From Sentences to Words: LLM-Assisted Multilingual Alignment of Buddhist Texts

要解决的问题 Our Questions

本工作的两项辅助判断 Two aids we built first

既有对勘的接入：投影、直连、重组 Using alignments we already have: project, link, recombine

此法不另行发现平行，仅将既有对勘映射至 CBETA 句号。 Here we are not looking for new parallels. We simply place alignments that already exist onto CBETA’s sentence numbers.

1. 1 投影：将 BCBS、DEDU、法句及人工锚点映射至 CBETA 句号，作为金标准层。 Project. Take BCBS, DEDU, the Dharmapada tables, and a few hand-made anchors, and put them on CBETA sentence numbers. This is our gold standard.
2. 2 MITRA 直连：将已下载的梵汉／藏汉句对挂接经文。属机器句对，可回放至原 TSV。 Link. Place the downloaded Sanskrit–Chinese and Tibetan–Chinese pairs on the texts. These are machine pairs; we can always go back to the original TSV.
3. 3 重组：同梵异汉、同藏异汉、梵藏桥接（Bridge），由跨语句对推导汉—汉关系。V2 默认不计入正式覆盖。 Recombine. If two Chinese texts share the same Sanskrit, or the same Tibetan, or if Sanskrit and Tibetan can be bridged (Bridge), we infer a Chinese–Chinese link. Version 2 stays a candidate; we do not treat it as settled.

证据强度自上而下递减：人工投影最强，MITRA 直连次之，Bridge 为推导候选。 The evidence gets thinner as we go down: a human projection is the strongest, a MITRA link comes next, and Bridge is only an inference.

向量发现 Looking with vectors

本工作用以发现尚未登记、或尚未落实到句的平行候选。实施上分为三条路径。 This is how we look for parallels that no one has listed yet, or that have not yet been broken into sentences. We do it in three ways.

1. 1 对应经典向量匹配：在既有汉文关系范围内，以 MITRA-E 检索相似句。 Within a pair of Chinese texts we already know are related, we search for similar sentences with MITRA-E.
2. 2 全部汉文向量匹配：一部经对 CBETA 全库检索一次并缓存，编码器仍为 MITRA-E。 We also search one text against the whole of CBETA, once, and keep the result. Still MITRA-E.
3. 3 全部经典向量匹配：梵、巴、藏、汉逐句以千问 3.7 编码并建全库索引；逐句检索近邻，再聚合为可能平行的部与段落。 For a wider net we encode Sanskrit, Pali, Tibetan, and Chinese sentence by sentence with Qwen 3.7, put them in one index, look up near neighbours, and gather them into possible works and passages.
4. 4 指定两部之后，改以该两部句向量作一对一细对。汉—汉按翻译／其他大正藏／其他汉文／梵藏巴分层认经，以免注疏淹没异译。 When we have named two works, we compare those two on their own sentence vectors. On the Chinese side we sort works into layers—translations, the rest of the Taishō, other Chinese texts, then Sanskrit, Tibetan, and Pali—so a commentary does not push another translation off the list.

所得均为候选，尚未作语义审定。相似度分数不等于语义置信度。 None of this is a final judgement. A high score only means the sentences look alike; it does not mean they say the same thing.

模型审定 Asking a model to judge

审定对象并非全库近邻，而是第 3 节关系层中已登记的汉—汉范围。双方保持原序，模型仅返回句号。 We do not send the model everything that looks nearby. We only send Chinese–Chinese ranges already on the map in §3. Both sides stay in their original order; the model may return sentence numbers, and nothing else.

1. 1 双方按 CBETA v3 切句并编短编号。模型仅见编号与原文，不得改写，不得自造编号。 We split both texts with CBETA v3 and give each sentence a short number. The model sees the numbers and the wording. It must not rewrite the text, and it must not invent numbers.
2. 2 长经切为滑窗，连续提交大模型判定句级平行，再拼合为全书。 A long work is cut into overlapping windows, sent one after another, then put back together.
3. 3 先以人工金标准（法句、维摩等）验收，通过后再扩大范围。 We try this first on human gold standards—the Dharmapada, the Vimalakīrti, and a few others—and only then go further.
4. 4 结构校验已通过：A 侧每句恰好一次，B 编号合法。语义尚未审定。界面称为「AI匹配」。 The skeleton is in place: each sentence on side A appears once, and every number on side B is real. We have not yet read through the meaning. On the site this is called “AI matching”.