What Is Document Classification?什么是文档分类？
For the full topic map and the neighboring methods that support this workflow, continue with the unstructured data processing and document intelligence guide.
如需查看完整主题结构以及支撑本流程的相邻方法，请继续阅读非结构化数据处理与文档智能指南。
Document classification is the assignment of one or more predefined labels to a document, page or document bundle so another process can organize, prioritize, review or route it. The classifier may use native text, OCR text, layout, images, metadata and surrounding workflow context. Its output should include the predicted label or labels, confidence or decision evidence when available, model and taxonomy versions, and an explicit way to abstain.
文档分类是为文档、页面或文档包分配一个或多个预定义标签，使其他流程能够组织、排序、复核或路由它们。分类器可以使用原生文本、OCR文本、版面、图像、元数据和周边工作流上下文。输出应包含预测标签、可用时的置信度或决策证据、模型与分类体系版本，以及明确的拒识方式。
Automated document classification is useful when manual triage is repetitive, incoming volume is high, or downstream processing depends on document type. It is not the same as discovering topics without predefined classes, and it does not make every downstream decision safe. A label is a controlled prediction. Production systems must preserve uncertainty, route exceptions and keep the source available for verification.
当人工分拣重复、输入量较高，或下游处理依赖文档类型时，自动文档分类很有价值。它不同于在没有预定义类别时发现主题，也不会让每个下游决策自动变得安全。标签是一项受控预测；生产系统必须保留不确定性、路由异常，并保留源文档以供验证。
Quick answer: define the unit and allowed labels, create a versioned taxonomy and labeling guide, build a representative corpus, establish a simple baseline, evaluate every class on untouched data, set risk-based accept/review/reject thresholds, and monitor the full decision path after launch.
快速回答：定义分类单元和允许标签，建立版本化分类体系与标注指南，准备代表性语料，先建立简单基线，在未参与训练的数据上逐类评估，再按风险设置接受、复核与拒绝阈值，并在上线后监控完整决策路径。
Separate Classification from OCR, Extraction, Clustering and Routing区分分类、OCR、提取、聚类与路由
| Task任务 | Primary output主要输出 | Critical boundary关键边界 |
|---|---|---|
| OCR | Recognized text and coordinates识别文本与坐标 | Supplies signals; does not decide the business class提供信号，但不决定业务类别 |
| Extraction提取 | Fields, entities, tables or relationships字段、实体、表格或关系 | Answers what is inside, not which controlled label applies回答文档中有什么，而不是应分配哪个受控标签 |
| Clustering聚类 | Groups based on similarity基于相似性的分组 | May reveal structure but does not enforce approved classes可以发现结构，但不执行已批准类别 |
| Classification分类 | One or more controlled labels一个或多个受控标签 | Must represent unknown and ambiguous inputs必须表示未知与模糊输入 |
| Routing路由 | A destination or workflow action目标位置或工作流动作 | Consumes labels plus policy, permissions and context消费标签，并结合策略、权限与上下文 |
Keeping these contracts separate makes failures diagnosable. If a scan is unreadable, better routing rules cannot repair it. If the taxonomy overlaps, a more powerful model cannot create a stable truth. If classification is correct but a policy sends the label to the wrong queue, the routing layer—not the classifier—failed.
保持这些契约分离，才能诊断失败。如果扫描件不可读，更好的路由规则无法修复它；如果分类体系重叠，更强的模型也无法创造稳定真值；如果分类正确但策略把标签发送到错误队列，失败的是路由层而不是分类器。
Build the Taxonomy Before You Select a Model先建立分类体系，再选择模型
- Name the decision.命名决策。 State what the label controls and who relies on it. A category that has no consumer may not be worth automating.说明标签控制什么、谁依赖它。没有消费者的类别可能不值得自动化。
- Write inclusion and exclusion rules.编写包含与排除规则。 Give positive, boundary and counterexamples. Avoid definitions that merely repeat the label name.提供正例、边界例和反例，避免只重复标签名称的定义。
- Assign ownership.分配负责人。 A business owner approves meaning; data and model owners implement it; operations owns exceptions.业务负责人批准含义；数据与模型负责人实施；运营负责人处理异常。
- Version every change.为每次变化建立版本。 Record additions, merges, splits, deprecations, effective dates and migration rules.记录新增、合并、拆分、弃用、生效日期与迁移规则。
Test the taxonomy with double annotation before training. Low inter-annotator agreement often signals ambiguous policy, missing context or an unsuitable unit. Resolve disagreements through the guide and an adjudication process; do not hide them by majority vote alone.
