Merge remote-tracking branch 'origin/master'
This commit is contained in:
wangjinlei
@@ -113,6 +113,68 @@ class PubmedService
|
||||
return $abbr !== '' ? $abbr : null;
|
||||
}
|
||||
|
||||
/**
|
||||
* 按书目信息检索 PubMed（标题 + 第一作者 + 年份）
|
||||
*/
|
||||
public function searchByBibliographic($title, $author = '', $year = ''): ?array
|
||||
{
|
||||
$title = trim((string)$title);
|
||||
if ($title === '') {
|
||||
return null;
|
||||
}
|
||||
|
||||
$terms = ['(' . $this->quoteTerm($title) . '[Title])'];
|
||||
$author = trim((string)$author);
|
||||
if ($author !== '') {
|
||||
$parts = preg_split('/[,;]/', $author);
|
||||
$first = trim((string)($parts[0] ?? ''));
|
||||
if ($first !== '') {
|
||||
$terms[] = '(' . $this->quoteTerm($first) . '[Author])';
|
||||
}
|
||||
}
|
||||
$year = trim((string)$year);
|
||||
if ($year !== '' && preg_match('/^(19|20)\d{2}$/', $year)) {
|
||||
$terms[] = '(' . $year . '[pdat])';
|
||||
}
|
||||
|
||||
$pmid = $this->esearch(implode(' AND ', $terms));
|
||||
if (!$pmid) {
|
||||
return null;
|
||||
}
|
||||
|
||||
$info = $this->fetchByPmid($pmid);
|
||||
if (!$info) {
|
||||
return null;
|
||||
}
|
||||
$info['pmid'] = $pmid;
|
||||
$info['doi'] = $this->extractDoiFromPmidRecord($pmid);
|
||||
return $info;
|
||||
}
|
||||
|
||||
private function quoteTerm($text)
|
||||
{
|
||||
return str_replace('"', '', trim((string)$text));
|
||||
}
|
||||
|
||||
private function extractDoiFromPmidRecord($pmid)
|
||||
{
|
||||
$url = $this->base . 'efetch.fcgi?' . http_build_query([
|
||||
'db' => 'pubmed',
|
||||
'id' => $pmid,
|
||||
'retmode' => 'xml',
|
||||
'tool' => $this->tool,
|
||||
'email' => $this->email,
|
||||
]);
|
||||
$xml = $this->httpGet($url);
|
||||
if ($xml === '') {
|
||||
return '';
|
||||
}
|
||||
if (preg_match('/<ArticleId IdType="doi">([^<]+)<\/ArticleId>/i', $xml, $m)) {
|
||||
return trim($m[1]);
|
||||
}
|
||||
return '';
|
||||
}
|
||||
|
||||
// ----------------- Internals -----------------
|
||||
|
||||
private function esearch(string $term): ?string
|
||||
|
||||
File diff suppressed because it is too large
Load Diff
@@ -3,10 +3,12 @@
|
||||
namespace app\common\mq;
|
||||
|
||||
use think\Db;
|
||||
use app\common\ReferenceCheckService;
|
||||
use app\common\DbReconnectHelper;
|
||||
use app\common\ReferenceRelevanceCheckService;
|
||||
|
||||
/**
|
||||
* RabbitMQ 消费：按文章串行，文章内 reference_no 升序逐条校对（含低分同步二轮）
|
||||
* RabbitMQ 消费（队列 reference_check / ref_check.article）：
|
||||
* 全局文章串行，文章内 reference_no 升序链式逐条「主题相关性」校对。
|
||||
*/
|
||||
class ReferenceCheckArticleWorker
|
||||
{
|
||||
@@ -15,18 +17,20 @@ class ReferenceCheckArticleWorker
|
||||
const BATCH_DONE = 2;
|
||||
const BATCH_PARTIAL_FAILED = 3;
