diff --git a/OSPP__toWord/document_sample.md b/OSPP__toWord/document_sample.md
new file mode 100644
index 0000000000..f9996220ee
--- /dev/null
+++ b/OSPP__toWord/document_sample.md
@@ -0,0 +1,248 @@
+
+
+Original Text: 新媒体与社会
+
+New Media and Society
+
+ISSN, CN
+
+# Online First Paper of New Media and Society
+
+Title: Formal Trust and Progressive Reductio ad Absurdum: Research on the Credibility Evaluation of Artificial Intelligence-Generated Content
+
+Authors: Yang Yanni, Wu Leilong, Xiang Anling, Zhang Jiacheng
+
+Online First Date: 2025-09-11
+
+Citation Format: Yang Yanni, Wu Leilong, Xiang Anling, Zhang Jiacheng. Formal Trust and Progressive Reductio ad Absurdum: Research on the Credibility Evaluation of Artificial Intelligence-Generated Content [J/OL]. New Media and Society.
+
+https://link.cnki.net/urlid/CN.20250910.1710.008
+
+
+
+
+First Online Release: In the workflow of an editorial department, a manuscript undergoes several stages from acceptance to publication, including final acceptance draft, typeset draft, and complete issue compilation draft. The final acceptance draft refers to a manuscript whose content has been finalized and has been approved for publication after peer review and final approval by the editor-in-chief. The typeset draft refers to the manuscript that has been typeset according to the specific layout of the journal (including the online presentation format), with the publication year, volume, issue, and page numbers temporarily undetermined. The complete issue compilation draft refers to the complete compilation of manuscripts for print or digital publication, with the publication year, volume, issue, and page numbers all determined. The content of manuscripts released online as final acceptance drafts must comply with the relevant provisions of the "Regulations on Publication Administration" and the "Regulations on the Administration of Journal Publication": academic research results must be innovative, scientific, and advanced, meeting the editorial department's acceptance requirements for published articles, and free from academic misconduct and other infringing acts. The content of the manuscript should generally conform to the national technical standards for book and journal editing and publication, with correct and unified use of language, symbols, numbers, foreign letters, legal units of measurement, and map annotations. To ensure the seriousness of the online first release of final acceptance drafts, once published, the title, author(s), institutional affiliation(s), and academic content of the paper cannot be modified; only minor textual revisions based on editorial norms are allowed.
+
+Publication Confirmation: The editorial department of a print journal, by signing an agreement with the Electronic Magazine Co., Ltd. of "China Academic Journals (CD Edition)," establishes an online edition on the "China Academic Journals (Network Edition)" publication and dissemination platform that is consistent in content with the print journal. The final acceptance drafts, typeset drafts, and complete issue compilation drafts of papers are published in the form of individual articles or complete issues before print publication. Since the "China Academic Journals (Network Edition)" is a network continuous publication approved by the State Administration of Press, Publication, Radio, Film and Television (ISSN 2096-4188, CN 11-6037/Z), papers released online first in the network edition of the contracted journal are considered formally published.
+
+# Formal Trust and Hierarchical Fallacy: Research on the Credibility Assessment of Artificial Intelligence-Generated Content\*
+
+Yanni Yang, Leilong Wu, Anling Xiang, Jiacheng Zhang\*\*
+
+Abstract Establishing a scientific and effective credibility evaluation system for Artificial Intelligence-Generated Content (AIGC) is an important topic in current human-machine trust research. This study integrates the examination of AIGC's media attributes from the perspective of intelligent communication and constructs a comprehensive and systematic AIGC credibility evaluation system from the bottom up, ranging from the underlying technical architecture and intermediary media channels to user-interactive information content. Based on hierarchical analysis and thorough expert 论证 (demonstration/discussion), the main influencing factors of AIGC credibility are clarified. The research shows that content credibility is the primary factor affecting users' perception of AIGC credibility. Output results and training data have a significant impact on system credibility, while neutrality and social influence affect media credibility, and substantive and formal credibility influence content credibility. Optimization strategies are proposed to enhance AIGC credibility from aspects such as system technology, media platforms, content cues, and user reverse verification, providing references for the optimization design of AIGC products, user trust building, and information ecosystem governance.
+
+Keywords Artificial Intelligence-Generated Content, credibility evaluation framework, formal trust, hierarchical fallacy
+
+Keywords: AIGC; Credibility; Evaluation Framework; Formal Trust; Hierarchical Fallacy
+
+## I. Research Background: The Challenge of Trustworthy Communication in the Context of AIGC
+
+Currently, the production of global information content is undergoing a new generational revolution. With the support of technologies such as large language models, Artificial Intelligence Generated Content (AIGC) has surpassed traditional Professional Generated Content (PGC) and User Generated Content (UGC) in terms of production scale and output efficiency, gradually becoming the new engine of media content production in the Web3.0 era. As AIGC accelerates its penetration into industrial applications and mass communication, the human-centered paradigm of content production and dissemination has begun to shift towards an ecosystem of human-machine collaboration and symbiosis. The quasi-subjective status of AI in agenda-setting, content gatekeeping, and information distribution has become prominent, leading to imbalances and reconfigurations in information production modes and information power. The resulting human-machine trust issues have become a bottleneck for the development of AIGC.
+
+Particularly in the supply of trustworthy content, current AIGC products have faced significant criticism. On the one hand, limited by the openness of training corpus data for large language models and their probabilistic content generation logic, errors such as conceptual confusion, factual splicing, temporal and spatial dislocation, and false causation frequently occur in the content generated by some AIGC products, making fact-checking a "hard injury" for most AIGC products. On the other hand, although AIGC-related algorithms themselves have no inherent moral attributes, this technology has been widely used for negative purposes such as political manipulation and unfair business competition [1], such as AI-generated fake news and mass-generated false information [2]. Additionally, multimodal AIGC, including images, videos, and audio, possesses strong visual ambiguity, further exacerbating the spread of false and misleading information, posing new challenges to international politics, society, and human rights development [3].
+
+In response to issues such as content inaccuracy and human-machine trust arising from AIGC, multiple countries and regions have strengthened institutional constraints and legal protections at the top-level design level. China has also introduced policy documents such as the "Ethical Guidelines for the New Generation of Artificial Intelligence," "Provisions on the Administration of Deep Synthesis in Internet Information Services," and the "Administrative Measures for Generative Artificial Intelligence Services (Draft for Comments)," explicitly stating that "content generated using generative artificial intelligence should be true and accurate." Both domestic and international AIGC platforms have also improved content quality through multiple constraints such as algorithmic iteration, model correction, and manual review, but these efforts are mostly limited to political information and sensitive content. Trustworthy communication in general domains remains a pain point for industry development. From the user perspective, compared to the critical attitude of users in European and American countries towards AIGC, Chinese users exhibit a more positive attitude towards AIGC applications [4], with their perceived credibility of algorithm-generated news even surpassing their trust in human journalists [5]. If this high perceived trust and optimistic orientation towards AIGC lack rational guidance, it may facilitate the spread of false information, posing new risks to the public opinion environment and information security.
+
+With the accelerated penetration of AIGC products into the mass application market, questions arise regarding users' perceived trust in these products, the factors influencing this trust, and how to effectively evaluate the credibility of AIGC. These issues require further exploration. Against this backdrop, this study takes current typical text-based AI-generated content domestically and internationally as the research object and constructs an AIGC credibility evaluation system from three dimensions—system credibility, media credibility, and content credibility—based on the Trustworthy AI Framework, Media Trust, and Dual-Process Theory. This aims to uncover subjective perceptual factors and objective technological factors influencing credibility, providing theoretical references and practical guidance for intelligent communication and information ecosystem governance.
+
+## Literature Review: Trustworthy Systems, Trustworthy Media, and Trustworthy Content
+
+The research theme of credibility exhibits significant interdisciplinary characteristics, with disciplines such as communication studies, management, computer science, and psychology conducting relevant explorations. Scholars' interpretations of credibility vary. Overall, the analysis of AIGC credibility mainly unfolds from the following perspectives: First, credibility is viewed as an entity composed of multiple facets or dimensions. Wu Dan et al. analyzed the impact of content integrity, question-answering interactivity, and tool anthropomorphism on the credibility of generative intelligent search from three dimensions [6]. Jiang Zhongbo et al. measured credibility using five items: accuracy, trustworthiness, bias, authenticity, and authority. Second, credibility is explained through its sources [5]. Fogg et al. proposed interpreting credibility from the perspective of information sources, such as presumed credibility derived from stereotypes, reputational credibility generated by third-party endorsements, surface credibility obtained from external characteristics, and experiential credibility judged from experience [7]. Third, the subjects of credibility are categorized for analysis. Song Shijie et al. proposed constructing an AIGC credibility research framework from four perspectives: information sources, interaction modes, social actors, and algorithmic metaphors [8]. Liu Haiming and Li Jiaguai argued that in human-machine interaction scenarios, trust originates from the reliability, credibility, and responsible content generation of algorithmic systems on the one hand, and is reflected in the anthropomorphic characteristics of chatbots that bring positive responses to user interaction processes on the other hand [9]. Additionally, some scholars name credibility based on the research objects involved, such as review credibility and advertisement credibility.
+
+Building on credibility research across multiple traditional disciplines, the credibility of AIGC has emerged as a new research topic. With the widespread application of artificial intelligence technology, besides evaluating AI's credibility from a system technology perspective by treating it as a software system and analyzing content credibility from an information content perspective as in traditional research, scholars have recognized the value and importance of AI as a medium. Therefore, in the interpretation of AIGC credibility, it is also necessary to examine its credibility from a media dimension. A comprehensive review of typical domestic and international research findings on credibility from the three dimensions of system credibility, media credibility, and content credibility is presented in Table 1, with only a few examples listed due to space constraints.
