The Interpretation of Nothing: 14 Million Chinese Judgments Cite a Rule Whose Name Was Stripped

Chinese judgments cite Supreme People’s Court interpretations by their full titles, and those titles nest the statute they interpret inside a second pair of brackets: 《最高人民法院关于适用〈中华人民共和国民事诉讼法〉的解释》, the Interpretation on the Application of the Civil Procedure Law. In a very large share of judgment texts the inner part is simply gone. What remains is 《最高人民法院关于适用的解释》: an interpretation on the application of nothing.

We noticed this while measuring statute citation drift for the Civil Code cutover and gave it one section there. It deserves its own page, because any team that extracts statute citations from Chinese judgments, builds a citation graph, tags records by legal regime or evaluates retrieval on “which interpretation does this ruling apply” is keying on exactly the characters that are missing. Three findings, all measured on 3 October 2026 against our own index of 160,291,678 records:

The rest of this page shows the counts, where the stripped citations live, which other interpretations lose their statute name the same way, what the variants look like, and a repair recipe that flags rather than rewrites.

What a nested title is, and what is left when it breaks

Most SPC interpretations follow one naming pattern: 关于适用《X》的解释 (“Interpretation on the application of X”), 关于适用《X》若干问题的解释 (“Interpretation on several issues in the application of X”), or a variant that adds a chapter or a topic: 关于适用《X》时间效力的若干规定, 关于适用《X》有关担保制度的解释, 关于适用《X》婚姻家庭编的解释. When the whole title is itself quoted in a judgment, Chinese typography demands single angle brackets for the inner title: 《最高人民法院关于适用〈中华人民共和国民事诉讼法〉的解释》.

In the broken records, the inner title and its brackets are gone and the two halves of the outer title are joined: 关于适用的解释, 关于适用若干问题的解释, 关于适用时间效力的若干规定. These strings are grammatical enough that nothing downstream raises an error. A regex that looks for 民事诉讼法 inside a citation finds nothing; a parser that maps titles to a statute table finds a title that is not in the table; a reader who knows the context knows exactly what was meant.

Which statute was meant is usually recoverable, but not from the title. That is the point of this page.

How we measured

• Index. Elasticsearch, 160,291,678 PRC judgment records, body analysed with ik_max_word. All counts are match_phrase on the body with the analyzer forced to ik_max_word; we checked with _analyze that 关于适用的解释 tokenises as four consecutive tokens (关于 / 适用 / 的 / 解释), so the phrase matches the literal string and nothing else. We use ik_max_word at query time because the default ik_smart search analyzer returns zero for many phrases.
• Full titles. The complete inner title spelled out without brackets (关于适用中华人民共和国民事诉讼法的解释). Brackets are not tokens, so this phrase matches every bracket variant of the full form.
• Stripped-only. Records that match the stripped phrase and do not match any of the relevant full titles anywhere in the body. A record that cites the interpretation twice, once each way, counts as resolved, not stripped.
• Sample. 2,000 stripped records drawn with random_score, seed 20261003 on _seq_no, and boost_mode: replace. We first drew the sample without boost_mode and it over-represented civil judgments by two to one against the population; the default multiplies the random score by relevance, so short documents that mention the phrase often get drawn more. The sample reported here matches the population: 52.0% enforcement cases against 53.3% in the full count.
• Dates. Year buckets use the indexed year field. For some batches that field is a publication date rather than the decision date; we use it here only to show that the loss is present every year, not to time anything.
• No exclusions for the headline counts. One batch of 4,414,931 records has no text and contributes 24 citing records; it is shown in the batch table for completeness.

The procedural interpretations: 42% of citations have no statute

The stripped string 关于适用的解释 is the tail shared by three interpretations: the Civil Procedure Law interpretation (2015, 552 articles), the Criminal Procedure Law interpretation (2012, 548 articles; 2021, 655 articles) and the Administrative Litigation Law interpretation (2018, 163 articles). All three nest the statute name, so all three break into the same string.

Put differently: the Civil Procedure Law interpretation is the most-cited rule in the corpus, and a parser keyed on its name sees a little over half of its citations. The 637,791 records that carry both forms are the only ones where the title can be repaired from the same document.

The same loss, other titles

Any interpretation whose title nests a statute name loses it the same way. We checked the common tails. Each stripped tail collapses several interpretations into one string, so the full-title column lists the main members.