|
||||
|
||||
/** @var ReferenceCheckService */
|
||||
/** @var ReferenceRelevanceCheckService */
|
||||
private $svc;
|
||||
|
||||
public function __construct()
|
||||
{
|
||||
$this->svc = new ReferenceCheckService();
|
||||
$this->svc = new ReferenceRelevanceCheckService();
|
||||
}
|
||||
|
||||
public function handleMessage(array $payload)
|
||||
{
|
||||
DbReconnectHelper::ensure();
|
||||
$pArticleId = intval(isset($payload['p_article_id']) ? $payload['p_article_id'] : 0);
|
||||
$batchId = intval(isset($payload['batch_id']) ? $payload['batch_id'] : 0);
|
||||
$trigger = isset($payload['trigger']) ? (string)$payload['trigger'] : 'enqueue';
|
||||
if ($pArticleId <= 0 || $batchId <= 0) {
|
||||
$this->svc->log('ReferenceCheckArticleWorker invalid payload');
|
||||
return;
|
||||
@@ -34,7 +38,11 @@ class ReferenceCheckArticleWorker
|
||||
|
||||
if (!$this->canStartArticleWork($batchId)) {
|
||||
$this->svc->log('ReferenceCheckArticleWorker defer batch_id=' . $batchId . ' other article running');
|
||||
(new ReferenceCheckMqPublisher())->publishArticleStart($pArticleId, $batchId, isset($payload['trigger']) ? $payload['trigger'] : 'enqueue');
|
||||
(new ReferenceCheckMqPublisher())->publishArticleStart(
|
||||
$pArticleId,
|
||||
$batchId,
|
||||
isset($payload['trigger']) ? $payload['trigger'] : 'enqueue'
|
||||
);
|
||||
sleep(3);
|
||||
return;
|
||||
}
|
||||
@@ -46,6 +54,11 @@ class ReferenceCheckArticleWorker
|
||||
}
|
||||
}
|
||||
|
||||
$this->svc->recoverQueueRowsForArticle($pArticleId);
|
||||
if ($trigger !== 'recheck_pending_only'
|
||||
&& ReferenceRelevanceCheckService::PREPARE_LITERATURE_BEFORE_CHECK) {
|
||||
$this->svc->prepareLiteratureContentByArticle($pArticleId);
|
||||
}
|
||||
$this->svc->log('ReferenceCheckArticleWorker start p_article_id=' . $pArticleId . ' batch_id=' . $batchId);
|
||||
|
||||
$done = 0;
|
||||
@@ -59,7 +72,7 @@ class ReferenceCheckArticleWorker
|
||||
if ($checkId <= 0) {
|
||||
continue;
|
||||
}
|
||||
$result = $this->processOneRow($checkId, $row);
|
||||
$result = $this->processOneRow($checkId, $row, $trigger === 'recheck_pending_only');
|
||||
if ($result === 'ok') {
|
||||
$done++;
|
||||
} elseif ($result === 'failed') {
|
||||
@@ -75,7 +88,7 @@ class ReferenceCheckArticleWorker
|
||||
|
||||
private function canStartArticleWork($batchId)
|
||||
{
|
||||
$running = Db::name('article_reference_check_batch')
|
||||
$running = Db::name('article_reference_relevance_check_batch')
|
||||
->where('batch_status', self::BATCH_RUNNING)
|
||||
->where('id', '<>', intval($batchId))
|
||||
->count();
|
||||
@@ -85,7 +98,7 @@ class ReferenceCheckArticleWorker
|
||||
private function claimBatch($batchId)