+
+Table 1 Research Findings Related to Credibility Evaluation at Home and Abroad
+| Evaluation Perspective | Evaluation Dimension | Researcher | Evaluation Approach |
| Source | System Credibility | EU[10] | Accountability, Inclusiveness, Autonomy, Fairness, Privacy, Robustness, Security, Transparency |
| OECD[11] | Sustainable Development, Values, Fairness, Transparency, Explainability, Robustness, Security, Safety |
| Singh[12] | Fairness, Explainability, Robustness, Privacy, Security, Appropriateness |
| Fujii[1] | Integrity, Robustness, System Quality, Agility |
| He Jifeng[14] | Robustness, Self-Reflection, Adaptability, and Fairness |
| Channel | Media Credibility | Schweiger [15] | Media Type, Media Subcategory, Media Product, Editorial Unit, Information Creator, Information Presenter |
| Metzger[16] | Source, Information, and Channel Transmitting Information |
| Li Xiaojing[17] | Credibility of Media Organizations/Journalists (Source), Credibility of News Reports (Information), etc. |
| Zhang Hongzhong and Ren Wujiong[18] | Public Trust in Mass Media, Trust in Social Media, Human-Machine Trust Based on Machine Identity, Human-Machine Trust Based on Language Dialogue |
| Content | Content Credibility | Sundar[19] | MAIN Model (From Technological Affordance to Credibility Judgment) |
| Flanagin and Metzger [20] | Dual-Process Model (Heuristic, Systematic) |
| | Hilligoss and Rieh [21] | Integrated Framework for Credibility Assessment (Construction Layer, Exploration Layer, Interaction Layer) |
| | | 互层） |
Figure 1 AIGC Credibility Evaluation System
+
+### 1. System Credibility
+
+From the perspective of system technology, "trustworthiness" is a system metric developed based on concepts such as "reliability" and "security." It represents an overall evaluation by humans of various trustworthy attributes during the research, development, and application of systems [23]. It reflects not only the objective performance of the system itself but also the subjective perception of users towards the system. Compared to early trustworthy systems that focused on hardware devices, trustworthy artificial intelligence (trustworthy AI) places more emphasis on trustworthy attributes at the software level. It refers to AI systems that, during the design, development, deployment, and use processes, aim to gain the trust and acceptance of users and society based on attributes such as product performance and risk assurance [24].
+
+The trustworthiness of AI systems relies on a substantial amount of high-quality, reliable data support, appropriate algorithm model applications, and expected output results. Considering the characteristics of AIGC technology and existing relevant research, for AIGC products, the evaluation of system credibility should comprehensively consider three aspects: underlying training data, algorithm models, and output results. The noise, biases, and privacy infringement risks hidden in the underlying data significantly impact the credibility of the generated content. At the algorithm model level, generative AI, through reinforcement learning based on human feedback, aligns to some extent with human common sense, cognition, needs, and even values. It also introduces filtering mechanisms to block sensitive issues (such as violence, crime, discrimination, etc.), reducing dissemination risks by refusing to answer or providing relatively neutral and safe responses. This ethical constraint and optimization mechanism at the algorithmic level ensure content security to a certain extent. However, the "malleable" nature of algorithms also increases the "discipline" risks of AI itself, and the risk of content untrustworthiness caused by malicious training cannot be ignored. Additionally, limited by the underlying computational logic of statistical language models and the effectiveness of interactive instructions (prompts), AI output results also exhibit instability.
+
+Based on relevant research, this study focuses on analyzing the training data layer from perspectives such as dataset transparency, source reliability, data robustness, data coverage, data scalability, noise removal, and source security. The algorithm model layer is evaluated based on model interpretability, robustness, stability, adaptability, privacy protection, and malleability. For output results, indicators such as the accuracy, error rate, precision, recall, and F-score of the generated content need to be analyzed.
+
+### 2. Media Credibility
+
+In the context of AIGC, the boundaries between media organizations, algorithm platforms, and content producers tend to blur. AI algorithms, as a technological means, embed capabilities for content production, distribution, and agenda-setting, gradually demonstrating a trend of deep mediatization. As an emerging media form, if technological credibility considers AIGC from its underlying operational logic, media credibility evaluates AIGC more as a media carrier and communication entity in terms of its credibility.
+
+Media credibility refers to whether technological channels and specific information dissemination organizations are trustworthy [25]. In early research on media credibility, the competency (i.e., the degree of professionalization) of communicators was a core factor influencing credibility [26]. Competency includes not only the depth and authority of media in professional fields but also their professionalism in information gatekeeping, packaging, and distribution. It directly determines whether the media can output correct and trustworthy information and is also related to the media's brand endorsement, i.e., whether it can take responsibility for the authenticity of the information content. Subsequent related research has added factors such as unselfishness and consistency [27,28] based on media competency. "Unselfishness" relates to whether the media's communication motives are legitimate and pure, i.e., whether the media itself has any 利益 (interest) associations with the subjects it reports on or the issues it involves. For content output platforms, consistency is more reflected in whether there are value biases and contradictions. In addition to ensuring professionalism, unselfishness, and consistency in content output, external means such as effective interaction with audiences, providing feedback loops, establishing transparency and ethical standards, and certification by third-party independent organizations can also help media enhance their credibility [29][30].
+
+Based on relevant literature reviews and considering the technological characteristics of AIGC, this study evaluates its media credibility from four aspects: professional service capabilities, social influence, neutrality of stance, and the establishment of safeguard mechanisms. Professional service capabilities mainly focus on the content accumulation of AIGC platforms in professional fields, collaborations with professional media, and standardized process mechanisms. Social influence includes the platform's social visibility, social evaluation, third-party certifications, and endorsements. Neutrality of stance encompasses the platform's unselfishness, consistency, independence, and profitability. The establishment of safeguard mechanisms involves the platform's user feedback mechanisms, accountability mechanisms, privacy mechanisms/data security mechanisms, etc., and relates to whether the platform can ensure information quality and security.
+
+### 3. Content Credibility
+
+According to the dual-process theory, humans often activate dual cognitive systems when processing and reasoning information: one is a fast, intuitive, and emotional processing mode, and the other is a slow, controlled, and rational thinking mode [3]. The former can process a large amount of information in a short time and quickly make judgments and generate answers, while the latter requires logical reasoning and abstract thinking to form judgments and verify answers. Corresponding to trust mechanisms, intuitive information processing often 激发 (elicits) "form-based trust," while controlled information processing forms "substantive trust." Form-based trust is more based on 认同感 (identification) generated by the surface features and credible cues of the content. Different credible cues form cognitive shortcuts by activating users' existing cognitive structures and attitude tendencies, further helping users judge the credibility of the content. From the perspective of content attributes, this includes the degree of emotionalization of information, the use of multimedia forms in the content, whether other users are mentioned, the reputation of external links/sources included in the information, and the use of specific vocabulary [32]. From the perspective of external cues, this includes social endorsement cues (such as data on reposts, likes, and comments), temporal cues (such as update frequency, timeliness, etc.), and reputation cues (such as historical performance, third-party certifications, integrity records, etc.). Form-based trust often determines users' initial credibility of the content, and this initial credibility plays a significant mediating role in users' subsequent belief changes [33].
+
+Compared to form-based trust, substantive trust mainly explores the credibility of the essence of the content. This includes the objectivity and accuracy of the content, whether there are obvious biases or stance tendencies, and whether there are obvious errors or false information [34]. Elements such as the completeness, consistency, and timeliness of the content also determine its essential credibility, while users' own knowledge levels, experiences, and judgment abilities directly influence their perception differences [35]. The formation of substantive trust often requires users to engage in thoughtful analysis and judgment of the content and progressive 溯源推理 (溯源 reasoning). Both in terms of trust depth and trust duration, substantive trust is stronger than form-based trust. Faced with issues such as widespread errors, factual inconsistencies, political biases, information security, and ethical risks in current AIGC products, analyzing credible cues at both the form and substantive levels and engaging in progressive falsification are considered necessary.
+
+Based on the above discussion, this study divides the content credibility of AIGC into two aspects: formal credibility and substantive credibility. Formal credibility is primarily analyzed from the structural and linguistic characteristics of AIGC, encompassing aspects such as citation norms, response consistency, expression precision, source professionalism, authoritative endorsement, and information noise. Substantive credibility, on the other hand, is mainly evaluated based on the essential attributes of the content, including content authenticity, content accuracy, content objectivity, information timeliness, information completeness, information consistency, quality robustness, content security, and content fairness.
+
+## (II) AIGC Credibility Evaluation Methods
+
+Scholars both domestically and internationally have proposed various subjective or objective evaluation methods for decision analysis targeting specific objectives. The Analytic Hierarchy Process (AHP), introduced by American operations researcher Professor T.L. Saaty in the 1970s, quickly gained widespread application across various fields for its ability to quantitatively analyze qualitative problems through a simple and flexible multi-criteria decision-making approach. Its core idea is to decompose complex decision-making problems into multiple levels and determine the weight coefficients of various factors by constructing a judgment matrix and conducting consistency checks. In this study, we first assembled an expert team to ensure the consistency and effectiveness of team evaluations. Subsequently, based on the Analytic Hierarchy Process, we established a hierarchical structure model. The expert team evaluated the indicators within the aforementioned credibility evaluation framework, constructed a judgment matrix, and conducted consistency checks to calculate and determine the weights of indicators at all levels.
+
+### 1. Expert Team Assembly and Training
+
+The expertise and experience of evaluation experts directly influence the accuracy and reliability of annotation results. To this end, this study selected five experts with extensive experience in the fields of artificial intelligence and natural language processing, including two algorithm engineers, two research scholars in related disciplines, and one senior media practitioner. These experts possess in-depth experience with AIGC products and can effectively understand various instructional documents. Prior to conducting evaluation annotations, staff members introduced the purpose and background of this study, explained the definitions of each indicator, and provided examples and simulated scenarios to ensure that all experts had a clear understanding of the evaluation methods. To ensure the stability and reliability of the evaluation results, this study employed Cohen's Kappa standard measurement method to conduct consistency checks on expert scores. The results showed a Kappa value of 0.647, with P < 0.001, indicating that the experts' understanding and evaluation of each indicator were generally consistent, allowing for subsequent credibility evaluations. Although the scoring process involves subjective judgment, this study ensured the accuracy and reliability of the evaluation results as much as possible by strictly adhering to the diversity of the expert team and the standardization of the scoring methods.
+
+### 2. Indicator Evaluation and Weight Determination
+
+## (1) Establishing a Hierarchical Structure Model
+
+Based on the previously constructed AIGC credibility evaluation framework, the Analytic Hierarchy Process was employed to determine the weights of primary and secondary indicators. First, a hierarchical structure was established, divided into the objective layer, criterion layer, and sub-criterion layer. The objective layer is the AIGC credibility evaluation (U), and the criterion layer includes three primary indicators: system credibility (U1), media credibility (U2), and content credibility (U3). These three main factors correspond to several sub-factors, forming the sub-criterion layer, which includes nine sub-factors such as training data credibility ($\mathbf{U}_{11}$), algorithm model credibility ($\mathrm{U}_{12}$), and output result credibility ($\mathbf{U}_{13}$). The hierarchical structure results are shown in Table 2.