训练前用双人标注测试分类体系。低标注者一致性通常表明策略模糊、上下文缺失或分类单元不合适。应通过指南与裁决流程解决分歧，不能只用多数票掩盖问题。
Prepare Representative Document Classification Training Data准备有代表性的文档分类训练数据
Collect only documents you are permitted to use. Preserve provenance, source system, time range, language, template family, page count, scan quality and known duplicates. Sample normal traffic plus rare, high-impact, ambiguous, unknown and damaged inputs. Masking or synthetic examples can help with privacy or scarcity, but they must be tracked separately because they may not represent production variation.
只收集获准使用的文档。保留来源、源系统、时间范围、语言、模板家族、页数、扫描质量和已知重复项。样本应覆盖正常流量，以及稀有、高影响、模糊、未知和损坏输入。脱敏或合成样本可以缓解隐私与稀缺问题，但必须单独跟踪，因为它们可能不代表生产变化。
Each labeled record should preserve:每条标注记录应保留：
- stable document ID and content hash;稳定文档ID与内容哈希；
- taxonomy and guideline version;分类体系与指南版本；
- labeler, timestamp and permitted evidence;标注者、时间戳与允许使用的证据；
- assigned labels, uncertainty and adjudication status;已分配标签、不确定性与裁决状态；
- source and grouping keys used to prevent leakage.用于防止泄漏的来源与分组键。
No universal training count proves readiness. Service minimums only allow a job to run. Use learning curves and class-level error analysis to decide whether more examples help. If added data does not improve a class, revisit label definition, signal availability, annotation quality or method choice.
不存在能证明就绪的通用训练样本数。服务最低数量只代表任务可以运行。应通过学习曲线与逐类错误分析判断增加样本是否有帮助；如果新增数据不能改善某个类别，应重新检查标签定义、信号可用性、标注质量或方法选择。
Compare Rules, Machine Learning, Embeddings and Generative Models比较规则、机器学习、嵌入与生成式模型
| Method方法 | Strength优势 | Limitation局限 | Good starting case适合起点 |
|---|---|---|---|
| Deterministic rules确定性规则 | Traceable, fast and easy to constrain可追溯、快速、易约束 | Brittle across wording and layout change对措辞与版面变化脆弱 | Stable identifiers or mandated phrases稳定标识符或规定短语 |
| Classical text ML传统文本机器学习 | Strong baseline, efficient and inspectable基线强、效率高、可检查 | Needs representative labels; may miss layout需要代表性标签，可能忽略版面 | Clear classes with usable text类别清晰且文本可用 |
| Embedding similarity嵌入相似度 | Flexible retrieval and few-example prototypes灵活检索与少样本原型 | Similarity is not calibrated business confidence相似度不是校准后的业务置信度 | Candidate generation or semantic categories候选生成或语义类别 |
| Specialized document model专用文档模型 | Can combine language, layout and visual signals可组合语言、版面与视觉信号 | More training, serving and monitoring complexity训练、服务与监控更复杂 | Visually distinct document families视觉差异明显的文档家族 |
| Generative zero/few-shot生成式零/少样本 | Rapid experiments and rich label instructions快速实验，可使用丰富标签指令 | Output variability, cost, privacy and version drift输出波动、成本、隐私与版本漂移 | Taxonomy exploration or reviewed low volume分类体系探索或有复核的低流量任务 |
| Hybrid混合方法 | Combines hard policy with learned coverage组合硬性策略与学习覆盖 | Needs precedence and conflict rules需要优先级与冲突规则 | Regulated or heterogeneous intake受监管或异构入口 |
Build at least one interpretable baseline before adopting a complex model. A baseline reveals whether the labels are learnable from permitted signals and gives a meaningful comparison for latency, cost, review workload and maintainability. A hybrid design often lets exact policy rules handle known anchors while a model covers variable language and layout.