|
||||
{
|
||||
$now = date('Y-m-d H:i:s');
|
||||
$affected = Db::name('article_reference_check_batch')
|
||||
$affected = Db::name('article_reference_relevance_check_batch')
|
||||
->where('id', intval($batchId))
|
||||
->whereIn('batch_status', [self::BATCH_WAITING, self::BATCH_RUNNING])
|
||||
->update([
|
||||
@@ -97,15 +110,15 @@ class ReferenceCheckArticleWorker
|
||||
|
||||
private function getBatch($batchId)
|
||||
{
|
||||
return Db::name('article_reference_check_batch')->where('id', intval($batchId))->find();
|
||||
return Db::name('article_reference_relevance_check_batch')->where('id', intval($batchId))->find();
|
||||
}
|
||||
|
||||
private function fetchNextPendingRow($pArticleId)
|
||||
{
|
||||
return Db::name('article_reference_check_result')
|
||||
return Db::name('article_reference_relevance_check_result')
|
||||
->where('p_article_id', intval($pArticleId))
|
||||
->where('queue_status', ReferenceCheckService::QUEUE_PENDING)
|
||||
->where('status', ReferenceCheckService::RECORD_PENDING)
|
||||
->where('queue_status', ReferenceRelevanceCheckService::QUEUE_PENDING)
|
||||
->where('status', ReferenceRelevanceCheckService::RECORD_PENDING)
|
||||
->order('reference_no asc,am_id asc,text_start asc,id asc')
|
||||
->find();
|
||||
}
|
||||
@@ -113,44 +126,44 @@ class ReferenceCheckArticleWorker
|
||||
/**
|
||||
* @return string ok|failed|skip
|
||||
*/
|
||||
private function processOneRow($checkId, array $row)
|
||||
private function processOneRow($checkId, array $row, $skipLiteratureFetch = false)
|
||||
{
|
||||
$claimed = Db::name('article_reference_check_result')
|
||||
DbReconnectHelper::ensure();
|
||||
$claimed = Db::name('article_reference_relevance_check_result')
|
||||
->where('id', intval($checkId))
|
||||
->where('queue_status', ReferenceCheckService::QUEUE_PENDING)
|
||||
->update(['queue_status' => ReferenceCheckService::QUEUE_RUNNING]);
|
||||
->where('queue_status', ReferenceRelevanceCheckService::QUEUE_PENDING)
|
||||
->update(['queue_status' => ReferenceRelevanceCheckService::QUEUE_RUNNING]);
|
||||
if (intval($claimed) <= 0) {
|
||||
return 'skip';
|
||||
}
|
||||
|
||||
$retryCount = intval(isset($row['retry_count']) ? $row['retry_count'] : 0);
|
||||
try {
|
||||
$this->svc->runReferenceCheckOnce($checkId);
|
||||
$amId = intval(isset($row['am_id']) ? $row['am_id'] : 0);
|
||||
if ($amId > 0) {
|
||||
$this->svc->syncAmRefCheckStatus($amId);
|
||||
}
|
||||
$this->svc->markQueueRuntime($checkId, ReferenceCheckService::QUEUE_COMPLETED, $retryCount);
|
||||
$this->svc->runCheckOnce($checkId, $skipLiteratureFetch);
|
||||
$this->svc->markQueueRuntime($checkId, ReferenceRelevanceCheckService::QUEUE_COMPLETED, $retryCount);