+
+Table 2 Hierarchical Structure of AIGC Credibility Evaluation System
+| Objective Layer | Criterion Layer | Sub-criterion Layer | Description of Sub-criterion Layer Indicators |
| AIGC | System Credibility | Training Data Credibility | Dataset Transparency, Source Reliability, Data Robustness, Da |
| Credibility Assessment U | (U1) | (U11) | ta Coverage, Data Scalability, Noise Removal, and Source Security |
| Algorithm Model Credibility $\left(\mathrm{U}_{12}\right)$ | Model Interpretability, Robustness, Stability, Adaptability, Privacy Protection, and Plasticity |
| Output Result Credibility (U13) | Accuracy, Error Rate, Precision, Recall, and F-measure of Generated Content |
| Media Credibility (U2) | Professional Service Capability $\left(\mathrm{U}_{21}\right)$ | Content Accumulation in Professional Fields, Professional Media Cooperation, Standardized Process Mechanisms |
| Social Influence $\left(\mathrm{U}_{22}\right)$ | Platform's Social Awareness, Social Evaluation, Third-party Certification and Endorsement |
| Neutrality of Stance $\left(\mathrm{U}_{23}\right)$ | Platform's Selflessness, Consistency, Independence, Profitability |
| Guarantee Mechanism Construction (U24) | Platform's User Feedback Mechanism, Accountability Mechanism, Privacy Mechanism/Data Security Mechanism |
| Content Credibility (U3) | Formal Credibility (U31) | Citation Normativity, Response Consistency, Expression Precision, Source Professionalism, Authoritative Endorsement, Information Noise |
| Substantive Credibility (U32) | Content Authenticity, Content Correctness, Content Objectivity, Information Timeliness, Information Completeness, Information Consistency, Quality Robustness, Content Security, Content Fairness |
Table 3 Nine-level Scaling Method
+| Relative Importance Level | Definition | Description |
| 1 | Equally Important | The importance of the two indicators is the same |
| 3 | Slightly Important | Based on experience or judgment, one indicator is slightly more important |
| 5 | Quite Important | Based on experience or judgment, one indicator is quite important |
| 7 | Extremely Important | In practice, one indicator is extremely important |
| 6 | Absolutely Important | There is sufficient evidence to show that one indicator is absolutely important |
| 2, 4, 6, 8 | Intermediate Value of Adjacent Judgments | Used when a compromise judgment on indicators is required |
Table 4 Average Consistency Index (RI) Value
+| n | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 |
| RI | 0 | 0 | 0.58 | 0.90 | 1.12 | 1.24 | 1.32 | 1.41 | 1.46 |
Table 5 $\mathrm{U{-}U_{k}}$ judgment matrix
+| $\mathrm{U{-}U_{k}}$ | $\mathrm{U}_{1}$ | $\mathrm{U}_{2}$ | $\mathrm{U}_{3}$ |
| $\mathrm{U}_{1}$ | 1 | 2 | 1/2$ |
| $\mathrm{U}_{2}$ | 1/2 | 1 | 1/5$ |
| $\mathrm{U}_{3}$ | 2 | 5 | 1 |
Table 6 $\scriptstyle\mathrm{U_{1}-U_{1i}}$ Judgment Matrix
+| $\mathrm{U_{1}\mathrm{-U_{1i}}}$ | $\mathrm{U_{11}}$ | $\mathrm{U}_{12}$ | $\mathrm{U}_{13}$ |
| $\mathrm{U_{11}}$ | | $2$ | $1/2$ |
| $\mathrm{U_{12}}$ | $1/2$ | $1$ | $1/4$ |
| $\mathrm{U_{13}}$ | | $4$ | $1$ |
Table 7 $U_{2}-U_{2i}$ judgment matrix
+| $\mathrm{U_{2}\mathrm{-U_{2i}}}$ | $\mathrm{U}_{21}$ | $\mathrm{U}_{22}$ | $\mathrm{U}_{23}$ | $\mathrm{U}_{24}$ |
| $\mathrm{U}_{21}$ | 1 | 1/5 | 1/6 | $1/3$ |
| $\mathrm{U}_{22}$ | 5 | 1 | 1 | 3 |
| $\mathrm{U}_{23}$ | 6 | 1 | 1 | 3 |
| $\mathrm{U}_{24}$ | 3 | 1/3 | 1/3 | 1 |
| $\mathrm{U_{3}\mathrm{-U_{3i}}}$ | $\mathrm{U}_{31}$ | $\mathrm{U}_{32}$ |
| $\mathrm{U}_{31}$ | 1 | $1/7$ |
| $\mathrm{U}_{32}$ | 7 | 1 |
Table 9 Hierarchical Weights of Trustworthy Factors for AIGC
+| U | System Credibility $\mathrm{U}_{1}$ | Medium Credibility $\mathrm{U}_{2}$ | Content Credibility $\mathrm{U}_{3}$ | Overall Ranking Weight |
| 0.277 | 0.128 | 0.595 |
| Training Data Credibility $\mathrm{U_{11}}$ | 0.286 | | | 0.079 |
| Algorithm Model Credibility $\mathrm{U}_{12}$ | 0.143 | | | 0.040 |
| Output Result Credibility $\mathrm{U_{13}}$ | 0.571 | | | 0.158 |
| Professional Service Capability $\mathrm{U}_{21}$ | | 0.065 | | 0.008 |
| Social Influence $\mathrm{U}_{22}$ | | 0.384 | | 0.049 |
| Neutrality of Stance $\mathrm{U}_{23}$ | (missing value corrected to) 0.401 | 0.051 | | Guarantee Mechanism Construction $\mathrm{U}_{24}^{-}$ |
| 0.150 | | 0.019 | | Formal Credibility $\mathrm{U}_{31}$ |
| 0.125 | | | 0.074 | Substantive Credibility $\mathrm{U}_{32}$ |
| 0.875 | | | 0.521 | 0.521 |
Table 1 Research Findings Related to Credibility Evaluation at Home and Abroad
+| Evaluation Perspective | Evaluation Dimension | Researcher | Evaluation Approach |
| Source | System Credibility | EU[10] | Accountability, Inclusiveness, Autonomy, Fairness, Privacy, Robustness, Security, Transparency |
| OECD[11] | Sustainable Development, Values, Fairness, Transparency, Explainability, Robustness, Security, Safety |
| Singh[12] | Fairness, Explainability, Robustness, Privacy, Security, Appropriateness |
| Fujii[1] | Integrity, Robustness, System Quality, Agility |
| He Jifeng[14] | Robustness, Self-Reflection, Adaptability, and Fairness |
| Channel | Media Credibility | Schweiger [15] | Media Type, Media Subcategory, Media Product, Editorial Unit, Information Creator, Information Presenter |
| Metzger[16] | Source, Information, and Channel Transmitting Information |
| Li Xiaojing[17] | Credibility of Media Organizations/Journalists (Source), Credibility of News Reports (Information), etc. |
| Zhang Hongzhong and Ren Wujiong[18] | Public Trust in Mass Media, Trust in Social Media, Human-Machine Trust Based on Machine Identity, Human-Machine Trust Based on Language Dialogue |
| Content | Content Credibility | Sundar[19] | MAIN Model (From Technological Affordance to Credibility Judgment) |
| Flanagin and Metzger [20] | Dual-Process Model (Heuristic, Systematic) |
| | Hilligoss and Rieh [21] | Integrated Framework for Credibility Assessment (Construction Layer, Exploration Layer, Interaction Layer) |
| | | 互层） |
Figure 1 AIGC Credibility Evaluation System
+
+### 1. System Trustworthiness
+
+From the perspective of system technology, "trustworthiness" is a system metric developed on the basis of concepts such as "reliability" and "security." It represents an overall evaluation by humans of various trustworthy attributes in the process of system research, development, and application [23]. It reflects not only the objective performance of the system itself but also the subjective perception of users towards the system. Compared to early trustworthy systems that focused on hardware devices, trustworthy artificial intelligence (trustworthy AI) places more emphasis on trustworthy attributes at the software level. It refers to AI systems that, during the processes of design, development, deployment, and use, are based on attributes such as product performance and risk assurance to gain the trust and acceptance of users and society [24].
+
+The trustworthiness of AI systems relies on the support of a large amount of high-quality, reliable data, appropriate algorithm model applications, and compliance with
+
+1
+
+Expected output results. Considering the characteristics of AIGC technology and existing relevant research, the evaluation of system credibility for AIGC products requires a comprehensive assessment from three aspects: underlying training data, algorithmic models, and output results. The noise, biases, and privacy infringement risks hidden in the underlying data significantly impact the credibility of the generated content. At the algorithmic model level, generative artificial intelligence, through reinforcement learning based on human feedback, aligns itself to a certain extent with human common sense, cognition, needs, and even values. Simultaneously, filtering mechanisms are introduced to block sensitive issues (such as violence, crime, discrimination, etc.), reducing dissemination risks by refusing to answer or providing relatively neutral and safe responses. These ethical constraints and optimization mechanisms at the algorithmic level ensure content security to a certain extent. However, the "malleable" nature of algorithms also increases the risk of "disciplining" AI itself, and the risk of content untrustworthiness due to malicious training cannot be ignored. Additionally, limited by the underlying computational logic of statistical language models and the effectiveness of interactive instructions (prompts), AI output results also exhibit instability.
+
+Based on relevant research, this study focuses on analyzing the training data layer from perspectives such as dataset transparency, source reliability, data robustness, data coverage, data scalability, noise removal, and source security. The algorithmic model layer is evaluated based on model interpretability, robustness, stability, adaptability, privacy protection, and malleability. For output results, indicators such as the accuracy, error rate, precision, recall rate, and F-score of the generated content need to be analyzed.
+
+### 2. Media Credibility
+
+In the context of AIGC, the boundaries between media organizations, algorithmic platforms, and content producers tend to blur. AI algorithms, as a technological means, embed capabilities for content production, distribution, and agenda-setting, gradually exhibiting a trend of deep mediatization. As an emerging media form, if technological credibility considers AIGC from its underlying operational logic, media credibility assesses AIGC's credibility more as a media carrier and communication subject.
+
+Media credibility refers to whether technological channels and specific information dissemination organizations are trustworthy [25]. In early research on media credibility (media credibility), the competence (competency) of communicators, i.e., their level of professionalization, was a core factor influencing credibility [26]. Competence includes not only the depth and authority of media in their professional fields but also their level of professionalization in communication processes such as information gatekeeping, packaging, and distribution. It directly determines whether the media can output correct and credible information and is also related to the media's own brand endorsement, i.e., its ability to take responsibility for the authenticity of the information content. Subsequent related research added factors such as unselfishness and consistency [27,28] based on media competence. "Unselfishness" relates to whether the media's communication motives are legitimate and pure, i.e., whether the media itself has any利益 (Note: "利益" is left untranslated here as it is a Chinese term embedded in the English context for illustrative purposes; in a complete translation, it would be translated as "interests" or "vested interests") connections with the subjects it reports on or the issues it involves. For content output platforms, consistency is more reflected in whether there are value biases and inconsistencies over time. In addition to ensuring professionalism, unselfishness, and consistency at the content output level, external means such as effective interaction with audiences, providing feedback loops, establishing transparency and ethical standards, and certification by third-party independent organizations can also help media enhance their credibility [29][30].
