zm/tougao
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

参考文献优化

Browse Source
This commit is contained in:
wyn
2026-07-16 17:56:52 +08:00
parent 9b9eec24c2
commit ddb9c9f603
2 changed files with 27 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
application/common/ReferenceRelevanceCheckService.php
View File

@@ -581,15 +581,18 @@ class ReferenceRelevanceCheckService
}
$claims = $this->decodeClaimsJson(isset($row['claims_json']) ? $row['claims_json'] : '');
$author_comment = $this->resolveAuthorCommentFromRow($row, $claims);
$author_comment = $author_comment?"<b>[文献 ".intval($row['reference_no'])."]: </b>".$author_comment:$author_comment;
$g['records'][] = [
'check_id' => intval($row['id']),
'reference_no' => intval($row['reference_no']),
'am_id' => intval($row['am_id']),
'status' => $st,
'is_relevant' => intval($row['is_relevant']),
'relevance_score' => $score,
'is_pass' => $isPass,
'reason' => (string)$row['reason'],
'author_comment' => $this->resolveAuthorCommentFromRow($row, $claims),
'author_comment' => $author_comment,
'combined_relevance_score' => floatval($row['combined_relevance_score']),
'combined_reason' => (string)$row['combined_reason'],
'cite_group_refs' => (string)$row['cite_group_refs'],