|
||||
return 'ok';
|
||||
} catch (\Exception $e) {
|
||||
$this->svc->log('ReferenceCheckArticleWorker check_id=' . $checkId . ' err=' . $e->getMessage());
|
||||
if ($retryCount < ReferenceCheckService::QUEUE_MAX_RETRY) {
|
||||
$this->svc->markQueueRuntime($checkId, ReferenceCheckService::QUEUE_PENDING, $retryCount + 1);
|
||||
return $this->processOneRow($checkId, array_merge($row, ['retry_count' => $retryCount + 1]));
|
||||
DbReconnectHelper::ensure();
|
||||
if ($retryCount < ReferenceRelevanceCheckService::QUEUE_MAX_RETRY) {
|
||||
$this->svc->markQueueRuntime($checkId, ReferenceRelevanceCheckService::QUEUE_PENDING, $retryCount + 1);
|
||||
return $this->processOneRow($checkId, array_merge($row, ['retry_count' => $retryCount + 1]), $skipLiteratureFetch);
|
||||
}
|
||||
try {
|
||||
$this->svc->updateCheckResult($checkId, [
|
||||
'status' => ReferenceCheckService::RECORD_FAILED,
|
||||
'error_msg' => $e->getMessage(),
|
||||
]);
|
||||
$this->svc->markQueueRuntime($checkId, ReferenceCheckService::QUEUE_FAILED, $retryCount);
|
||||
$fresh = Db::name('article_reference_relevance_check_result')->where('id', intval($checkId))->find();
|
||||
$groupRows = !empty($fresh) ? $this->svc->findCitationGroupRowsForWorker($fresh) : [];
|
||||
if (!empty($groupRows)) {
|
||||
$this->svc->failGroupWithQueue($groupRows, $e->getMessage(), $retryCount);
|
||||
} else {
|
||||
$this->svc->updateCheckResult($checkId, [
|
||||
'status' => ReferenceRelevanceCheckService::RECORD_FAILED,
|
||||
'error_msg' => $e->getMessage(),
|
||||
]);
|
||||
$this->svc->markQueueRuntime($checkId, ReferenceRelevanceCheckService::QUEUE_FAILED, $retryCount);
|
||||
}
|
||||
} catch (\Exception $e2) {
|
||||
\think\Log::error('ReferenceCheckArticleWorker markFailed: ' . $e2->getMessage());
|
||||
}
|
||||
$amId = intval(isset($row['am_id']) ? $row['am_id'] : 0);
|
||||
if ($amId > 0) {
|
||||
$this->svc->syncAmRefCheckStatus($amId);
|
||||
}
|
||||
return 'failed';
|
||||
}
|
||||
}
|
||||
@@ -166,7 +179,7 @@ class ReferenceCheckArticleWorker
|
||||
if ($failed > 0) {
|
||||
$status = self::BATCH_PARTIAL_FAILED;
|
||||
}
|
||||
Db::name('article_reference_check_batch')->where('id', intval($batchId))->update([
|
||||
Db::name('article_reference_relevance_check_batch')->where('id', intval($batchId))->update([
|
||||
'batch_status' => $status,
|
||||
'done_count' => intval($done),
|
||||
'failed_count' => intval($failed),
|
||||
@@ -179,7 +192,7 @@ class ReferenceCheckArticleWorker
|
||||
|
||||
private function publishNextWaitingBatch()
|
||||
{
|
||||
$next = Db::name('article_reference_check_batch')
|
||||
$next = Db::name('article_reference_relevance_check_batch')