+
+Based on relevant literature reviews and considering the technological characteristics of AIGC, this study evaluates from the perspectives of professional service capabilities, social influence, and
+
+1
+
+Its media credibility is evaluated from four dimensions: professional service capability, social influence, neutrality of stance, and the establishment of a safeguard mechanism. Professional service capability primarily focuses on the content accumulation of AIGC platforms in professional fields, collaborations with professional media, and standardized procedural mechanisms. Social influence encompasses the platform's social recognition, public evaluation, third-party certifications, and endorsements. Neutrality of stance includes the platform's selflessness, consistency, independence, and profitability. The establishment of a safeguard mechanism involves aspects such as the platform's user feedback mechanisms, accountability mechanisms, and privacy/data security mechanisms, all of which are crucial for ensuring information quality and security on the platform.
+
+### 3. Content Credibility
+
+According to the dual-process theory, humans often activate dual cognitive systems when processing and reasoning information: one is a fast, intuitive, and emotional processing mode, while the other is a slow, controlled, and rational thinking mode [3]. The former can process a large amount of information in a short time and quickly make judgments and generate answers, while the latter requires logical reasoning and abstract thinking to form judgments and verify answers. Corresponding to trust mechanisms, intuitive information processing often triggers "formal trust," while controlled information processing forms "substantive trust." Formal trust is more based on the identification generated by the surface characteristics and credible cues of the content. Different credible cues form cognitive shortcuts by activating users' existing cognitive structures and attitude tendencies, further helping users judge the credibility of the content. From the perspective of content attributes, this includes the emotional intensity of the information, the use of multimedia forms in the content, whether other users are mentioned, the reputation of external links/sources included in the information, and the use of specific vocabulary [32]. From the perspective of external cues, this includes social recommendation cues (such as data on reposts, likes, and comments), temporal cues (such as update frequency and timeliness), and reputation cues (such as historical performance, third-party certifications, and integrity records). Formal trust often determines the initial credibility of the content for users, and this initial credibility plays a significant mediating role in the subsequent changes in users' beliefs [33].
+
+Compared to formal trust, substantive trust primarily explores the credibility of the essence of the content. This includes the objectivity and accuracy of the content, the presence or absence of obvious biases or stance tendencies, and the presence or absence of obvious errors or false information [34]. Elements such as the completeness, consistency, and timeliness of the content also determine its essential credibility, while users' own knowledge levels, experiences, and judgment abilities directly influence their perception differences [35]. The formation of substantive trust often requires users to engage in thoughtful analysis and judgment of the content, as well as progressive溯源 (Note: "溯源" is translated as "溯源 reasoning," meaning tracing back to the source for reasoning, to convey the idea of tracing the origin or cause of something for logical deduction) reasoning. It is stronger than formal trust in terms of both the depth and duration of trust. Faced with the prevalent issues of errors and confusion, factual inconsistencies, political biases, information security, and ethical risks in AIGC products, analyzing credible cues at both formal and substantive levels and engaging in progressive reductio ad absurdum are deemed necessary.
+
+Based on the above discussion, this study divides the content credibility of AIGC into two aspects: formal credibility and substantive credibility. Formal credibility is primarily analyzed from the structural and linguistic characteristics of AIGC, including citation norms, response consistency, expression precision, source professionalism, authoritative endorsements, and information noise. Substantive credibility primarily evaluates the essential attributes of the content, including content authenticity, content correctness, content objectivity, information timeliness, information completeness, information consistency, quality robustness, content security, and content fairness.
+
+1
+
+## (II) AIGC Credibility Evaluation Method
+
+For decision-making analysis targeting specific objectives, scholars both domestically and internationally have proposed various subjective or objective evaluation methods. The Analytic Hierarchy Process (AHP), introduced by American operations researcher Professor T.L. Saaty in the 1970s, quickly gained widespread application across various fields for its ability to conduct quantitative analysis of qualitative problems through a simple and flexible multi-criteria decision-making approach. The core idea is to decompose complex decision-making problems into multiple levels and determine the weight coefficients of various factors by constructing a judgment matrix and conducting consistency checks. In this study, an expert team was first established to ensure the consistency and effectiveness of team evaluations. Subsequently, based on the Analytic Hierarchy Process, a hierarchical structure model was constructed. The expert team evaluated the indicators within the credibility evaluation framework established above, constructed a judgment matrix, and conducted consistency checks to calculate and determine the weights of indicators at all levels.
+
+### 1. Expert Team Formation and Training
+
+The expertise and experience of evaluation experts directly influence the accuracy and reliability of annotation results. To address this, this study selected five experts with extensive experience in the fields of artificial intelligence and natural language processing, including two algorithm engineers, two research scholars in related disciplines, and one senior media practitioner. These experts had in-depth experience using AIGC products and were well-versed in understanding various instructional documents. Prior to conducting evaluation annotations, staff members introduced the purpose and background of this study, explained the definitions of each indicator, and provided examples and simulated scenarios to ensure that all experts had a clear understanding of the evaluation methods. To ensure the stability and reliability of the evaluation results, this study employed Cohen's Kappa standard measurement method to conduct consistency checks on expert scores. The results showed a Kappa value of 0.647, with P < 0.001, indicating that the experts' understanding and evaluation of each indicator were generally consistent, allowing for subsequent credibility evaluations. Although the scoring process involved subjective judgments, this study ensured the accuracy and reliability of the evaluation results as much as possible by strictly adhering to the diversity of the expert team and the standardization of the scoring methods.
+
+### 2. Indicator Evaluation and Weight Determination
+
+## (1) Establishment of a Hierarchical Structure Model
+
+Based on the AIGC credibility evaluation framework constructed earlier, the Analytic Hierarchy Process was employed to determine the weights of primary and secondary indicators. First, a hierarchical structure was established, divided into the objective level, criterion level, and sub-criterion level. The objective level is the AIGC credibility evaluation (U), and the criterion level includes three primary indicators: system credibility (U1), media credibility (U2), and content credibility (U3). These three main factors correspond to several sub-factors, forming the sub-criterion level, including nine sub-factors such as training data credibility ($\mathbf{U}_{11}$), algorithm model credibility ($\mathrm{U}_{12}$), and output result credibility ($\mathbf{U}_{13}$). The hierarchical structure results are shown in Table 2.
+
+Table 2 Hierarchical Structure of AIGC Credibility Evaluation System
+| Objective Layer | Criterion Layer | Sub-criterion Layer | Indicator Description of Sub-criterion Layer |
| AIGC | System Credibility | Training Data Credibility | Dataset Transparency, Source Reliability, Data Robustness, Da |
| Credibility Assessment U | (U1) | (U11) | ta Coverage, Data Scalability, Noise Removal, and Source Security |
| Algorithm Model Credibility $\left(\mathrm{U}_{12}\right)$ | Model Interpretability, Robustness, Stability, Adaptability, Privacy Protection, and Plasticity |
| Output Result Credibility (U13) | Accuracy, Error Rate, Precision, Recall, and Completeness of Generated Content |
| Media Credibility (U2) | Professional Service Capability $\left(\mathrm{U}_{21}\right)$ | Content Accumulation in Professional Fields, Professional Media Cooperation, Standardized Process Mechanisms |
| Social Influence $\left(\mathrm{U}_{22}\right)$ | Platform's Social Awareness, Social Evaluation, Third-Party Certification, and Endorsement |
| Neutrality of Stance $\left(\mathrm{U}_{23}\right)$ | Platform's Selflessness, Consistency, Independence, and Profitability |
| Construction of Guarantee Mechanisms (U24) | Platform's User Feedback Mechanism, Accountability Mechanism, Privacy Mechanism/Data Security Mechanism |
| Content Credibility (U3) | Formal Credibility (U31) | Citation Normativity, Response Consistency, Expression Precision, Source Professionalism, Authoritative Endorsement, Information Noise |
| Substantive Credibility (U32) | Content Authenticity, Content Correctness, Content Objectivity, Information Timeliness, Information Completeness, Information Consistency, Quality Robustness, Content Security, Content Fairness |
Table 3 Nine-level Scaling Method
+| Relative Importance Level | Definition | Description |
| 1 | Equally Important | The importance of the two indicators is the same |
| 3 | Slightly Important | Based on experience or judgment, one indicator is slightly more important |
| 5 | Quite Important | Based on experience or judgment, one indicator is quite important |
| 7 | Extremely Important | In practice, one indicator is extremely important |
| 6 | Absolutely Important | There is sufficient evidence to show that one indicator is absolutely important |
| 2, 4, 6, 8 | Intermediate Value of Adjacent Judgments | Used when a compromise judgment on indicators is required |
Table 4 Average Consistency Index (RI) Value
+| n | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 |
| RI | 0 | 0 | 0.58 | 0.90 | 1.12 | 1.24 | 1.32 | 1.41 | 1.46 |
Table 5 $\mathrm{U{-}U_{k}}$ judgment matrix
+| $\mathrm{U{-}U_{k}}$ | $\mathrm{U}_{1}$ | $\mathrm{U}_{2}$ | $\mathrm{U}_{3}$ |
| $\mathrm{U}_{1}$ | 1 | 2 | 1/2$ |
| $\mathrm{U}_{2}$ | 1/2 | 1 | 1/5$ |
| $\mathrm{U}_{3}$ | 2 | 5 | 1 |
Table 6 $\scriptstyle\mathrm{U_{1}-U_{1i}}$ Judgment Matrix
+| $\mathrm{U_{1}\mathrm{-U_{1i}}}$ | $\mathrm{U_{11}}$ | $\mathrm{U}_{12}$ | $\mathrm{U}_{13}$ |
| $\mathrm{U_{11}}$ | | $2$ | $1/2$ |
| $\mathrm{U_{12}}$ | $1/2$ | $1$ | $1/4$ |
| $\mathrm{U_{13}}$ | | $4$ | $1$ |
Table 7 $U_{2}-U_{2i}$ judgment matrix
+| $\mathrm{U_{2}\mathrm{-U_{2i}}}$ | $\mathrm{U}_{21}$ | $\mathrm{U}_{22}$ | $\mathrm{U}_{23}$ | $\mathrm{U}_{24}$ |
| $\mathrm{U}_{21}$ | 1 | 1/5 | 1/6 | $1/3$ |
| $\mathrm{U}_{22}$ | 5 | 1 | 1 | 3 |
| $\mathrm{U}_{23}$ | 6 | 1 | 1 | 3 |
| $\mathrm{U}_{24}$ | 3 | 1/3 | 1/3 | 1 |
| $\mathrm{U_{3}\mathrm{-U_{3i}}}$ | $\mathrm{U}_{31}$ | $\mathrm{U}_{32}$ |
| $\mathrm{U}_{31}$ | 1 | $1/7$ |
| $\mathrm{U}_{32}$ | 7 | 1 |
Table 9 Hierarchical Weights of Trustworthy Factors for AIGC
+| U | System Credibility $\mathrm{U}_{1}$ | Medium Credibility $\mathrm{U}_{2}$ | Content Credibility $\mathrm{U}_{3}$ | Overall Ranking Weight |
| 0.277 | 0.128 | 0.595 |
| Training Data Credibility $\mathrm{U_{11}}$ | 0.286 | | | 0.079 |
| Algorithm Model Credibility $\mathrm{U}_{12}$ | 0.143 | | | 0.040 |
| Output Result Credibility $\mathrm{U_{13}}$ | 0.571 | | | 0.158 |
| Professional Service Capability $\mathrm{U}_{21}$ | | 0.065 | | 0.008 |
| Social Influence $\mathrm{U}_{22}$ | | 0.384 | | 0.049 |
| Neutrality of Stance $\mathrm{U}_{23}$ | (missing value corrected to) 0.401 | 0.051 | | Guarantee Mechanism Construction $\mathrm{U}_{24}^{-}$ |
| 0.150 | | 0.019 | | Formal Credibility $\mathrm{U}_{31}$ |
| 0.125 | | | 0.074 | Substantive Credibility $\mathrm{U}_{32}$ |
| 0.875 | | | 0.521 | 0.521 |
| page | header | footer | footnote | page_number |
|---|
| 1 | | | | |
| 2 | https://link.cnki.net/urlid/CN.20250910.1710.008 | | *Funding Project: Youth Project of the National Natural Science Foundation of China, "Research on Risk Identification and Governance Strategies for AI-Generated Content" (Project Number: 72304290).
