元数据不完整条件下企业制度问答的条款有效性判断与选择性回答
Year: 2026
Type: preprint
Abstract: 企业制度的修订通常只涉及部分条款,发布日期、生效日期与适用部门又可能分散在不同文件中。此时,问答系统即使检索到最新通知,仍可能提前使用尚未生效的标准,或丢失没有被修订的旧条款。本文讨论元数据不完整条件下的条款有效性判断,将回答决策表述为对证据相容补全结果的稳定性检验:缺失信息的不同合理取值不改变答案时,可以回答;用户补充时间或部门即可消除歧义时,应请求澄清;关键制度依据缺失且无法由用户条件解决时,应暂缓回答。围绕该判据,建立条款关系建模、候选证据筛选和引用核验流程,并采用风险与覆盖率联合评价。在120题的评价设置下,条款级方案的一个操作点为回答90题,其中7题错误,对应75.00%的覆盖率和7.78%的选择性风险。与单纯提高拒答比例相比,该方案强调缺失信息与当前问题的关联性。方法仍依赖制度库覆盖和条款抽取质量,其跨领域有效性有待进一步检验。 English title: Clause Validity Assessment and Selective Answering for Enterprise Policy Question Answering with Incomplete Metadata English abstract: Revisions to enterprise policies often affect only some clauses, while publication dates, effective dates, and applicable departments may be distributed across different documents. Even after retrieving the latest notice, a question-answering system may apply a standard before it takes effect or lose older clauses that were never revised. This paper examines clause validity under incomplete metadata and formulates the answering decision as a stability test over evidence-compatible completions. An answer may be given when different reasonable values of missing information leave it unchanged; clarification is appropriate when the user can resolve ambiguity by supplying a date or department; and answering should be deferred when essential policy evidence is missing and user conditions cannot resolve the gap. The design combines clause-relation modeling, candidate-evidence filtering, and citation verification, evaluated jointly through risk and coverage. In the 120-question evaluation setting, one operating point of the clause-level approach answers 90 questions with 7 errors, corresponding to 75.00% coverage and 7.78% selective risk. Rather than merely increasing abstention, the approach emphasizes whether missing information matters to the current question. Its validity still depends on policy-corpus coverage and clause-extraction quality, and generalization across domains requires further investigation. Version 2 adds a full English translation PDF while retaining the original Chinese Word manuscript. AI assistance was used for translation and typesetting. No new experimental results were generated during translation. Original figure artwork is accompanied by English captions and label keys.
Author: ZHOU, YANG
Language: English
Subfield: Law
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7225448594