采用复杂模型前至少建立一个可解释基线。基线能揭示标签是否可从允许的信号中学习，并为延迟、成本、复核工作量和可维护性提供有意义的比较。混合设计通常可以让精确策略规则处理已知锚点，同时由模型覆盖变化的语言与版面。
Split Training, Validation and Test Data Without Leakage在无泄漏条件下划分训练、验证与测试数据
Training data fits model parameters. Validation data selects features, prompts, thresholds and model versions. The test set estimates final performance and should remain untouched until the design is frozen. Never tune repeatedly on the test set and still call it independent.
训练数据用于拟合模型参数；验证数据用于选择特征、提示词、阈值和模型版本；测试集估计最终表现，在设计冻结前应保持未触碰。不能反复在测试集上调优后仍称其独立。
Leakage traps: near-duplicate documents crossing splits; pages from one bundle placed in different sets; the same customer or template family appearing on both sides; labels encoded in filenames or folder paths; post-decision metadata unavailable at inference time; future documents used to predict the past.
泄漏陷阱：近重复文档跨越不同集合；同一文档包的页面被拆到不同集合；同一客户或模板家族同时出现在两侧；文件名或文件夹路径编码标签；推理时不可用的事后元数据；用未来文档预测过去。
Group by the strongest dependency key and use time-based holdouts when production change over time matters. Keep a challenge set for rare formats, poor scans, short pages, mixed languages and out-of-scope inputs. It complements—rather than replaces—the representative test set.
应按最强依赖键分组；当生产环境随时间变化很重要时，使用按时间留出的测试。另建挑战集覆盖稀有格式、低质量扫描、短页、混合语言和范围外输入。挑战集是代表性测试集的补充，而不是替代。
Evaluate Document Classification by Class and Risk Slice按类别与风险切片评估文档分类
Overall accuracy can hide failure on rare or consequential classes. Start with a confusion matrix, which shows which true classes become which predicted classes. Report precision, recall, F1 and support for every class. Macro averages weight classes equally; weighted averages reflect class frequency. For multi-label tasks, also inspect per-label confusion and whether complete label sets are correct.
总体准确率可能掩盖稀有或高影响类别的失败。先使用混淆矩阵查看真实类别被预测成哪些类别，再报告每个类别的精确率、召回率、F1和支持数。宏平均对各类别等权；加权平均反映类别频率。多标签任务还要检查逐标签混淆以及完整标签集合是否正确。
Of items predicted as a class, how many truly belong? Emphasize it when false acceptance is costly.
被预测为某类别的项目中，有多少真实属于该类？误接受代价高时应重点关注。
Of true items in a class, how many were found? Emphasize it when missing a class is costly.
某类别的真实项目中，有多少被发现？漏掉该类代价高时应重点关注。
Harmonic balance of precision and recall; useful, but it does not encode business cost.
精确率与召回率的调和平衡，有用但不编码业务成本。
Review rate, correction time, queue delay, failed files, downstream reversals and incident severity.
复核率、更正时间、队列延迟、失败文件、下游撤销与事件严重度。
Slice results by source, language, template, page count, image quality, time period and protected or high-risk group where lawful and relevant. A single aggregate can pass while one source fails. The official scikit-learn classification report documentation explains per-class precision, recall, F1 and macro or weighted averages; its confusion matrix documentation defines the true-versus-predicted layout.