+**Yang Yanni, Associate Professor at the School of Literature and Media, China Three Gorges University; Wu Leilong, Master's Student at the School of Literature and Media, China Three Gorges University;
+Xiang Anling, Lecturer at the School of Journalism and Communication, Minzu University of China; Zhang Jiacheng, Postdoctoral Fellow at the School of Journalism and Communication, Tsinghua University. | |
| 3 | | | | |
| 4 | | | | |
| 5 | | | | |
| 6 | | | | |
| 7 | | | | |
| 8 | | | | |
| 9 | | | | |
| 10 | | | | |
| 11 | | | | |
| 12 | | | | |
| 13 | | | | |
| 14 | | | | |
| 15 | | | | |
| 16 | | | | |
| 17 | | | | |
| 18 | | | | |
| 19 | | | | |
| 20 | | | | |
diff --git a/OSPP__toWord/pdf_to_json_to_latex.py b/OSPP__toWord/pdf_to_json_to_latex.py
new file mode 100644
index 0000000000..f1d2251869
--- /dev/null
+++ b/OSPP__toWord/pdf_to_json_to_latex.py
@@ -0,0 +1,380 @@
+# Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import json, os, re
+
+
+def escape_latex(s: str) -> str:
+ """
+ 转义 LaTeX 特殊字符（对普通文本使用）
+ """
+ if not s:
+ return ""
+ return (
+ s.replace("\\", "\\textbackslash{}")
+ .replace("&", "\\&")
+ .replace("%", "\\%")
+ .replace("$", "\\$")
+ .replace("#", "\\#")
+ .replace("_", "\\_")
+ .replace("{", "\\{")
+ .replace("}", "\\}")
+ .replace("~", "\\textasciitilde{}")
+ .replace("^", "\\textasciicircum{}")
+ )
+
+
+def get_image_width_from_md(
+ md_path: str, image_name: str, default_ratio: float = 0.8
+) -> float:
+ """
+ 从 Markdown/HTML 文件中提取 ![]()
的 width（仅解析百分比，如 10% -> 0.10）。
+ 支持 width="10%"、width=10%、style="width:10%;" 等格式。
+ 返回 0 < ratio <= 1.0，找不到则返回 default_ratio。
+ """
+ if not md_path or not os.path.exists(md_path):
+ return default_ratio
+
+ with open(md_path, "r", encoding="utf-8") as f:
+ content = f.read()
+
+ flags = re.I | re.DOTALL
+
+ # 优先匹配 width 属性
+ pat_width_attr = re.compile(
+ rf']*src\s*=\s*["\']?[^"\'>]*{re.escape(image_name)}[^"\'>]*["\']?[^>]*\bwidth\s*=\s*["\']?(\d+)\s*%?["\']?',
+ flags,
+ )
+ m = pat_width_attr.search(content)
+ if m:
+ try:
+ val = int(m.group(1))
+ if val > 0:
+ return min(max(val / 100.0, 0.01), 1.0)
+ except:
+ pass
+
+ return default_ratio
+
+
+def escaped_paragraph_text(s: str) -> str:
+ """
+ 处理 text block：
+ - 普通文本转义
+ - 保留公式原样
+ """
+ paragraphs = re.split(r"\n\s*\n", s)
+ processed_paras = []
+
+ for p in paragraphs:
+ p = p.strip()
+ if not p:
+ continue
+
+ # 临时占位公式
+ placeholders = []
+
+ def placeholder_repl(m):
+ placeholders.append(m.group(0))
+ return f"@@FORMULA{len(placeholders)-1}@@"
+
+ formula_pattern = re.compile(r"(\$\$.*?\$\$|\$.*?\$|\\\[.*?\\\])", re.DOTALL)
+ temp_text = formula_pattern.sub(placeholder_repl, p)
+ temp_text = escape_latex(temp_text)
+
+ # 替换回公式
+ for i, formula in enumerate(placeholders):
+ temp_text = temp_text.replace(f"@@FORMULA{i}@@", formula)
+
+ processed_paras.append("\\par " + temp_text)
+
+ return "\n\n".join(processed_paras) + "\n\n"
+
+
+def generate_image_latex(block, image_base_path, md_base_path) -> str:
+ """
+ 图像/表格处理函数
+ """
+ bbox = block.get("block_bbox", [0, 0, 0, 0])
+ try:
+ x1, y1, x2, y2 = map(int, bbox)
+ except:
+ x1, y1, x2, y2 = 0, 0, 0, 0
+
+ filenames = [
+ f"img_in_chart_box_{x1}_{y1}_{x2}_{y2}.jpg",
+ f"img_in_image_box_{x1}_{y1}_{x2}_{y2}.jpg",
+ block.get("file_name") or block.get("image") or "",
+ ]
+ image_path = None
+ for fname in filenames:
+ if fname:
+ candidate = os.path.join(image_base_path or "", fname)
+ if os.path.exists(candidate):
+ image_path = os.path.abspath(candidate)
+ break
+
+ if not image_path:
+ return f"% [Image not found: {filenames[0]} / {filenames[1]}]\n\n"
+
+ caption_text = escape_latex(block.get("caption", "").strip())
+
+ # 获取宽度
+ width_ratio = 0.8
+ if md_base_path:
+ page = int(block.get("page", 0) or 0)
+ md_candidates = [
+ os.path.join(md_base_path, f"page_{page}.md"),
+ os.path.join(md_base_path, f"document_sample_{page}.md"),
+ os.path.join(md_base_path, f"{page}.md"),
+ ]
+ for mdp in md_candidates:
+ if os.path.exists(mdp):
+ width_ratio = get_image_width_from_md(
+ mdp, os.path.basename(image_path), default_ratio=width_ratio
+ )
+ break
+ width_ratio = max(0.01, min(float(width_ratio), 1.0))
+
+ return (
+ f"\\begin{{figure}}[h]\n"
+ f"\\centering\n"
+ f"\\includegraphics[width={width_ratio:.2f}\\linewidth]{{{image_path}}}\n"
+ f"\\caption*{{{caption_text}}}\n"
+ f"\\end{{figure}}\n\n"
+ )
+
+
+def generate_table_latex(block) -> str:
+
+ from bs4 import BeautifulSoup
+
+ content = block.get("block_content", "")
+ if "{\\raggedright\\arraybackslash}X" for _ in range(col_count)]
+ )
+
+ latex = "\\begin{center}\n\\renewcommand{\\arraystretch}{1.5}\n"
+ latex += f"\\begin{{tabularx}}{{\\textwidth}}{{{col_format}}}\n\\toprule\n"
+ for i, row in enumerate(norm_rows):
+ latex += " & ".join(row) + " \\\\\n"
+ if i == 0:
+ latex += "\\midrule\n"
+ latex += "\\bottomrule\n\\end{tabularx}\n\\end{center}\n\n"
+ return latex
+
+
+def block_to_latex(block: dict, image_base_path: str = None, md_base_path: str = None):
+ """
+ 单个 block 转 LaTeX
+ """
+ label = block.get("block_label", "")
+ raw_content = block.get("block_content", "") or ""
+
+ if label == "doc_title":
+ content = escape_latex(raw_content.strip())
+ return f"\\begin{{center}}\n{{\\Huge {content}}}\\end{{center}}\n\n", None
+
+ if label in ["header", "footer"]:
+ return "", None
+
+ if label == "abstract":
+ content = raw_content.strip()
+ if not content:
+ return "", None
+ return (
+ f"\\begin{{abstract}}\n{escape_latex(content)}\n\\end{{abstract}}\n\n",
+ None,
+ )
+
+ if label == "paragraph_title":
+ content = escape_latex(raw_content.strip())
+ return f"\\section*{{{content}}}\n\n", None
+
+ if label == "reference":
+ lines = [line.strip() for line in raw_content.split("\n") if line.strip()]
+ bibitems = []
+ for line in lines:
+ content = escape_latex(re.sub(r"^\[\d+\]\s*", "", line))
+ key = f"ref{abs(hash(line)) % 100000}"
+ bibitems.append(f"\\bibitem{{{key}}} {content}")
+ return "\n".join(bibitems) + "\n", "\n".join(bibitems) + "\n"
+
+ if label == "text":
+ return escaped_paragraph_text(raw_content), None
+ if label == "content":
+ lines = [line.rstrip() for line in raw_content.splitlines()]
+ latex_lines_content = [
+ escape_latex(line) + " \\\\" for line in lines if line.strip()
+ ]
+ return "\n".join(latex_lines_content) + "\n\n", None
+
+ if label == "formula":
+ return f"\\[\n{raw_content.strip()}\n\\]\n\n", None
+
+ if label == "algorithm":
+ return "\\begin{verbatim}\n" + raw_content + "\n\\end{verbatim}\n\n", None
+
+ if label in ["image", "chart", "seal"]:
+ return generate_image_latex(block, image_base_path, md_base_path), None
+
+ if label == "table":
+ return generate_table_latex(block), None
+
+ if label in ["figure_title", "chart_title", "table_title"]:
+ content = escape_latex(raw_content.strip())
+ if not content:
+ return "", None
+ return f"\\begin{{center}}\n{{\\small {content}}}\\end{{center}}\n\n", None
+
+ return f"% 未知标签 {label}: {escape_latex(raw_content)}\n\n", None
+
+
+def blocks_to_latex(
+ json_path: str, tex_output_path: str, image_base_path: str, md_base_path: str = None
+):
+ if not os.path.exists(json_path):
+ print(f"❌ JSON 文件不存在: {json_path}")
+ return
+
+ with open(json_path, "r", encoding="utf-8") as f:
+ try:
+ blocks = json.load(f)
+ except Exception as e:
+ print("❌ 读取 JSON 失败：", e)
+ return
+
+ # 分页
+ pages = {}
+ for b in blocks:
+ p = int(b.get("page", 0) or 0)
+ pages.setdefault(p, []).append(b)
+
+ # LaTeX 文档头
+ latex_lines = [
+ "\\documentclass[12pt]{article}",
+ "\\usepackage{xeCJK}",
+ "\\usepackage{fontspec}",
+ "\\usepackage{graphicx}",