The time-effect rules are the serious case for anyone splitting the corpus into pre-Code and Civil Code regimes. Of the 4,056,401 records that cite them in any form, 1,837,085 (45.3%) contain only the stripped form and never mention 民法典 by name in that title. A rule of the form “cites 中华人民共和国民法典 somewhere” still catches most of them through other citations, but a rule that keys on the time-effect title does not. We covered the regime-tagging consequence in the cutover page.

The 若干问题的解释 tail is the ambiguous one. Five different pre-Code interpretations share it, three of them (Marriage, Guarantee, Contract) with hundreds of thousands of citations each, and 790,753 records carry it without any full title. Case type does not separate them: they are all civil. The article number and the words around it do, most of the time.

Not one batch, not one year

When a defect looks like this, the first suspicion is a single collection pass that mangled brackets. The numbers do not support that. Our index is assembled from fourteen ingest batches collected over several years by more than one pipeline; every batch that has text sits between 34% and 50% stripped-only.

By year the picture is the same once the interpretation exists. The Civil Procedure Law interpretation took effect in February 2015, so 2015 is a partial year with few citations and most of them intact; from 2017 onwards the stripped share sits between 40% and 50% and has not improved in the most recent records.

Two batches collected years apart by different pipelines, ws_new and 2021, sit at 48.0% and 50.1%. The oldest batch, ws, is lowest at 33.8%, and the newest, 2025, is at 43.7%. We do not know the cause. Within the 107 sampled records that contain both a stripped and an intact nested title, the intact ones use three different inner-bracket characters (〈 in 72 records, 《 in 27, fullwidth ＜ in 16), and in our earlier reading of 200 records an ASCII < also survived in some. Whatever removed the inner titles did not remove one bracket style consistently, so we do not claim it was tag stripping, and we do not claim it happened on our side or the source’s. We flag it and move on.

Where the stripped citations live

By case type, the stripped records are mostly enforcement rulings, which is where the Civil Procedure Law interpretation is cited most mechanically.

The last row is a reminder that the case-type field itself carries several vocabularies; 1,134,670 records say 裁定 (ruling) and 274,133 say 判决 (judgment) where a case type should be.

The article numbers that follow the stripped title confirm what is being cited. In the 2,000-record sample, 1,968 records (98.4%) had an article number immediately after the title, 2,214 citations in all. The most frequent, in the numbering of the 2015 Civil Procedure Law interpretation (which the 2022 revision kept at 552 articles):

Article 519 alone accounts for a third of the stripped citations in the sample. Every one of those is a ruling that ends an enforcement round for want of assets, and every one of them is invisible to a citation extractor keyed on the statute name. If your product counts how often courts terminate enforcement, or builds a graph of which rules enforcement courts apply, this single string is most of your data.

Three interpretations, one string: how far case type and article number get you

Because the civil, criminal and administrative procedure interpretations all collapse to 关于适用的解释, resolving a stripped citation means deciding among three candidates. We applied two cheap signals to the sample: the case type, and the article number against each interpretation’s article count (552 civil, 655 or 548 criminal, 163 administrative).

So case type alone settles 88.7% of the sample, and the article number is consistent with case type in every stratum we could check. The remaining 11.4% are the records whose case-type field holds a document type, and for those the article number does the work: 519 cannot be the administrative interpretation, and in a ruling that says 终结本次执行程序 it is not the criminal one either. A dictionary of the fifty most-cited articles with their interpretation and a rule for the case-type vocabularies covers nearly all of it. What it cannot do is tell you which version of the Criminal Procedure Law interpretation a 2019 ruling meant, since the numbering changed in 2021; for that you need the decision date, with its own caveats.

What the variants look like

Of the 2,000 sampled records, 1,988 contain the stripped string exactly. The other twelve matched the phrase through punctuation the analyzer ignores, and they show the loss in the act:

• 《最高人民法院关于适用》的解释第五百一十九条: the outer closing bracket moved to where the inner title used to end.
• 《最高人民法院关于适用，的解释》第五百一十九条 and 关于适用,的解释: a comma left behind where the inner title sat.
• 依照《最高人民法院关于适用 — and the body ends. Three of the twelve stop inside the citation, which is also where the operative part of the ruling would have started. A truncation that lands on a citation is still a truncation; we check for it separately.

None of these are false positives for the purpose of this page: they are all citations of a procedural interpretation with the statute name missing. They are false negatives for anyone matching the exact stripped string, which is why the repair should work on the tail (的解释 preceded by 关于适用 within a few characters) rather than the exact string.