应按来源、语言、模板、页数、图像质量、时间段以及在合法且相关时的受保护或高风险群体切片。单一汇总指标可能通过，但某个来源仍会失败。官方scikit-learn分类报告文档解释了逐类精确率、召回率、F1以及宏平均和加权平均；其混淆矩阵文档定义了真实类别与预测类别的排列方式。
Set Confidence Thresholds, Unknown Classes and Abstention设置置信阈值、未知类别与拒识
A score is not automatically a probability, and a probability is not automatically a business decision. Check calibration: among predictions near a stated confidence, does roughly that proportion prove correct on representative data? Then translate scores into policy bands. High-confidence low-impact cases may be accepted; intermediate cases reviewed; low-confidence, conflicting or out-of-scope cases rejected or quarantined.
分数不自动等于概率，概率也不自动等于业务决策。应检查校准：在代表性数据上，接近某个置信水平的预测是否大致有相同比例正确？然后把分数转换为策略区间。高置信且低影响案例可以接受；中间案例复核；低置信、冲突或范围外案例应拒绝或隔离。
Use class-specific thresholds when error costs differ. The official Google Cloud Document AI evaluation guide documents how a higher confidence threshold generally raises precision while reducing recall. Do not copy its automatically optimized threshold into another system: select your own threshold against validation data, review capacity and downstream loss.
当错误成本不同时，应使用逐类阈值。官方Google Cloud Document AI评估指南说明，提高置信阈值通常会提升精确率但降低召回率。不要把其自动优化阈值照搬到其他系统；应根据验证数据、复核能力和下游损失选择自己的阈值。
Hypothetical Example: Classify a Mixed Finance Intake假设示例：分类混合财务入口
Assumption example—not a customer result: a finance team receives invoices, credit notes, bank statements and unrelated attachments. It needs one document-type label, optional sensitivity tags and an unknown path before extraction.
假设示例，并非客户结果：某财务团队接收发票、贷项通知单、银行对账单和无关附件；在提取前，它需要一个文档类型标签、可选敏感标签以及未知路径。
- Define outputs.定义输出。 The type is single-label; sensitivity is multi-label; unreadable and unrelated items must abstain.类型采用单标签；敏感属性采用多标签；不可读和无关项目必须拒识。
- Build the corpus.建立语料。 Group splits by supplier and template family, then hold out the latest period to test change.按供应商与模板家族分组划分，再留出最近时间段测试变化。
- Create baselines.建立基线。 Exact tax identifiers and stable phrases form rules; a text-layout model covers variable evidence.税务标识和稳定短语形成规则；文本版面模型覆盖变化证据。
- Tune decisions.调节决策。 Low-confidence type, rule/model conflict and any high-impact sensitivity tag enter review.低置信类型、规则与模型冲突以及任何高影响敏感标签进入复核。
- Accept end to end.端到端验收。 Measure per-class errors, review labor, queue delay and whether approved labels invoke the correct extractor and destination.衡量逐类错误、复核工作量、队列延迟，以及获批标签是否调用正确提取器与目标位置。
The example deliberately provides no fabricated accuracy, throughput or savings. Those values must come from the organization’s representative corpus, deployed configuration and measured workflow.
本示例刻意不提供虚构的准确率、吞吐量或节省数字。这些数值必须来自组织自己的代表性语料、部署配置与实测工作流。
Build a Traceable Production Classification Path建立可追溯的生产分类路径
- Intake and validate.接收与验证。 Authenticate the source, scan where required, validate format and preserve the original plus a stable identifier.验证来源，在需要时进行安全扫描，验证格式，并保留原件与稳定标识符。
- Prepare signals.准备信号。 Use native parsing or OCR, layout and permitted metadata. Record extraction failures instead of silently classifying empty input.使用原生解析或OCR、版面与获准元数据。记录提取失败，不要静默分类空输入。
- Infer and apply policy.推理并应用策略。 Store model, rule, prompt, taxonomy and threshold versions with candidate labels and evidence.随候选标签和证据保存模型、规则、提示词、分类体系与阈值版本。
- Review exceptions.复核异常。 Queue ambiguous, unknown, conflicting or high-impact cases with reason codes and controlled overrides.将模糊、未知、冲突或高影响案例按原因码排队，并受控覆盖。
- Route approved outputs.路由获批输出。 Only approved labels trigger extraction, retention, access or downstream analysis. Make actions idempotent and reversible where feasible.只有获批标签才能触发提取、保留、访问或下游分析；在可行时让动作幂等且可撤销。
Microsoft documents that its custom classifier can identify page ranges and, depending on version and settings, split multi-document files. Treat that as a product-specific behavior to test, not a universal feature. Review the current Azure Document Intelligence custom classifier documentation before relying on defaults.