+ "\\usepackage{amsmath}",
+ "\\usepackage{geometry}",
+ "\\usepackage{fancyhdr}",
+ "\\usepackage{indentfirst}",
+ "\\usepackage{caption}",
+ "\\usepackage{tabularx, booktabs}",
+ "\\usepackage{amssymb}",
+ "\\usepackage{amsfonts}",
+ "\\geometry{a4paper, margin=1in}",
+ "\\setCJKmainfont{Droid Sans Fallback}",
+ "\\setmainfont{DejaVu Serif}",
+ "\\setsansfont{Lato}",
+ "\\setmonofont{Latin Modern Mono}",
+ "\\pagestyle{fancy}",
+ "\\setlength{\\parindent}{2em}",
+ "\\begin{document}\n",
+ ]
+
+ in_bibliography = False
+ pending_references = []
+
+ for page_num in sorted(pages.keys()):
+ page_blocks = sorted(
+ pages[page_num],
+ key=lambda b: (
+ b.get("block_bbox", [0, 0, 0, 0])[1] if b.get("block_bbox") else 0
+ ),
+ )
+ header_blocks = [b for b in page_blocks if b.get("block_label") == "header"]
+ footer_blocks = [b for b in page_blocks if b.get("block_label") == "footer"]
+ page_header = " ".join(b.get("block_content", "") for b in header_blocks)
+ page_footer = " ".join(b.get("block_content", "") for b in footer_blocks)
+
+ latex_lines.append(f"% ==== page {page_num} header/footer ====")
+ latex_lines.append(f"\\fancyhead[L]{{{escape_latex(page_header)}}}")
+ latex_lines.append(f"\\fancyfoot[C]{{{escape_latex(page_footer)}}}\n")
+
+ for block in page_blocks:
+ lbl = block.get("block_label", "")
+ if lbl == "reference_title":
+ if not in_bibliography:
+ latex_lines.append("\\begin{thebibliography}{99}")
+ in_bibliography = True
+ continue
+
+ tex_fragment, bib_item = block_to_latex(
+ block, image_base_path=image_base_path, md_base_path=md_base_path
+ )
+ if lbl == "reference" and bib_item:
+ if in_bibliography:
+ latex_lines.append(bib_item)
+ else:
+ pending_references.append(bib_item)
+ continue
+ if tex_fragment:
+ latex_lines.append(tex_fragment)
+
+ latex_lines.append("\\clearpage\n")
+
+ # 写入 pending references
+ if pending_references:
+ if not in_bibliography:
+ latex_lines.append("\\begin{thebibliography}{99}")
+ latex_lines.extend(pending_references)
+ if not in_bibliography:
+ latex_lines.append("\\end{thebibliography}\n")
+ elif in_bibliography:
+ latex_lines.append("\\end{thebibliography}\n")
+
+ latex_lines.append("\\end{document}")
+ latex_text = "\n".join(latex_lines)
+
+ os.makedirs(os.path.dirname(tex_output_path), exist_ok=True)
+ with open(tex_output_path, "w", encoding="utf-8") as f:
+ f.write(latex_text)
+
+ print(f"✅ LaTeX 文件已保存至: {tex_output_path}")
diff --git a/OSPP__toWord/pdf_to_json_to_latex_test.py b/OSPP__toWord/pdf_to_json_to_latex_test.py
new file mode 100644
index 0000000000..476b1b9534
--- /dev/null
+++ b/OSPP__toWord/pdf_to_json_to_latex_test.py
@@ -0,0 +1,57 @@
+# Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# -*- coding: utf-8 -*-
+import pdf_to_json_to_word as toword
+from paddlex import create_pipeline
+import pdf_to_json_to_latex as tolatex
+
+pipeline = create_pipeline(pipeline="PP-DocTranslation")
+
+input_path = "mypaddle/input/en.pdf"
+output_path = "mypaddle/upgit/output"
+
+# 该部分不需要用到翻译模块，故去掉翻译部分
+
+if input_path.lower().endswith(".md"):
+ ori_md_info_list = pipeline.load_from_markdown(input_path)
+else:
+ visual_predict_res = pipeline.visual_predict(
+ input_path,
+ use_doc_orientation_classify=False,
+ use_doc_unwarping=False,
+ use_common_ocr=True,
+ use_seal_recognition=False,
+ use_table_recognition=True,
+ )
+
+ ori_md_info_list = []
+ # 收集每页解析到的json数据
+ json_list = []
+
+ for res in visual_predict_res:
+ layout_parsing_result = res["layout_parsing_result"]
+ # 每页的json
+ json_data = layout_parsing_result._to_json()
+ json_list.append(json_data)
+
+ layout_parsing_result.markdown
+ layout_parsing_result.save_to_markdown(output_path)
+
+
+merged_json_path = toword.merge_block(json_list, output_path=output_path)
+
+tolatex.blocks_to_latex(
+ merged_json_path, f"{output_path}/output.tex", f"{output_path}/imgs", output_path
+)
diff --git a/OSPP__toWord/pdf_to_json_to_word.py b/OSPP__toWord/pdf_to_json_to_word.py
new file mode 100644
index 0000000000..a66ff6e238
--- /dev/null
+++ b/OSPP__toWord/pdf_to_json_to_word.py
@@ -0,0 +1,320 @@
+# Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import List, Dict
+import json, os, re, copy
+
+TRANSLATABLE_LABELS = {"chart", "image", "seal", "number"}
+SPLIT_TOKEN = "¥$¥"
+
+
+# --- 样式设置 ---
+# 设置段落的字体、字号、加粗、对齐方式和首行缩进
+def set_paragraph_style(
+ para,
+ font_name="Times New Roman",
+ font_size_pt=12,
+ bold=False,
+ indent=False,
+ alignment=None,
+):
+ from docx.oxml.ns import qn
+ from docx.shared import Inches
+ from docx.shared import Pt
+ from docx.enum.text import WD_ALIGN_PARAGRAPH
+
+ run = para.runs[0] if para.runs else para.add_run()
+ run.font.name = font_name
+ run._element.rPr.rFonts.set(qn("w:eastAsia"), "宋体")
+ run.font.size = Pt(font_size_pt)
+ run.bold = bold
+ if alignment is None:
+ alignment = WD_ALIGN_PARAGRAPH.LEFT
+ para.alignment = alignment
+ if indent:
+ para.paragraph_format.first_line_indent = Inches(0.3)
+
+
+# 设置 run 的字体，包括中英文字体和字号
+def set_run_font(
+ run,
+ font_name_en="Times New Roman",
+ font_name_cn="宋体",
+ font_size_pt=10.5,
+ bold=False,
+):
+ from docx.oxml.ns import qn
+ from docx.shared import Pt
+
+ run.font.name = font_name_en
+ run._element.rPr.rFonts.set(qn("w:eastAsia"), font_name_cn)
+ run.font.size = Pt(font_size_pt)
+ run.bold = bold
+
+
+# 清空并设置 section 的页眉或页脚内容，居中显示
+def set_section_part_text(section_part, text):
+ from docx.enum.text import WD_ALIGN_PARAGRAPH
+
+ for _ in range(len(section_part.paragraphs)):
+ p = section_part.paragraphs[0]
+ p._element.getparent().remove(p._element)
+ para = section_part.add_paragraph()
+ para.alignment = WD_ALIGN_PARAGRAPH.CENTER
+ run = para.add_run(text)
+ set_run_font(run)
+
+
+# --- 内容格式 ---
+# 根据块的标签和内容，向文档添加对应格式的段落或标题
+def format_block_style(doc, label, content):
+ from docx.enum.text import WD_ALIGN_PARAGRAPH
+
+ style_map = {
+ "doc_title": {
+ "level": 0,
+ "size": 20,
+ "bold": True,
+ "align": WD_ALIGN_PARAGRAPH.CENTER,
+ },
+ "header": {"size": 16, "bold": True, "align": WD_ALIGN_PARAGRAPH.CENTER},
+ "abstract_title": {
+ "level": 1,
+ "size": 14,
+ "bold": True,
+ "align": WD_ALIGN_PARAGRAPH.CENTER,
+ },
+ "content_title": {
+ "level": 1,
+ "size": 14,
+ "bold": True,
+ "align": WD_ALIGN_PARAGRAPH.LEFT,
+ },
+ "reference_title": {
+ "level": 1,
+ "size": 14,
+ "bold": True,
+ "align": WD_ALIGN_PARAGRAPH.LEFT,
+ },
+ "paragraph_title": {
+ "level": 2,
+ "size": 14,
+ "bold": True,
+ "align": WD_ALIGN_PARAGRAPH.LEFT,
+ },
+ "abstract": {"size": 12, "align": WD_ALIGN_PARAGRAPH.JUSTIFY},
+ "text": {"size": 12, "align": WD_ALIGN_PARAGRAPH.JUSTIFY, "indent": True},
+ "figure_title": {"size": 10, "align": WD_ALIGN_PARAGRAPH.CENTER},
+ "table_title": {"size": 10, "align": WD_ALIGN_PARAGRAPH.CENTER},
+ "chart_title": {"size": 10, "align": WD_ALIGN_PARAGRAPH.CENTER},
+ "reference": {"size": 12, "align": WD_ALIGN_PARAGRAPH.JUSTIFY},
+ "algorithm": {
+ "font": "Courier New",
+ "size": 11,
+ "align": WD_ALIGN_PARAGRAPH.LEFT,
+ },
+ "formula": {"size": 12, "align": WD_ALIGN_PARAGRAPH.CENTER},
+ "vision_footnote": {"size": 9, "align": WD_ALIGN_PARAGRAPH.LEFT},
+ "number": {"size": 9, "align": WD_ALIGN_PARAGRAPH.CENTER},
+ "footer": {"size": 9, "align": WD_ALIGN_PARAGRAPH.CENTER},
+ }
+
+ config = style_map.get(label, {"size": 12, "align": WD_ALIGN_PARAGRAPH.LEFT})
+ para = (
+ doc.add_heading(content, level=config["level"])
+ if "level" in config
+ else doc.add_paragraph(content)
+ )
+ set_paragraph_style(
+ para,
+ font_name=config.get("font", "Times New Roman"),
+ font_size_pt=config["size"],
+ bold=config.get("bold", False),