|
||||
->where('batch_status', self::BATCH_WAITING)
|
||||
->order('id asc')
|
||||
->find();
|
||||
@@ -193,8 +206,8 @@ class ReferenceCheckArticleWorker
|
||||
isset($next['trigger']) ? $next['trigger'] : 'enqueue'
|
||||
);
|
||||
} catch (\Exception $e) {
|
||||
$this->svc->log('publishNextWaitingBatch failed: ' . $e->getMessage());
|
||||
\think\Log::error('publishNextWaitingBatch: ' . $e->getMessage());
|
||||
$this->svc->log('ReferenceCheck publishNextWaitingBatch failed: ' . $e->getMessage());
|
||||
\think\Log::error('ReferenceCheck publishNextWaitingBatch: ' . $e->getMessage());
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
@@ -28,18 +28,17 @@ class LLMService
|
||||
* @param string $contextText 正文引用处句子
|
||||
* @param string $referText 参考文献条目（或 refer 格式化文本）
|
||||
* @param bool $isAgain 是否为 DOI 二次复核
|
||||
* @param string|null $doiBlock 可选：系统抓取到的 DOI 真实文献内容（仅二次复核使用）
|
||||
* @param string|null $doiBlock 可选：系统抓取到的 DOI 真实文献内容（仅二次复核使用）
|
||||
* @param string $citeGroupRefs 引用文献组，如 1,2 或 4,5,6
|
||||
* @param string $localContext 本引用位置附近上下文（可选）
|
||||
* @return array{results:array,request_failed?:bool}
|
||||
*/
|
||||
public function checkReference($contextText, $referText, $isAgain = false, $doiBlock = null)
|
||||
public function checkReference($contextText, $referText, $isAgain = false, $doiBlock = null, $citeGroupRefs = '', $localContext = '')
|
||||
{
|
||||
// request_failed=true 表示"LLM 通讯/解析层面的失败"（可重试，区别于业务上的"未命中"）；
|
||||
// 上游 runReferenceCheckOnce 会据此把 DB.status 置为 3(失败) 并抛异常触发 MQ worker 重试
|
||||
$fallback = [
|
||||
'can_support' => false,
|
||||
'is_match' => false,
|
||||
'confidence' => 0.0,
|
||||
'reason' => 'LLM not configured or request failed',
|
||||
'results' => [],
|
||||
'request_failed' => true,
|
||||
'reason' => 'LLM not configured or request failed',
|
||||
];
|
||||
if ($this->url === '' || $this->model === '') {
|
||||
\think\Log::warning('ReferenceCheck LLM: url or model not configured');
|
||||
@@ -47,15 +46,16 @@ class LLMService
|
||||
}
|
||||
|
||||
$contextText = trim($contextText);
|
||||
\think\Log::info('llm checkReference:' . $contextText);
|
||||
$referText = trim($referText);
|
||||
\think\Log::info('llm referText:' . $referText);
|
||||
$doiBlock = trim((string)$doiBlock);
|
||||
$citeGroupRefs = trim((string)$citeGroupRefs);
|
||||
$localContext = trim((string)$localContext);
|
||||
if ($contextText === '' || $referText === '') {
|
||||
// 空文本是入参问题，不是 LLM 故障，不需要重试
|
||||
return [
|
||||
'can_support' => false,
|
||||
'is_match' => false,
|
||||
'confidence' => 0.0,
|
||||
'reason' => 'Empty citation context or reference text',
|
||||
'results' => [],
|
||||