32
application/common/service/ReferenceRelevanceLlmService.php
View File

@@ -464,7 +464,7 @@ Claim 具体内容见顶层 `claims`，combined_reason **只用字母指代**（
"is_relevant": 0,
"relevance_score": 0.45,
"reason": "系统综述，主语一致。Claim覆盖：A✔ B✔ C部分 D✘。类型部分匹配。因缺 D 且 C 仅背景，未达 0.65，故 0.45。",
"author_comment": "编辑意见：该文献与正文主张仅部分对应，关键论点仍缺直接证据。请改引更能支撑核心结论的文献，或补充说明引用依据。"
"author_comment": "该文献与正文主张仅部分对应，关键论点仍缺直接证据。请改引更能支撑核心结论的文献，或明确现有引用依据。"
},
{
"reference_no": 2,
@@ -478,7 +478,7 @@ Claim 具体内容见顶层 `claims`，combined_reason **只用字母指代**（
**顶层 claims**：本引用组共用，只写一次；键为 A/B/C…（按实际拆解数量），值为各 Claim 中文具体内容。
**results 每条仅含**：reference_no、is_relevant、relevance_score、reason、author_comment。
`author_comment` 规则：当 `relevance_score <= 0.65` 时必须生成（中文、**以期刊编辑审稿口吻**、礼貌可执行，**80–100字**）；当 `relevance_score > 0.65` 时必须返回空字符串 `""`。
`author_comment` 规则：当 `relevance_score <= 0.65` 时必须生成（中文、**以期刊编辑审稿口吻**、礼貌可执行，**80–100字**，句尾需句号）；当 `relevance_score > 0.65` 时必须返回空字符串 `""`。
注意：`author_comment` 禁止出现 A/B/C/D、✔/✘、"Claim覆盖"、具体分值（如 0.65）等技术表达，需改写为作者易读的编辑批注。
**禁止**在 results 各条中写 combined_relevance_score、combined_reason、cite_group_refs、claims。
PROMPT;
@@ -502,7 +502,7 @@ PROMPT;
if ($abstractText !== '') {
$parts[] = "【文献摘要/清洗后内容（Europe PMC·PubMed·Crossref·PDF）】\n" . $abstractText;
}
$tail = '请严格按六步标准校对流程执行：①提取Claim→②提取文献证据→③Claim Mapping（✔/部分/✘）→④计算覆盖率→⑤判断支持类型→⑥输出分值与理由。先在 JSON 顶层 claims 写出各 Claim 具体内容（多事实背景句须拆出比较型/成因型/进展型/数据型独立 Claim），再对每篇文献给出单条 relevance_score（弱相关文献不得因联合抬高），最后写顶层 combined_relevance_score 与 combined_reason。results 每条只写 reference_no、is_relevant、relevance_score、reason、author_comment。author_comment 规则：score<=0.65 时输出以期刊编辑审稿口吻的中文批注（礼貌、可执行，80–100字）；score>0.65 时固定空字符串；author_comment 禁止出现 A/B/C/D、✔/✘、Claim覆盖、具体分值。禁止 reason_en、combined_reason_en、relevance_level、relevance_role、combined_is_relevant 等多余字段。';
$tail = '请严格按六步标准校对流程执行：①提取Claim→②提取文献证据→③Claim Mapping（✔/部分/✘）→④计算覆盖率→⑤判断支持类型→⑥输出分值与理由。先在 JSON 顶层 claims 写出各 Claim 具体内容（多事实背景句须拆出比较型/成因型/进展型/数据型独立 Claim），再对每篇文献给出单条 relevance_score（弱相关文献不得因联合抬高），最后写顶层 combined_relevance_score 与 combined_reason。results 每条只写 reference_no、is_relevant、relevance_score、reason、author_comment。author_comment 规则：score<=0.65 时输出以期刊编辑审稿口吻的中文批注（礼貌、可执行，80–100字，句尾需句号）；若出现“建议补充/补充文献/补充说明”等措辞，需改写为“建议替换/改引/明确现有依据”，避免引导新增文献编号；score>0.65 时固定空字符串；author_comment 禁止出现 A/B/C/D、✔/✘、Claim覆盖、具体分值。禁止 reason_en、combined_reason_en、relevance_level、relevance_role、combined_is_relevant 等多余字段。';
if ($this->hasMixedReferTypes($referTypeMap)) {
$tail .= ' 注意：本组为图书与期刊混排引用，务必按上方【各编号文献类型标注】分轨判断——图书/教材走「书目推断」轨（据书名、副标题、作者、ISBN、出版社判断主题，缺摘要不得默认 0.25）；期刊/原始研究走「摘要核对」轨（据摘要/清洗内容逐项核对 claim）。同组内两类分别按各自标准独立打分，切勿用同一把尺子。';
} elseif ($this->allReferType($referTypeMap, 'book')) {
@@ -1612,28 +1612,28 @@ PROMPT;
if ($authorComment !== '') {
$authorComment = $this->sanitizeAuthorCommentText($authorComment);
if ($authorComment !== '') {
return mb_substr($authorComment, 0, 100);
return $this->finalizeAuthorComment($authorComment);
}
}
$fallback = '编辑意见：该条参考文献与正文匹配度有限，关键论点仍缺直接证据。请改引更能支撑核心结论的文献，或补充说明引用依据。';
$fallback = '该条参考文献与正文匹配度有限，关键论点仍缺直接证据。请改引更能支撑核心结论的文献，或明确现有引用依据。';
$reason = trim((string)$reason);
if ($reason === '') {
return $fallback;
return $this->finalizeAuthorComment($fallback);
}
// 去掉显式分数结论，保留给作者可读的编辑批注
$reason = preg_replace('/Claim覆盖[:：].*/u', '', $reason);
$reason = preg_replace('/故\s*(?:联合分?)?\s*[01](?:\.\d+)?[。.]?/u', '', $reason);
$reason = trim((string)$reason);
if ($reason === '') {
return $fallback;
return $this->finalizeAuthorComment($fallback);
}
$reason = $this->sanitizeAuthorCommentText($reason);
if ($reason === '') {
return $fallback;
return $this->finalizeAuthorComment($fallback);
}
return mb_substr('编辑意见：' . $reason, 0, 100);
return $this->finalizeAuthorComment($reason);
}
private function sanitizeAuthorCommentText($text)
@@ -1647,12 +1647,26 @@ PROMPT;
$text = preg_replace('/[✔✘]/u', '', $text);
$text = preg_replace('/\b0?\.\d{1,2}\b/u', '', $text);
$text = preg_replace('/\b\d{1,3}%\b/u', '', $text);
$text = preg_replace('/建议\s*补充/u', '建议替换', $text);
$text = preg_replace('/补充文献/u', '替换文献', $text);
$text = preg_replace('/补充说明/u', '明确依据', $text);
$text = preg_replace('/\s{2,}/u', ' ', $text);
$text = trim((string)$text, " \t\n\r\0\x0B;；，,");
return $text;
}
private function finalizeAuthorComment($text)
{
$text = mb_substr(trim((string)$text), 0, 100);
$text = trim($text, " \t\n\r\0\x0B;；，,。");
if ($text === '') {
return '';
}
return $text . '。';
}
private function boolVal($v)
{
if (is_bool($v)) {