enables the system to understand the emphasis and de-emphasis logic implied in the PPT layout, thereby helping to improve the human-like level of the evaluation. Furthermore, the fusion of text semantic vectors and visual semantic vectors simulates the way humans read PPTs by combining text and images, allowing the system to recognize, like a human, that titles are more important than footnotes during evaluation, thus providing results that are more in line with human intuition in subsequent evaluations.

[0137] Based on the above embodiments, step 120 includes:

[0138] The source document is segmented into paragraphs and parsed hierarchically to obtain a multi-branch text tree. The nodes in the multi-branch text tree correspond to the text fragments in the source document, and there is a hierarchical relationship between the nodes.

[0139] Encode each node in the multi-branch text tree to obtain the semantic vector corresponding to each node;

[0140] A multi-branch semantic tree is constructed based on the multi-branch text tree and the semantic vectors corresponding to each node.

[0141] Specifically, the process of semantically segmenting the source document to obtain a multi-branch semantic tree, as described above, may include:

[0142] First, the system can use natural language processing techniques to process the source document, such as identifying physical line breaks, punctuation marks, and style tags in the source document using a pre-trained document segmentation model. Based on this information, the system will segment the source document into paragraphs, breaking the continuous text stream into independent text blocks.

[0143] Simultaneously, the system performs hierarchical parsing, determining the logical hierarchy (hierarchical and sequential relationships) between these text blocks based on the heading level (chapter, section, subsection, etc.). Based on this relationship, the system constructs a multi-branch text tree. In this multi-branch text tree, the root node represents the document's topic, the branch nodes represent chapters at various levels, and the leaf nodes correspond to specific text paragraphs. At this point, each node in the tree still stores the original text string (text fragment), and the hierarchical relationship between nodes is clearly defined through pointers; for example, Section 1.1 is a child node of Chapter 1.

[0144] Next, to make these nodes computable, the system needs to convert the text into vectors. That is, it traverses each node in the multi-branch text tree, using conventional feature extraction models and large language models as encoders to encode each node, i.e., the corresponding text segment. Through this process, the original text string stored in each node is transformed into a high-dimensional dense vector, i.e., a semantic vector. This vector captures the core meaning, keywords, and contextual features of the text segment.

[0145] Finally, the system can construct a multi-branch semantic tree using the multi-branch text tree and the semantic vectors of each node. That is, it can reconstruct a multi-branch semantic tree using the hierarchical relationships of the multi-branch text tree and the semantic vectors of each node; alternatively, it can directly attach the semantic vectors of each node back to the corresponding node in the multi-branch text tree. Thus, the original multi-branch text tree containing only text strings is upgraded to a multi-branch semantic tree that includes both structural information and semantic vectors. In this final generated tree, each node not only knows its parent and child nodes (structural hierarchy) but also possesses a mathematical identity (semantic vector) that can be used for computation.

[0146] In this embodiment of the invention, the combination of structured and semantic analysis not only preserves the hierarchical relationship of the source document but also extracts its semantic vector, enabling subsequent evaluation to check both the accuracy of the content and the logical coherence. This greatly improves the processing capability for long documents. At the same time, by decomposing long documents into a tree structure, the system can break down the complex full-text comparison task into local comparison tasks targeting tree nodes, thereby significantly reducing computational complexity and making fine-grained evaluation of long documents of hundreds of pages possible.

[0147] Based on the above embodiments, a consistency quality assessment is performed on the presentation and source documents based on the page segment association matrix to obtain the quality assessment results. The process then includes:

[0148] Construct a reward function that includes accuracy, coverage, logicality, and redundancy;

[0149] The quality score of the presentation is determined based on the quality assessment results and the reward function;

[0150] Obtain feedback scores for the quality assessment results;

[0151] The reward function is updated based on the quality score and feedback score to ensure that the quality score determined based on the updated reward function and the corresponding feedback score meet the consistency condition.

[0152] Specifically, in order to make the evaluation criteria less rigid and more adaptive to changes in industry needs, an evolutionary mechanism based on feedback is introduced in this embodiment of the invention.

[0153] In detail, the system pre-constructs a multi-dimensional comprehensive scoring formula, namely the reward function. This function aims to simulate the scoring logic of human examiners, and it integrates the following four dimensions:

[0154] Accuracy: This examines whether the PPT content is faithful to the original text and whether there are any illusions.