微软文档说明其自定义分类器可以识别页码范围，并可根据版本与设置拆分多文档文件。这是需要测试的产品特定行为，不是所有分类器的通用功能。在依赖默认行为前，应查阅当前Azure Document Intelligence自定义分类器文档。
Monitor Data Drift, Taxonomy Change and Decision Quality监控数据漂移、分类体系变化与决策质量
Production monitoring needs more than latency and availability. Track input volume, source mix, language, template and quality distributions; predicted class and confidence distributions; unknown, review and override rates; extractor or destination rejection; and delayed ground-truth metrics when labels arrive. Compare current slices with the approved baseline.
生产监控不能只有延迟和可用性。应跟踪输入量、来源组合、语言、模板和质量分布；预测类别与置信度分布；未知、复核和覆盖率；提取器或目标位置的拒绝；以及真实标签延迟到达后的指标，并把当前切片与已批准基线比较。
Create triggers and responses before launch: sample a changed source, increase review, disable one route, roll back a model, remap deprecated labels or pause automation. The NIST AI Risk Management Framework provides a voluntary structure for governing, mapping, measuring and managing AI risk. Apply its ideas in proportion to impact and applicable obligations.
上线前就要定义触发条件与响应：抽样变化来源、提高复核比例、禁用某条路由、回滚模型、重映射弃用标签或暂停自动化。NIST人工智能风险管理框架为治理、映射、衡量和管理AI风险提供自愿结构；应按影响程度和适用义务采用。
Protect Documents, Labels and Model Evidence保护文档、标签与模型证据
- Minimize access: separate source, annotation, training, review and production permissions; restrict supplier access and log exports.最小化访问：分离源文档、标注、训练、复核与生产权限；限制供应商访问并记录导出。
- Control retention: define retention and deletion for originals, derivatives, prompts, labels, embeddings, logs and reviewer notes.控制保留：定义原件、副本、提示词、标签、嵌入、日志与复核备注的保留和删除。
- Test attacks and contamination: malformed files, embedded instructions, adversarial text, unsupported languages and poisoned corrections should fail safely.测试攻击与污染：畸形文件、嵌入指令、对抗文本、不支持语言和被污染更正都应安全失败。
- Preserve accountability: store who approved taxonomy, data, threshold and release changes, with rollback material.保留问责：记录谁批准了分类体系、数据、阈值与发布变化，并保存回滚材料。
A confidentiality label can itself be sensitive. Do not expose labels, rationales or class lists more widely than the source. When labels trigger access or retention, require policy validation rather than letting the classifier make the final authorization decision.
机密标签本身也可能敏感。标签、理由或类别清单的暴露范围不应大于源文档。当标签触发访问或保留时，应进行策略验证，而不能让分类器直接作出最终授权决定。
Recognize Common Document Classification Failures识别常见文档分类失败
Reviewers and models cannot consistently distinguish labels. Rewrite boundaries or separate independent dimensions.
复核人员和模型无法稳定区分标签。应重写边界或拆分独立维度。
Overall accuracy looks good while rare classes fail. Inspect per-class metrics and macro averages.
总体准确率良好但稀有类别失败。应检查逐类指标与宏平均。
The model confidently classifies missing or corrupt text. Validate signals and expose no-content outcomes.