+ indent=config.get("indent", False),
+ alignment=config["align"],
+ )
+
+
+# --- 表格解析 ---
+# 解析 HTML 表格字符串，提取为二维文本列表
+def parse_html_table(html):
+ from bs4 import BeautifulSoup
+
+ soup = BeautifulSoup(html, "html.parser")
+ return [
+ [cell.get_text(strip=True) for cell in tr.find_all(["td", "th"])]
+ for tr in soup.find_all("tr")
+ ]
+
+
+# --- 排序 ---
+# 根据页码和块的纵坐标，对块进行排序
+def sort_blocks_by_position(blocks):
+ return sorted(blocks, key=lambda b: (b.get("page", 0), b["block_bbox"][1]))
+
+
+# 从 markdown 文件中提取指定图片的宽度比例，默认 1.0
+def get_image_width_from_md(md_path, image_name):
+ with open(md_path, "r", encoding="utf-8") as f:
+ content = f.read()
+ pattern = re.compile(
+ rf']*src=["\'].*?{re.escape(image_name)}.*?["\'][^>]*width=["\'](\d+)%["\']',
+ re.I,
+ )
+ match = pattern.search(content)
+ return int(match.group(1)) / 100 if match else 1.0
+
+
+# 向文档插入图片，宽度根据比例缩放，居中显示
+def insert_image(doc, image_path, width_ratio):
+ from docx.enum.text import WD_ALIGN_PARAGRAPH
+ from docx.shared import Inches
+
+ para = doc.add_paragraph()
+ run = para.add_run()
+ run.add_picture(image_path, width=Inches(5.5 * width_ratio))
+ para.alignment = WD_ALIGN_PARAGRAPH.CENTER
+
+
+# --- 主流程函数 ---
+# 从 JSON 读取块列表，按页生成 Word 文档，支持页眉页脚、表格、图片等
+def blocks_to_word(
+ json_path, word_output_path, image_base_path, input_path, output_path
+):
+ from docx import Document
+ from docx.enum.section import WD_SECTION
+
+ with open(json_path, "r", encoding="utf-8") as f:
+ blocks = json.load(f)
+
+ doc = Document()
+ pages = {}
+ for block in blocks:
+ pages.setdefault(block.get("page", 0), []).append(block)
+
+ for page_num in sorted(pages.keys()):
+ page_blocks = sort_blocks_by_position(pages[page_num])
+ section = (
+ doc.add_section(WD_SECTION.NEW_PAGE) if page_num != 0 else doc.sections[0]
+ )
+
+ # 页眉
+ header_blocks = [b for b in page_blocks if b["block_label"] == "header"]
+ if header_blocks:
+ header_text = "\n".join(
+ b["block_content"].strip()
+ for b in header_blocks
+ if b.get("block_content")
+ )
+ section.header.is_linked_to_previous = False
+ set_section_part_text(section.header, header_text)
+
+ # 页脚
+ footer_blocks = [b for b in page_blocks if b["block_label"] == "footer"]
+ if footer_blocks:
+ footer_text = "\n".join(
+ b["block_content"].strip()
+ for b in footer_blocks
+ if b.get("block_content")
+ )
+ section.footer.is_linked_to_previous = False
+ set_section_part_text(section.footer, footer_text)
+
+ for block in page_blocks:
+ label = block["block_label"]
+ content = block.get("block_content", "").strip()
+ if not content and label not in ["chart", "image", "table", "seal"]:
+ continue
+
+ if label in ["chart", "image", "seal"]:
+ bbox = block.get("block_bbox", [0, 0, 0, 0])
+ x1, y1, x2, y2 = map(int, bbox)
+ filename1 = f"img_in_chart_box_{x1}_{y1}_{x2}_{y2}.jpg"
+ filename2 = f"img_in_image_box_{x1}_{y1}_{x2}_{y2}.jpg"
+ image_filename = (
+ filename1
+ if os.path.exists(os.path.join(image_base_path, filename1))
+ else filename2
+ )
+ image_path = os.path.join(image_base_path, image_filename)
+
+ if os.path.exists(image_path):
+ base_name = os.path.splitext(os.path.basename(input_path))[0]
+ md_path = f"{output_path}/{base_name}_{block.get('page')}.md"
+ width = get_image_width_from_md(md_path, image_filename)
+ insert_image(doc, image_path, width)
+ else:
+ doc.add_paragraph(f"[Image {image_filename} not found]")
+ continue
+
+ elif label == "table":
+ rows = (
+ parse_html_table(content)
+ if " str:
+ merged = []
+ for i, item in enumerate(json_list):
+ blocks = copy.deepcopy(item["res"]["parsing_res_list"])
+ for b in blocks:
+ b["page"] = i
+ merged.extend(blocks)
+
+ merged_path = os.path.join(output_path, "merged_translated.json")
+ with open(merged_path, "w", encoding="utf-8") as f:
+ json.dump(merged, f, ensure_ascii=False, indent=4)
+ return merged_path
+
+
+def md2cn_word(json_list: List[Dict], input_path, output_path):
+ # 将所有的json中的block抽取出来，合并成一个json
+ merged_json_path = merge_block(json_list, output_path=output_path)
+
+ # 调用转换成 Word 的函数
+ base_name = os.path.splitext(os.path.basename(input_path))[0]
+
+ blocks_to_word(
+ json_path=merged_json_path,
+ word_output_path=f"{output_path}/{base_name}_toword.docx",
+ image_base_path=f"{output_path}/imgs",
+ input_path=input_path,
+ output_path=output_path,
+ )
diff --git a/OSPP__toWord/pdf_to_json_to_word_test.py b/OSPP__toWord/pdf_to_json_to_word_test.py
new file mode 100644
index 0000000000..96f386be59
--- /dev/null
+++ b/OSPP__toWord/pdf_to_json_to_word_test.py
@@ -0,0 +1,64 @@
+# Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# -*- coding: utf-8 -*-
+import pdf_to_json_to_word as toword
+import os
+from paddlex import create_pipeline
+
+pipeline = create_pipeline(pipeline="PP-DocTranslation")
+
+input_path = "mypaddle/upgit/document_sample.pdf"
+output_path = "mypaddle/upgit/output"
+
+# 该部分不需要用到翻译模块，故去掉翻译部分
+
+if input_path.lower().endswith(".md"):
+ ori_md_info_list = pipeline.load_from_markdown(input_path)
+else:
+ visual_predict_res = pipeline.visual_predict(
+ input_path,
+ use_doc_orientation_classify=False,
+ use_doc_unwarping=False,
+ use_common_ocr=True,
+ use_seal_recognition=False,
+ use_table_recognition=True,
+ )
+
+ ori_md_info_list = []
+ # 收集每页解析到的json数据
+ json_list = []
+
+ for res in visual_predict_res:
+ layout_parsing_result = res["layout_parsing_result"]
+ # 每页的json
+ json_data = layout_parsing_result._to_json()
+ json_list.append(json_data)
+
+ layout_parsing_result.markdown
+ layout_parsing_result.save_to_markdown(output_path)
+
+# 将所有的json中的block抽取出来，合并成一个只含block的json
+merged_json_path = toword.merge_block(json_list, output_path=output_path)
+
+# 调用转换成 Word 的函数
+base_name = os.path.splitext(os.path.basename(input_path))[0]
+
+toword.blocks_to_word(
+ json_path=merged_json_path,
+ word_output_path=f"{output_path}/{base_name}_toword.docx",
+ image_base_path=f"{output_path}/imgs",
+ input_path=input_path,
+ output_path=output_path,
+)
diff --git a/OSPP__toWord/pdf_to_md_to_latex.py b/OSPP__toWord/pdf_to_md_to_latex.py
new file mode 100644
index 0000000000..971f503a5e
--- /dev/null
+++ b/OSPP__toWord/pdf_to_md_to_latex.py
@@ -0,0 +1,274 @@
+# Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os, re
+
+
+def escape_latex_outside_formula(s: str) -> str:
+ """
+ 转义 LaTeX 特殊字符，但保留公式原样
+ """
+ if not s:
+ return ""
+
+ placeholders = []
+
+ def repl(m):
+ placeholders.append(m.group(0))
+ return f"@@FORMULA{len(placeholders)-1}@@"
+
+ # 提取公式
+ formula_pat = re.compile(
+ r"(\$\$.*?\$\$|\$.*?\$|\\\[.*?\\\]|\\\(.*?\\\))", re.DOTALL
+ )
+ tmp = formula_pat.sub(repl, s)
+
+ # 转义
+ tmp = (
+ tmp.replace("\\", "\\textbackslash{}")
+ .replace("&", "\\&")
+ .replace("%", "\\%")
+ .replace("$", "\\$")
+ .replace("#", "\\#")
+ .replace("_", "\\_")
+ .replace("{", "\\{")
+ .replace("}", "\\}")
+ .replace("~", "\\textasciitilde{}")
+ .replace("^", "\\textasciicircum{}")
+ )
+
+ # 恢复公式
+ for i, f in enumerate(placeholders):
+ tmp = tmp.replace(f"@@FORMULA{i}@@", f)
+ return tmp
+
+
+def get_image_width_from_md_line(line, default_ratio=0.8):
+ """
+ 解析图片 width 属性
+ """
+ m = re.search(r'width\s*=\s*["\']?(\d+)%?["\']?', line)
+ if m:
+ val = int(m.group(1))
+ return max(0.01, min(val / 100.0, 1.0))
+ m2 = re.search(r"width\s*:\s*(\d+)%", line)
+ if m2:
+ val = int(m2.group(1))
+ return max(0.01, min(val / 100.0, 1.0))
+ return default_ratio
+
+
+def process_table_html(content) -> str:
+
+ from bs4 import BeautifulSoup
+
+ """
+ 表格处理
+ """
+ if "{\\raggedright\\arraybackslash}X" for _ in range(col_count)]
+ )
+
+ latex = "\\begin{center}\n\\renewcommand{\\arraystretch}{1.5}\n"