What this does to a legal AI pipeline

• Citation extraction undercounts, unevenly. The Civil Procedure Law interpretation loses 42% of its citations; the Civil Code time-effect rules lose 45%; the pre-Code 若干问题的解释 family loses an amount that depends on which member you ask about. A citation-frequency ranking of interpretations built from the full titles alone is wrong in rank order, not just in level.
• Enforcement disappears from citation statistics. 53% of stripped records are enforcement rulings. Enforcement is the largest procedural category in the corpus, and its most-cited rule is cited in the stripped form more often than not.
• Regime tagging misfires. A record whose only Civil Code citation is the stripped time-effect title never mentions 民法典 in that title. Rules of the form “cites the Civil Code” built on title matching misclassify it as an old-law judgment.
• Retrieval evaluation is biased. A query for “rulings applying the Civil Procedure Law interpretation” that matches on the statute name has a recall ceiling near 58% on this corpus, and the misses are concentrated in enforcement. An evaluation set built from the hits inherits the bias.
• Training data teaches the wrong citation. A model fine-tuned on judgment text learns that 《最高人民法院关于适用的解释》 is how courts cite the rule, because fourteen million examples say so. It will then generate stripped titles itself.

Repair as a flag, not a rewrite

We do not rewrite judgment text. The record is what the court published, as collected; a corrected title written back into the body is a fabrication with good intentions. The repair belongs in fields next to the text:

1. Detect by tail. Match 关于适用 followed within a few characters (allowing 》, ，, a comma or a space) by one of the known tails: 的解释, 若干问题的解释, 若干问题的规定, 时间效力的若干规定, 有关担保制度的解释, 婚姻家庭编的解释, 总则编若干问题的解释. Record citation_title_stripped = true.
2. Resolve by candidates. For each tail, list the interpretations that produce it. Score candidates by case type (after normalising the vocabularies), by article number against each candidate’s article count, and by nearby cue words (执行, 终结本次执行程序, 举证, 被告人, 行政行为). Store resolved_interpretation and a confidence, and leave it empty when two candidates remain.
3. Prefer same-document evidence. In 637,791 records the full title appears elsewhere in the same text. Use it before any dictionary.
4. Keep both counts. Report citation statistics as “full-title” and “full-title plus resolved stripped” with the share resolved. Never publish the first as if it were the second.
5. Version by date where numbering changed. For the Criminal Procedure Law interpretation, attach the 2012 or 2021 numbering by decision date, and flag records whose date field is a publication date.

How to reproduce this on any delivery

# Elasticsearch, body analysed with ik_max_word.
# 1. Confirm the stripped string tokenises as consecutive tokens
POST judgments/_analyze  {"analyzer":"ik_max_word","text":"关于适用的解释"}

# 2. Stripped title (all case types)
GET judgments/_count
{"query":{"match_phrase":{"body":{"query":"关于适用的解释","analyzer":"ik_max_word"}}}}

# 3. Stripped only: stripped AND NOT any full procedural title
{"query":{"bool":{
  "must":[{"match_phrase":{"body":{"query":"关于适用的解释","analyzer":"ik_max_word"}}}],
  "must_not":[
    {"match_phrase":{"body":{"query":"关于适用中华人民共和国民事诉讼法的解释","analyzer":"ik_max_word"}}},
    {"match_phrase":{"body":{"query":"关于适用中华人民共和国刑事诉讼法的解释","analyzer":"ik_max_word"}}},
    {"match_phrase":{"body":{"query":"关于适用中华人民共和国行政诉讼法的解释","analyzer":"ik_max_word"}}}]}}}

# 4. Break down by dataset / year / caseType with a terms aggregation on each
# 5. Uniform sample: function_score + random_score(seed, field:_seq_no) + boost_mode:replace
#    then regex  关于适用的解释[》〉>＞]?[）)]?\s*(第[一二三四五六七八九十百零〇]+条)?  on the body

Run step 2 against your own delivery before step 3. If the stripped count is near zero on a corpus of tens of millions of judgments, either the vendor repaired the text (ask how, and ask for the flag) or the phrase query is failing for analyzer reasons. Both are worth knowing before you sign.

What this page does not claim

• We do not know what removed the inner titles, or at what stage. The bracket evidence argues against a single consistent tag-stripping pass; it does not establish an alternative.
• Counts are of records in our index on 3 October 2026, not of judgments in China. Duplicate records across batches are counted as they stand.
• “Stripped only” means no full title of the same family anywhere in the body. A record may still be resolvable from other citations in the same text; we did not use those.
• Article descriptions are given only for the four articles we are confident of in the 2015 numbering; the others are left unverified in the table.
• The resolution rates come from a sample of 2,000 and describe the strata, not a production parser we have shipped.
• Year buckets use the indexed year field, which is a publication year for some batches.