模型对缺失或损坏文本作出高置信分类。应验证信号并暴露无内容结果。
Novel documents are routed as the nearest class. Add unknown detection, abstention and sampling.
新文档被路由为最接近类别。应增加未知检测、拒识与抽样。
Duplicates, filenames or future metadata inflate tests. Group splits and audit features available at decision time.
重复项、文件名或未来元数据抬高测试结果。应分组划分并审计决策时可用特征。
Templates and policy change without retraining or remapping. Assign triggers, owners and rollback paths.
模板和策略变化却不重训或重映射。应分配触发条件、负责人和回滚路径。
Analyze Approved Classified Documents with Related Sources把获批分类文档与相关来源共同分析
Before continuing, prepare rights-approved documents, their reviewed labels, stable IDs and any related database, audio or video sources you are permitted to analyze. Keep the classifier’s confidence, model version and review outcome as provenance rather than presenting a prediction as unquestioned fact.
继续之前，请准备权利获批的文档、已复核标签、稳定ID，以及获准分析的相关数据库、音频或视频来源。把分类器置信度、模型版本与复核结果作为来源记录保留，不要把预测呈现为无需质疑的事实。
Use InfiniSynapse to analyze approved documents alongside connected sources. It does not train the classifier, assign labels, manage the review queue or route records.
使用InfiniSynapse把获批文档与已连接来源共同分析。它不训练分类器、不分配标签、不管理复核队列，也不路由档案。
For adjacent workflow context, browse the InfiniSynapse data and document analysis articles and the current InfiniSynapse tools directory. Use only a destination whose visible capability matches the task; a conversion or analysis tool should not be treated as a classifier.
如需相邻工作流背景，可浏览InfiniSynapse数据与文档分析文章和当前InfiniSynapse工具目录。只使用可见能力与任务匹配的入口；不能把转换或分析工具当作文档分类器。
Document Classification FAQ文档分类常见问题
What is document classification?
什么是文档分类？
Document classification assigns one or more predefined labels to a document or page so a downstream system can organize, prioritize, review or route it. The task may use deterministic rules, supervised models, embeddings, generative models or a hybrid. A production design also needs an explicit unknown or review outcome; forcing every input into a known class hides risk.
文档分类为文档或页面分配一个或多个预定义标签，使下游系统能够组织、排序、复核或路由。任务可以使用确定性规则、监督模型、嵌入、生成式模型或混合方法。生产设计还需要明确的未知或复核结果；把所有输入强行放入已知类别会掩盖风险。
What is the difference between document classification and OCR?
文档分类与OCR有什么区别？
OCR converts pixels into text and may recover layout coordinates. Classification uses available signals—native text, OCR text, layout, images, metadata or context—to assign labels. OCR can be an input to a classifier, but it does not decide the business class, and a classifier does not necessarily extract the fields inside a document.
OCR把像素转换为文本，也可能恢复版面坐标。分类使用可用信号——原生文本、OCR文本、版面、图像、元数据或上下文——来分配标签。OCR可以是分类器的输入，但不决定业务类别；分类器也不一定提取文档内部字段。
Should document classification use single-label or multi-label mode?
文档分类应该使用单标签还是多标签模式？
Use single-label classification when exactly one mutually exclusive class must describe each unit, such as invoice, contract or statement. Use multi-label classification when several independent tags may be true at once, such as finance, confidential and urgent. Use a hierarchy when parent-child consistency matters. Decide the task before labeling because the target representation and evaluation differ.
当每个分类单元必须由一个互斥类别描述时使用单标签，例如发票、合同或对账单；当财务、机密、紧急等多个独立标签可同时成立时使用多标签；父子一致性重要时使用层级结构。标注前必须先决定任务，因为目标表示和评估方式不同。
How much training data does a document classifier need?
文档分类器需要多少训练数据？
There is no universal sample count. Need depends on class overlap, layout and language diversity, input quality, label consistency, model family and required error tolerance. Start with a representative labeled corpus and a simple baseline, inspect learning curves and class-level errors, then add examples where evidence is weak. Vendor minimums are service constraints, not proof of production readiness.