'reason' => 'Empty citation context or reference text',
|
||||
];
|
||||
}
|
||||
|
||||
@@ -63,27 +63,30 @@ class LLMService
|
||||
if (mb_strlen($contextText) > $maxContextLen) {
|
||||
$contextText = mb_substr($contextText, 0, $maxContextLen);
|
||||
}
|
||||
if (mb_strlen($referText) > 4000) {
|
||||
$referText = mb_substr($referText, 0, 4000);
|
||||
if (mb_strlen($localContext) > 3000) {
|
||||
$localContext = mb_substr($localContext, 0, 3000);
|
||||
}
|
||||
if (mb_strlen($doiBlock) > 4000) {
|
||||
$doiBlock = mb_substr($doiBlock, 0, 4000);
|
||||
if (mb_strlen($referText) > 6000) {
|
||||
$referText = mb_substr($referText, 0, 6000);
|
||||
}
|
||||
if (mb_strlen($doiBlock) > 8000) {
|
||||
$doiBlock = mb_substr($doiBlock, 0, 8000);
|
||||
}
|
||||
|
||||
if ($isAgain) {
|
||||
$system = $this->buildReferenceCheckSecondPassPrompt();
|
||||
$user = $this->buildReferenceCheckSecondPassUserPrompt($contextText, $referText, $doiBlock);
|
||||
$user = $this->buildReferenceCheckSecondPassUserPrompt($contextText, $referText, $doiBlock, $citeGroupRefs, $localContext);
|
||||
} else {
|
||||
$system = $this->buildReferenceCheckFirstPassPrompt();
|
||||
$user = $this->buildReferenceCheckFirstPassUserPrompt($contextText, $referText);
|
||||
$user = $this->buildReferenceCheckFirstPassUserPrompt($contextText, $referText, $citeGroupRefs, $localContext, $doiBlock);
|
||||
}
|
||||
|
||||
\think\Log::info('ReferenceCheck system head: ' . mb_substr($system, 0, 200));
|
||||
\think\Log::info('ReferenceCheck user head: ' . mb_substr($user, 0, 600));
|
||||
// \think\Log::info('ReferenceCheck system head: ' . mb_substr($system, 0, 200));
|
||||
// \think\Log::info('ReferenceCheck user head: ' . mb_substr($user, 0, 600));
|
||||
$payload = [
|
||||
'model' => $this->model,
|
||||
'model' => $this->model,
|
||||
'temperature' => 0,
|
||||
'messages' => [
|
||||
'messages' => [
|
||||
['role' => 'system', 'content' => $system],
|
||||
['role' => 'user', 'content' => $user],
|
||||
],
|
||||
@@ -101,23 +104,14 @@ class LLMService
|
||||
return $fallback;
|
||||
}
|
||||
|
||||
$canSupport = $this->parseCanSupportFromParsed($parsed);
|
||||
$confidence = $this->snapReferenceCheckConfidence(
|
||||
$this->normalizeConfidence(isset($parsed['confidence']) ? $parsed['confidence'] : 0),
|
||||
$canSupport
|
||||
);
|
||||
$reason = $this->cleanReason((string)(isset($parsed['reason']) ? $parsed['reason'] : ''));
|
||||
\think\Log::info(
|
||||
'ReferenceCheck result: can_support=' . ($canSupport ? '1' : '0')
|
||||
. ', confidence=' . $confidence
|
||||
. ', reason=' . $reason
|
||||
);
|
||||
return [
|
||||
'can_support' => $canSupport,
|
||||
'is_match' => $canSupport,
|
||||
'confidence' => $confidence,