+ latex += f"\\begin{{tabularx}}{{\\textwidth}}{{{col_format}}}\n\\toprule\n"
+ for i, row in enumerate(norm_rows):
+ latex += " & ".join(row) + " \\\\\n"
+ if i == 0:
+ latex += "\\midrule\n"
+ latex += "\\bottomrule\n\\end{tabularx}\n\\end{center}\n\n"
+ return latex
+
+
+def process_paragraph(s: str) -> str:
+ """
+ 处理文本段落，保留公式
+ """
+ paragraphs = re.split(r"\n\s*\n", s)
+ processed_paras = []
+ for p in paragraphs:
+ p = p.strip()
+ if not p:
+ continue
+ processed_paras.append("\\par " + escape_latex_outside_formula(p))
+ return "\n\n".join(processed_paras) + "\n\n"
+
+
+def process_md_line(line: str, output_path) -> str:
+
+ from bs4 import BeautifulSoup
+
+ """
+ 单行处理
+ """
+ line = line.strip()
+ if not line:
+ return ""
+
+ # 标题
+ if line.startswith("##### "):
+ return f"\\paragraph*{{{escape_latex_outside_formula(line[6:].strip())}}}\n\n"
+ if line.startswith("#### "):
+ return (
+ f"\\subsubsection*{{{escape_latex_outside_formula(line[5:].strip())}}}\n\n"
+ )
+ if line.startswith("### "):
+ return f"\\subsection*{{{escape_latex_outside_formula(line[4:].strip())}}}\n\n"
+ if line.startswith("## "):
+ return f"\\section*{{{escape_latex_outside_formula(line[3:].strip())}}}\n\n"
+ if line.startswith("# "):
+ return f"\\section*{{{escape_latex_outside_formula(line[2:].strip())}}}\n\n"
+
+ # 居中 div
+ if "1", md_text) if p.strip()]
+ if not pages:
+ print("❌ 没有有效内容")
+ return
+
+ # 最后一页作为页眉页脚信息
+ meta_info = pages[-1] if " Dict:
+
+ # HTML 表格头
+ html_lines = [
+ "",
+ "| page | header | footer | footnote | page_number |
|---|
",
+ ]
+
+ for page_idx, json_blocks in enumerate(json_list, start=1):
+
+ parsing_res_list = json_blocks.get("res", {}).get("parsing_res_list", [])
+
+ head_foot_dict = {"header": "", "footer": "", "footnote": "", "page_number": ""}
+
+ for block in parsing_res_list:
+ label = block.get("block_label", "").lower()
+ content = block.get("block_content", "").strip()
+ if not content:
+ continue
+
+ if label in {"header", "footer", "footnote", "number", "page_number"}:
+ if label == "number":
+ label = "page_number"
+ head_foot_dict[label] = content
+
+ # 构造表格行
+ html_lines.append(
+ f"| {page_idx} | "
+ f"{head_foot_dict['header']} | "
+ f"{head_foot_dict['footer']} | "
+ f"{head_foot_dict['footnote']} | "
+ f"{head_foot_dict['page_number']} |
"
+ )
+
+ html_lines.append("
")
+
+ # 将 HTML 行列表拼接成单个字符串
+ html_string = "".join(html_lines)
+
+ result = {
+ "markdown_images": {},
+ "page_index": 0,
+ "input_path": input_path,
+ "markdown_texts": html_string, # 单个字符串
+ "page_continuation_flags": (True, True), # 可根据需要调整
+ }
+ return result
+
+
+def process_md_page(document, md_text, output_path):
+ from bs4 import BeautifulSoup
+
+ """处理单页内容"""
+ lines = md_text.strip().split("\n")
+ for line in lines:
+ line = line.strip()
+ if not line:
+ continue
+
+ title_color = (0, 0, 255)
+ if line.startswith("##### "):
+ p = document.add_paragraph(line[6:])
+ set_paragraph_style(p, bold=True, font_size=10)
+ elif line.startswith("#### "):
+ p = document.add_paragraph(line[5:])
+ set_paragraph_style(p, bold=True, font_size=11)
+ elif line.startswith("### "):
+ p = document.add_paragraph(line[4:])
+ set_paragraph_style(p, bold=True, font_size=12)
+ elif line.startswith("## "):
+ p = document.add_paragraph(line[3:])
+ set_paragraph_style(p, bold=True, font_size=14)
+ elif line.startswith("# "):
+ p = document.add_paragraph(line[2:])
+ set_paragraph_style(p, bold=True, font_size=16)
+
+ # 居中内容处理
+ elif line.startswith("1", md_text) if p.strip()]
+
+ if not pages:
+ print("❌ 没有有效内容")
+ return
+
+ # 最后一页如果包含表格，作为页眉页脚信息
+ meta_info = pages[-1] if "1"，和页眉页脚。如果存在分隔符那么逐页进行 word 的生成，如果存在页眉页脚，则给每页插入页眉页脚。如果都没有，那么直接按照 md 中块的顺序生成 word。
+
+```
+具体调用方式请查看 pdf_to_md_to_word_test.py
+```
+
+## pdf_to_json_to_latex.py
+
+- 作用：pdf 解析得到的内容转 latex
+- 详情：和 pad_to_json_word 的逻辑大体一致，重点需要关注公式/特殊符号的转义，对不同的 block 内容按照块类型进行符合 latex 中语法的编辑
+
+```
+具体调用方式请查看 pdf_to_json_to_latex_test.py
+```
+
+## pdf_to_md_to_latex.py
+
+- 作用 1：pdf 版面解析之后，补充页眉页脚，经过模型翻译成为信息完整 md，并转为 word
+
+- 作用 2：如果直接输入 Md ，也可以直接翻译为 word
+
+- 详情：和 pdf_to_md_to_word 的逻辑大体一致。
+
+```
+具体调用方式请查看 pdf_to_md_to_latex_test.py
+```
diff --git a/paddlex/inference/pipelines/pp_doctranslation/pipeline.py b/paddlex/inference/pipelines/pp_doctranslation/pipeline.py
index d95b100821..ee502bd197 100644
--- a/paddlex/inference/pipelines/pp_doctranslation/pipeline.py
+++ b/paddlex/inference/pipelines/pp_doctranslation/pipeline.py
@@ -355,6 +355,7 @@ def translate(
glossary: Dict = None,
llm_request_interval: float = 0.0,
chat_bot_config: Dict = None,
+ use_flags: bool = False,
**kwargs,
):
"""
@@ -395,7 +396,12 @@ def translate(
and ori_md_info_list[0].get("page_index") is not None
):
# for multi page pdf
- ori_md_info_list = [self.concatenate_markdown_pages(ori_md_info_list)]
+ if use_flags:
+ ori_md_info_list = [
+ self.concatenate_markdown_pages_with_flags(ori_md_info_list)
+ ]
+ else:
+ ori_md_info_list = [self.concatenate_markdown_pages(ori_md_info_list)]
if not isinstance(llm_request_interval, float):
llm_request_interval = float(llm_request_interval)
@@ -477,6 +483,82 @@ def translate_func(text):
}
)
+ # 将下一页的第一句话拼接至上一页， 保证翻译时语义的完整性，并对每页进行分页处理
+
+ def concatenate_markdown_pages_with_flags(self, markdown_list: list) -> tuple:
+
+ markdown_texts = ""
+ previous_page_last_element_paragraph_end_flag = True
+ PAGE_PLACEHOLDER = "1"
+ SENTENCE_ENDINGS = "，,。！？.!?"
+
+ if len(markdown_list) == 0:
+ raise ValueError("The length of markdown_list is zero.")
+
+ for res in markdown_list:
+ # Get the paragraph flags for the current page
+ page_first_element_paragraph_start_flag: bool = res[
+ "page_continuation_flags"
+ ][0]
+ page_last_element_paragraph_end_flag: bool = res["page_continuation_flags"][
+ 1
+ ]
+
+ # Determine whether to add a space or a newline
+ if (
+ not page_first_element_paragraph_start_flag
+ and not previous_page_last_element_paragraph_end_flag
+ ):
+ # 提取下一页的第一个句子
+ first_sentence_match = re.search(
+ rf"^[^{SENTENCE_ENDINGS}]*[{SENTENCE_ENDINGS}]?",
+ res["markdown_texts"],
+ )
+ if first_sentence_match:
+ front_sentence = first_sentence_match.group()
+ remaining_text = res["markdown_texts"][first_sentence_match.end() :]
+ else:
+ front_sentence = res["markdown_texts"]
+ remaining_text = ""
+
+ last_char_of_markdown = markdown_texts[-1] if markdown_texts else ""
+ first_char_of_handler = (
+ res["markdown_texts"][0] if res["markdown_texts"] else ""
+ )
+
+ # Check if the last character and the first character are Chinese characters
+ last_is_chinese_char = (
+ re.match(r"[\u4e00-\u9fff]", last_char_of_markdown)
+ if last_char_of_markdown
+ else False
+ )
+ first_is_chinese_char = (
+ re.match(r"[\u4e00-\u9fff]", first_char_of_handler)
+ if first_char_of_handler
+ else False
+ )
+
+ if not (last_is_chinese_char or first_is_chinese_char):
+ markdown_texts += " " + front_sentence
+ else:
+ markdown_texts += front_sentence
+ markdown_texts += f"\n\n{PAGE_PLACEHOLDER}\n\n" + remaining_text
+ else:
+ markdown_texts += f"\n\n{PAGE_PLACEHOLDER}\n\n" + res["markdown_texts"]
+ # markdown_texts += "\n\n" + res["markdown_texts"]
+ previous_page_last_element_paragraph_end_flag = (
+ page_last_element_paragraph_end_flag
+ )
+
+ concatenate_result = {
+ "input_path": markdown_list[0]["input_path"],
+ "page_index": None,
+ "page_continuation_flags": (True, True),
+ "markdown_texts": markdown_texts,
+ }
+
+ return MarkdownResult(concatenate_result)
+
def concatenate_markdown_pages(self, markdown_list: list) -> tuple:
"""
Concatenate Markdown content from multiple pages into a single document.