不存在通用样本数。所需数量取决于类别重叠、版面和语言多样性、输入质量、标签一致性、模型家族与可容忍错误。先用代表性标注语料和简单基线开始，检查学习曲线与逐类错误，再在证据薄弱处补样。供应商最低数量是服务约束，不是生产就绪证明。
How should document classification accuracy be evaluated?
如何评估文档分类准确性？
Keep an independent, representative test set and report the confusion matrix plus precision, recall, F1 and support for every class. Add macro averages so rare classes remain visible, and inspect performance by source, language, layout, quality and risk. For multi-label tasks, evaluate each label and the complete label set. Measure review rate and downstream error cost as well as model metrics.
保留独立且有代表性的测试集，并报告混淆矩阵以及每个类别的精确率、召回率、F1和支持数。增加宏平均以保留稀有类别的可见性，并按来源、语言、版面、质量和风险切片检查。多标签任务要评估每个标签及完整标签集合。除模型指标外，还要衡量复核率和下游错误成本。
What confidence threshold should a document classifier use?
文档分类器应该使用什么置信阈值？
Choose thresholds from validation data and business loss, not from a universal percentage. A false acceptance may be costlier than a review for one class, while a missed urgent document may be costlier for another. Set class-specific accept, review and reject bands where appropriate, verify calibration, and retest thresholds after model, taxonomy or traffic changes.
应从验证数据和业务损失中选择阈值，而不是套用统一百分比。对某个类别，误接受可能比人工复核代价更高；对另一个类别，漏掉紧急文档可能更昂贵。可按类别设置接受、复核和拒绝区间，验证校准，并在模型、分类体系或流量变化后重新测试。
When is human review required for document classification?
文档分类何时需要人工复核？
Route low-confidence, conflicting, unknown, out-of-scope and high-impact cases to trained reviewers. Review may also be required for new sources, changed templates, sensitive decisions or regulated records. Give reviewers the source, predicted labels, confidence and permitted evidence; capture corrections with reason codes, but do not automatically treat every correction as trustworthy training data.
应把低置信、冲突、未知、范围外和高影响案例交给受训复核人员。新来源、模板变化、敏感决策或受监管记录也可能需要复核。向复核人员提供源文档、预测标签、置信度与允许使用的证据；用原因码记录更正，但不要自动把每次更正都当作可信训练数据。
Can InfiniSynapse replace a document classifier?
InfiniSynapse能替代文档分类器吗？
No. InfiniSynapse is publicly presented as a multi-source and multimodal analysis tool across databases, documents, audio and video. It can be a downstream analysis destination for approved documents and classification outputs, but it should not be described as a labeling tool, classifier-training platform, confidence review queue, routing engine, document management system or records system.
不能。InfiniSynapse公开定位为跨数据库、文档、音频和视频的多源多模态分析工具。它可以作为获批文档和分类结果的下游分析入口，但不能描述为标注工具、分类器训练平台、置信度复核队列、路由引擎、文档管理系统或档案系统。
Official Sources and Further Reading权威来源与延伸阅读
- Google Cloud Document AI: create, use and manage a custom document classifierGoogle Cloud Document AI：创建、使用和管理自定义文档分类器
- Google Cloud Document AI: evaluate processor performance and thresholdsGoogle Cloud Document AI：评估处理器表现与阈值
- Microsoft Azure Document Intelligence: custom classification modelMicrosoft Azure Document Intelligence：自定义分类模型
- AWS Comprehend: custom document classificationAWS Comprehend：自定义文档分类
- scikit-learn: classification report metricsscikit-learn：分类报告指标
- NIST Artificial Intelligence Risk Management FrameworkNIST人工智能风险管理框架
Vendor limits, features and defaults change. Verify the current documentation, region, service version, contract and your own measured results before implementation.
供应商限制、功能与默认值会变化。实施前应核对当前文档、地区、服务版本、合同与自己的实测结果。