|
||||
'reason' => $reason,
|
||||
];
|
||||
$results = $this->parseReferenceCheckResultsFromParsed($parsed, $citeGroupRefs, $localContext, $doiBlock);
|
||||
if (empty($results)) {
|
||||
\think\Log::warning('ReferenceCheck LLM: empty results array');
|
||||
return $fallback;
|
||||
}
|
||||
|
||||
\think\Log::info($results);
|
||||
return ['results' => $results];
|
||||
}
|
||||
|
||||
/**
|
||||
@@ -174,83 +168,541 @@ class LLMService
|
||||
$s = strtolower(trim((string)$value));
|
||||
return in_array($s, ['1', 'true', 'yes', 'support', 'supported'], true);
|
||||
}
|
||||
private function bulidReferenceCheckFirstPassPrompt(){
|
||||
return <<<'PROMPT'
|
||||
你是一名护理、医学与科研期刊的资深文献编辑，专门校对「正文引用句」与「对应参考文献条目」是否匹配。
|
||||
|
||||
/** 第一次校对：书目条目 vs 正文全文 */
|
||||
你的目标是严格识别错引、张冠李戴、方法不符、对象不符、结论不成立的问题。
|
||||
|
||||
宁可少判 true，也不要漏掉错引。
|
||||
|
||||
你只能依据用户提供的内容判断：
|
||||
1. 正文引用句
|
||||
2. 当前对应参考文献条目
|
||||
|
||||
禁止假设已阅读全文。
|
||||
禁止联网。
|
||||
禁止脑补文献内容。
|
||||
禁止根据学科常识推断研究结果。
|
||||
|
||||
====================
|
||||
【核心任务】
|
||||
|
||||
判断：
|
||||
|
||||
正文在该引用位置表达的核心观点、结论、方法、数据、定义、模型、研究发现、指南依据等，
|
||||
|
||||
是否能够被该条参考文献合理支撑。
|
||||
|
||||
你判断的是：
|
||||
|
||||
“引用是否成立”
|
||||
|
||||
不是：
|
||||
|
||||
“正文是否正确”。
|
||||
|
||||
====================
|
||||
【总原则（最高优先级）】
|
||||
|
||||
采用严格审稿标准：
|
||||
|
||||
边界不清时，一律判 false。
|
||||
|
||||
宁可误杀（人工复核），不要漏掉错引。
|
||||
|
||||
同领域 ≠ 匹配。
|
||||
|
||||
同关键词 ≠ 匹配。
|
||||
|
||||
相关 ≠ 能支撑。
|
||||
|
||||
====================
|
||||
【强制规则】
|
||||
|
||||
1. 严禁关键词硬匹配
|
||||
|
||||
不能因为出现：
|
||||
患者、护理、治疗、研究、模型、算法、深度学习、机器学习、焦虑、效果
|
||||
|
||||
等泛化词汇就判定匹配。
|
||||
|
||||
必须看：
|
||||
|
||||
- 核心对象
|
||||
- 研究问题
|
||||
- 方法
|
||||
- 场景
|
||||
- 结局指标
|
||||
- 核心论点
|
||||
|
||||
是否一致。
|
||||
|
||||
====================
|
||||
2. 方法学必须严格一致（极重要）
|
||||
|
||||
若正文明确提到：
|
||||
|
||||
- 算法
|
||||
- 模型
|
||||
- 聚类方法
|
||||
- 深度学习架构
|
||||
- 统计方法
|
||||
- 数学模型
|
||||
- 评价指标
|
||||
|
||||
必须要求文献与其存在明确关联。
|
||||
|
||||
例如：
|
||||
|
||||
不匹配：
|
||||
- fuzzy clustering ≠ deep learning
|
||||
- CNN ≠ LSTM
|
||||
- random forest ≠ SVM
|
||||
- 聚类 ≠ 分类
|
||||
- 特征选择 ≠ 分类预测
|
||||
- 风险因素分析 ≠ 干预研究
|
||||
|
||||
仅属于同一“大领域（AI/ML）”
|
||||
不能判定匹配。
|
||||
|
||||
若方法体系不同：
|
||||
|
||||
优先判 false + 0.10。
|
||||
|
||||
====================
|
||||
3. 医学护理引用严格一致
|
||||
|
||||
若正文涉及：
|
||||
|
||||
- 疾病
|
||||
- 人群
|
||||
- 护理场景
|
||||
- 干预措施
|
||||
- 结局指标
|
||||
|
||||
必须基本一致。
|
||||
|
||||
例如：
|
||||
|
||||
不匹配：
|
||||
- ICU ≠ 普通病房
|
||||
- 老年人 ≠ 儿童
|
||||
- 糖尿病 ≠ 高血压
|
||||
- 心理护理 ≠ 运动干预
|
||||
- 焦虑改善 ≠ 生存率提高
|
||||
|
||||
====================
|
||||
4. 强结论必须强证据
|
||||
|
||||
正文若出现：
|
||||
|
||||
- 显著改善
|
||||
- 明显降低
|
||||
- 证实
|
||||
- 优于