[0155] Coverage: This assesses whether the PPT has omitted any core knowledge points from the document.

[0156] Logicality: This examines whether the narrative order of the PPT is logical and coherent.

[0157] Redundancy: This examines whether the PPT contains repetitive or redundant content.

[0158] Next, the system can quantify and map the quality assessment results onto each dimension of the reward function, thereby obtaining scores for each dimension, namely accuracy score, coverage score, logicality score, and redundancy score; and by substituting these scores into the reward function, the quality score of the PPT can be obtained.

[0159] Here, the reward function can be expressed in the following form:

[0160]

[0161] in, Indicates quality score, Indicates the accuracy score. Indicates the coverage score. Indicates logical reasoning score. Indicates the redundancy score. , , and They are respectively , , and The corresponding weights.

[0162] Furthermore, to verify the accuracy of the system's scoring, this embodiment of the invention introduces external feedback, namely, feedback scores from human experts or more sophisticated review models on the same PPT. For example, although the system gives it 82 points, a human expert might believe that the PPT omits crucial risk warnings and only give it 60 points. This 60 points is the true value that the system needs to learn. Therefore, after receiving the feedback scores, the system can use reinforcement learning algorithms, such as Proximal Policy Optimization (PPO), to adjust the weights of the scores in each dimension of the reward function in reverse, based on the difference between the system's quality score (e.g., 82 points) and the feedback score (e.g., 60 points).

[0163] For example, if the system detects that its scores are inflated because it overemphasizes logical consistency while neglecting coverage, it will automatically lower them. The value, and increase The value of is adjusted so that the score calculated by the updated reward function can approximate the expert feedback score as closely as possible, thus satisfying the consistency condition.

[0164] Through continuous iteration, the reward function will become more and more accurate, eventually forming an automated scoring system that conforms to specific fields, such as rigorous financial risk control standards.

[0165] In this embodiment of the invention, a feedback-based evolutionary mechanism is introduced, which gives the system the ability to evolve on its own. This solves the problem of the one-size-fits-all approach of traditional fixed-rule evaluation, enabling the system to learn the preferences of different fields and users. For example, the medical field values ​​accuracy more, while the advertising field values ​​logic more, thus achieving dynamic alignment of evaluation standards. On this basis, as expert feedback data accumulates, the system's scoring will become closer and closer to the level of human experts, eventually replacing manual labor for high-quality automated quality inspection, significantly reducing labor costs.

[0166] The manuscript quality assessment device provided by the present invention is described below. The manuscript quality assessment device described below can be referred to in correspondence with the manuscript quality assessment method described above.

[0167] Figure 3 This is a schematic diagram of the manuscript quality assessment device provided by the present invention, as shown below. Figure 3 As shown, the device includes:

[0168] The acquisition unit 310 is used to acquire the presentation to be evaluated and the source document corresponding to the presentation;

[0169] The segmentation unit 320 is used to perform semantic segmentation on the source document to obtain a multi-branch semantic tree, wherein each node in the multi-branch semantic tree contains the semantic vector of the corresponding text segment in the source document.

[0170] The parsing unit 330 is used to perform multimodal parsing on the page sequence of the presentation to obtain a page vector sequence;

[0171] The association unit 340 is used to construct a page segment association matrix based on the similarity between each page vector in the page vector sequence and the semantic vector of each node in the multi-branch semantic tree;

[0172] Evaluation unit 350 is used to perform a consistency quality assessment on the presentation and the source document based on the page segment association matrix, and obtain a quality assessment result.

[0173] The document quality assessment device provided by this invention achieves a deep understanding of the logical structure of the source document and the multimodal content of the PPT by structuring the source document into a multi-branch semantic tree and parsing the PPT into a sequence of page vectors. Furthermore, based on the multi-branch semantic tree and the sequence of page vectors, a page segment association matrix is ​​constructed at a fine-grained semantic level, achieving precise semantic alignment from document paragraphs to PPT pages. Evaluation based on this page segment association matrix can not only effectively identify whether the content is accurate, but also locate the specific source of deviation, thereby significantly improving the accuracy and reliability of PPT quality assessment and providing strong technical support for consistency verification between PPT and source documents in office automation scenarios.

[0174] Based on the above embodiments, each node in the multi-branch semantic tree also contains a node weight, and the node weight represents the content importance of the corresponding text segment;

[0175] Evaluation unit 350 is used for:

[0176] Based on the page segment association matrix, each page in the page sequence of the presentation is associated with and located with each text segment in the source document to obtain the content location result;

[0177] Based on the node weights corresponding to each node in the multi-branch semantic tree, core nodes are determined, and based on the similarity of the core nodes in the page segment association matrix, important content coverage detection is performed on the presentation and the source document to obtain key coverage results.

[0178] Based on the page segment association matrix, conflict detection is performed on the text content of each page of the presentation and each text segment in the source document to obtain the conflict detection result.

[0179] Based on the content location results, the key coverage results, and the conflict detection results, the quality assessment results are determined.

[0180] Based on the above embodiments, the evaluation unit 350 is used for:

[0181] Based on the content positioning results, the key coverage results, and the conflict detection results, the content evaluation results are determined.

[0182] Extract the semantic evolution path of the page vector sequence of the presentation and perform a depth-first search on the multi-branch semantic tree to obtain the narrative logic path of the source document;

[0183] Based on the topological similarity between the semantic evolution path and the narrative logic path, the logical consistency between the presentation and the source document is evaluated to obtain the logical evaluation result.

[0184] The quality assessment result is determined based on the content assessment result and the logical assessment result.

[0185] Based on the above embodiments, the evaluation unit 350 is used for:

[0186] Based on the topological similarity between the semantic evolution path and the narrative logic path, the logical consistency between the presentation and the source document is evaluated to obtain the consistency evaluation result.

[0187] Based on the semantic transfer distance between the page vectors of adjacent pages in the page vector sequence, a logical coherence analysis is performed on the presentation and the source document to obtain the coherence analysis results.

[0188] Based on the consistency assessment results and the coherence analysis results, the logical assessment results are determined.

[0189] Based on the above embodiments, the evaluation unit 350 is used for:

[0190] If the semantic transfer distance exceeds the fault threshold, then the logical breakpoint between the corresponding adjacent pages is determined.

[0191] Using the preceding and following pages of the logical breakpoint as search criteria, a local search is performed in the multi-branch semantic tree;

[0192] If a node with semantic association with both the preceding and following pages is found in the multi-branch semantic tree, the coherence analysis result is determined to be a logical jump in the presentation; otherwise, the coherence analysis result is determined to be a logical missing element in the source document.

[0193] Based on the above embodiments, the parsing unit 330 is used for:

[0194] The page sequence of the presentation is structured and parsed to obtain the text content of each page in the page sequence, and the text content is mapped into a text semantic vector;

[0195] Extract the visual information of each page and project the visual information onto the vector space where the text semantic vector is located to obtain the visual semantic vector of each page;

[0196] By fusing the text semantic vector and visual semantic vector of each page, a page vector is obtained, and the page vector sequence is determined based on the page vector of each page.

[0197] Based on the above embodiments, the segmentation unit 320 is used for:

[0198] The source document is segmented into paragraphs and parsed hierarchically to obtain a multi-branch text tree. The nodes in the multi-branch text tree correspond to text segments in the source document, and there is a hierarchical relationship between the nodes.

[0199] Encode each node in the multi-branch text tree to obtain the semantic vector corresponding to each node;

[0200] Based on the multi-branch text tree and the semantic vectors corresponding to each node, a multi-branch semantic tree is constructed.

[0201] Based on the above embodiments, the evaluation unit 350 is further configured to:

[0202] Construct a reward function that includes accuracy, coverage, logicality, and redundancy;

[0203] Based on the quality assessment results and the reward function, the quality score of the presentation is determined;

[0204] Obtain feedback scores for the quality assessment results;

[0205] Based on the quality score and the feedback score, the reward function is updated so that the quality score determined based on the updated reward function and the corresponding feedback score meet the consistency condition.

[0206] Figure 4 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 4As shown, the electronic device may include: a processor 410, a communications interface 420, a memory 430, and a communication bus 440, wherein the processor 410, the communications interface 420, and the memory 430 communicate with each other through the communication bus 440. The processor 410 can call logical instructions in the memory 430 to execute a document quality assessment method. This method includes: acquiring a presentation to be assessed and the corresponding source document; performing semantic segmentation on the source document to obtain a multi-branch semantic tree, where each node in the multi-branch semantic tree contains a semantic vector of the corresponding text segment in the source document; performing multimodal parsing on the page sequence of the presentation to obtain a page vector sequence; constructing a page segment association matrix based on the similarity between each page vector in the page vector sequence and the semantic vector of each node in the multi-branch semantic tree; and performing a consistency quality assessment on the presentation and the source document based on the page segment association matrix to obtain a quality assessment result.

[0207] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0208] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program comprising program instructions, wherein when the program instructions are executed by a computer, the computer is able to execute the document quality assessment method provided by the above methods, the method comprising: acquiring a presentation to be evaluated and a source document corresponding to the presentation; performing semantic segmentation on the source document to obtain a multi-branch semantic tree, each node in the multi-branch semantic tree containing a semantic vector of a corresponding text segment in the source document; performing multimodal parsing on the page sequence of the presentation to obtain a page vector sequence; constructing a page segment association matrix based on the similarity between each page vector in the page vector sequence and the semantic vector of each node in the multi-branch semantic tree; and performing a consistency quality assessment on the presentation and the source document based on the page segment association matrix to obtain a quality assessment result.

[0209] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the document quality assessment method provided by the methods described above. This method includes: acquiring a presentation to be evaluated and a source document corresponding to the presentation; performing semantic segmentation on the source document to obtain a multi-branch semantic tree, where each node in the multi-branch semantic tree contains a semantic vector of a corresponding text segment in the source document; performing multimodal parsing on the page sequence of the presentation to obtain a page vector sequence; constructing a page segment association matrix based on the similarity between each page vector in the page vector sequence and the semantic vector of each node in the multi-branch semantic tree; and performing a consistency quality assessment on the presentation and the source document based on the page segment association matrix to obtain a quality assessment result.

[0210] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0211] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0212] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for evaluating manuscript quality, characterized in that, include: Obtain the presentation to be evaluated, as well as the source document corresponding to the presentation; The source document is semantically segmented to obtain a multi-branch semantic tree, and each node in the multi-branch semantic tree contains the semantic vector of the corresponding text segment in the source document; Multimodal parsing is performed on the page sequence of the presentation to obtain a page vector sequence; Based on the similarity between each page vector in the page vector sequence and the semantic vector of each node in the multi-branch semantic tree, a page segment association matrix is ​​constructed; Based on the page segment association matrix, a consistency quality assessment is performed on the presentation and the source document to obtain the quality assessment results; Each node in the multi-branch semantic tree also contains a node weight, which represents the content importance of the corresponding text segment; The consistency quality assessment of the presentation and the source document based on the page segment association matrix yields the quality assessment results, including: Based on the page segment association matrix, each page in the page sequence of the presentation is associated with and located with each text segment in the source document to obtain the content location result; Based on the node weights corresponding to each node in the multi-branch semantic tree, core nodes are determined, and based on the similarity of the core nodes in the page segment association matrix, important content coverage detection is performed on the presentation and the source document to obtain key coverage results. Based on the page segment association matrix, conflict detection is performed on the text content of each page of the presentation and each text segment in the source document to obtain the conflict detection result. Based on the content location results, the key coverage results, and the conflict detection results, the quality assessment results are determined.

2. The manuscript quality assessment method according to claim 1, characterized in that, The process of determining the quality assessment result based on the content location result, the key coverage result, and the conflict detection result includes: Based on the content positioning results, the key coverage results, and the conflict detection results, the content evaluation results are determined. Extract the semantic evolution path of the page vector sequence of the presentation and perform a depth-first search on the multi-branch semantic tree to obtain the narrative logic path of the source document; Based on the topological similarity between the semantic evolution path and the narrative logic path, the logical consistency between the presentation and the source document is evaluated to obtain the logical evaluation result. The quality assessment result is determined based on the content assessment result and the logical assessment result.

3. The manuscript quality assessment method according to claim 2, characterized in that, The logical consistency assessment of the presentation and the source document is performed based on the topological similarity between the semantic evolution path and the narrative logic path, yielding the logical assessment result, including: Based on the topological similarity between the semantic evolution path and the narrative logic path, the logical consistency between the presentation and the source document is evaluated to obtain the consistency evaluation result. Based on the semantic transfer distance between the page vectors of adjacent pages in the page vector sequence, a logical coherence analysis is performed on the presentation and the source document to obtain the coherence analysis results. Based on the consistency assessment results and the coherence analysis results, the logical assessment results are determined.

4. The manuscript quality assessment method according to claim 3, characterized in that, The logical coherence analysis of the presentation and the source document is performed based on the semantic transfer distance between page vectors of adjacent pages in the page vector sequence to obtain coherence analysis results, including: If the semantic transfer distance exceeds the fault threshold, then the logical breakpoint between the corresponding adjacent pages is determined. Using the preceding and following pages of the logical breakpoint as search criteria, a local search is performed in the multi-branch semantic tree; If a node with semantic association with both the preceding and following pages is found in the multi-branch semantic tree, the coherence analysis result is determined to be a logical jump in the presentation; otherwise, the coherence analysis result is determined to be a logical missing element in the source document.

5. The manuscript quality assessment method according to any one of claims 1 to 4, characterized in that, The step of performing multimodal parsing on the page sequence of the presentation to obtain a page vector sequence includes: The page sequence of the presentation is structured and parsed to obtain the text content of each page in the page sequence, and the text content is mapped into a text semantic vector; Extract the visual information of each page and project the visual information onto the vector space where the text semantic vector is located to obtain the visual semantic vector of each page; By fusing the text semantic vector and visual semantic vector of each page, a page vector is obtained, and the page vector sequence is determined based on the page vector of each page.

6. The manuscript quality assessment method according to any one of claims 1 to 4, characterized in that, The semantic segmentation of the source document to obtain a multi-branch semantic tree includes: The source document is segmented into paragraphs and parsed hierarchically to obtain a multi-branch text tree. The nodes in the multi-branch text tree correspond to text segments in the source document, and there is a hierarchical relationship between the nodes. Encode each node in the multi-branch text tree to obtain the semantic vector corresponding to each node; Based on the multi-branch text tree and the semantic vectors corresponding to each node, a multi-branch semantic tree is constructed.

7. The manuscript quality assessment method according to any one of claims 1 to 4, characterized in that, The process of performing a consistency quality assessment on the presentation and the source document based on the page segment association matrix to obtain a quality assessment result further includes: Construct a reward function that includes accuracy, coverage, logicality, and redundancy; Based on the quality assessment results and the reward function, the quality score of the presentation is determined; Obtain feedback scores for the quality assessment results; Based on the quality score and the feedback score, the reward function is updated so that the quality score determined based on the updated reward function and the corresponding feedback score meet the consistency condition.

8. A manuscript quality assessment device, characterized in that, include: The acquisition unit is used to acquire the presentation to be evaluated and the source document corresponding to the presentation; The segmentation unit is used to perform semantic segmentation on the source document to obtain a multi-branch semantic tree, wherein each node in the multi-branch semantic tree contains the semantic vector of the corresponding text fragment in the source document; The parsing unit is used to perform multimodal parsing on the page sequence of the presentation to obtain a page vector sequence; The association unit is used to construct a page segment association matrix based on the similarity between each page vector in the page vector sequence and the semantic vector of each node in the multi-branch semantic tree; An evaluation unit is used to perform a consistency quality assessment on the presentation and the source document based on the page segment association matrix, and obtain a quality assessment result; Each node in the multi-branch semantic tree also contains a node weight, which represents the content importance of the corresponding text segment; The consistency quality assessment of the presentation and the source document based on the page segment association matrix yields the quality assessment results, including: Based on the page segment association matrix, each page in the page sequence of the presentation is associated with and located with each text segment in the source document to obtain the content location result; Based on the node weights corresponding to each node in the multi-branch semantic tree, core nodes are determined, and based on the similarity of the core nodes in the page segment association matrix, important content coverage detection is performed on the presentation and the source document to obtain key coverage results. Based on the page segment association matrix, conflict detection is performed on the text content of each page of the presentation and each text segment in the source document to obtain the conflict detection result. Based on the content location results, the key coverage results, and the conflict detection results, the quality assessment results are determined.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the manuscript quality assessment method as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the manuscript quality assessment method as described in any one of claims 1 to 7.

11. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the manuscript quality assessment method as described in any one of claims 1 to 7.

Images