|
||||
- 有效预测
|
||||
- 危险因素
|
||||
- 因果关系
|
||||
|
||||
文献必须能合理支撑该强结论。
|
||||
|
||||
仅“应用研究”“相关研究”“观察研究”
|
||||
不能自动支持强结论。
|
||||
|
||||
否则 false。
|
||||
|
||||
====================
|
||||
5. 特定证据类型必须一致
|
||||
|
||||
正文若明确写：
|
||||
|
||||
- RCT/randomized trial
|
||||
- Meta-analysis
|
||||
- Guideline
|
||||
- Systematic review
|
||||
- Expert consensus
|
||||
|
||||
而参考文献类型明显不符：
|
||||
|
||||
直接 false。
|
||||
|
||||
====================
|
||||
6. 信息不足从严
|
||||
|
||||
若参考文献只有：
|
||||
|
||||
作者 + 年份
|
||||
|
||||
或信息过少，
|
||||
|
||||
无法建立明确关联：
|
||||
|
||||
false + 0.30
|
||||
|
||||
====================
|
||||
【判定逻辑】
|
||||
|
||||
只有同时满足以下条件，才能 true：
|
||||
|
||||
1. 主题一致
|
||||
2. 核心对象一致
|
||||
3. 核心论点一致
|
||||
4. 方法/研究方向一致
|
||||
5. 无明显错引风险
|
||||
|
||||
任意一点明显不符：
|
||||
|
||||
false。
|
||||
|
||||
====================
|
||||
【评分（只能四选一）】
|
||||
|
||||
只能输出：
|
||||
|
||||
0.90
|
||||
0.75
|
||||
0.30
|
||||
0.10
|
||||
|
||||
禁止任何其他分数。
|
||||
|
||||
评分规则：
|
||||
|
||||
0.90
|
||||
明确匹配：
|
||||
主题、对象、方法、核心论点均明显一致。
|
||||
|
||||
0.75
|
||||
基本匹配：
|
||||
整体支撑成立，但存在轻微概括或小范围表述差异。
|
||||
|
||||
0.30
|
||||
存疑：
|
||||
同领域但支撑不足；
|
||||
信息不足；
|
||||
需人工复核。
|
||||
|
||||
0.10
|
||||
明确错引：
|
||||
主题、对象、方法或核心论点明显不符。
|
||||
|
||||
硬规则：
|
||||
|
||||
is_match=true
|
||||
只能：
|
||||
0.75 或 0.90
|
||||
|
||||
is_match=false
|
||||
只能：
|
||||
0.10 或 0.30
|
||||
|
||||
====================
|
||||
【reason 要求】
|
||||
|
||||
仅说明：
|
||||
|
||||
1. 是否主题一致；
|
||||
2. 核心论点/方法是否能支撑。
|
||||
|
||||
禁止模糊措辞：
|
||||
“可能”
|
||||
“看起来”
|
||||
“应该”
|
||||
“疑似”
|
||||
|
||||
长度：
|
||||
|
||||
20~60字。
|
||||
|
||||
====================
|
||||
【输出要求】
|
||||
|
||||
仅输出一行 minified JSON。
|
||||
|
||||
禁止 markdown。
|
||||
禁止解释。
|
||||
禁止换行。
|
||||
禁止任何额外内容。
|
||||
|
||||
格式：
|
||||
|
||||
{"is_match":true|false,"confidence":0.10|0.30|0.75|0.90,"reason":"简体中文说明"}
|
||||
PROMPT;
|
||||
|
||||
}
|
||||
/** 第一次校对：参考文献真实性与支撑力度 */
|
||||
private function buildReferenceCheckFirstPassPrompt()
|
||||
{
|
||||
return <<<'PROMPT'
|
||||
你是文献引用校对助手。判断【正文全文】与【参考文献书目】是否相关、能否用于支撑正文中的引用。
|
||||
|
||||
【核心原则：从宽判断，避免误杀】
|
||||
默认倾向 can_support=true。只要文献与正文不是「风马牛不相及」，即判为相关、能支撑。
|
||||
不要求变量一致、不要求结论逐条对应、不要求研究设计相同。
|
||||
|
||||
【仅当以下情况才判 can_support=false（与正文明显无关）】
|
||||
- 学科/主题完全无关（如正文讲深度学习聚类，文献是糖尿病步态检测）。
|
||||
- 明显张冠李戴（正文断言 A 疗法的效果，文献研究的是完全不同的 B 问题且无关联）。
|
||||
- 文献条目与正文讨论的对象/场景毫无交集，且无法作背景或理论引用。
|
||||
|
||||
【以下情况均应 can_support=true】
|
||||
- 同一大领域或相邻方向（如护理、心理、管理、医学、统计、AI 等相近子领域）。
|
||||
- 可作背景文献、综述性引用、理论或方法的一般性依据。
|
||||
- 表述略宽、略有概括、变量名不完全一致，但大方向说得通。
|
||||
|
||||
【confidence 固定档位（禁止其它小数）】
|
||||
can_support=true：0.65（有关联但较泛）/ 0.78 / 0.85 / 0.92 / 0.98（非常确定相关）
|
||||
can_support=false：0.15（明确风马牛不相及）/ 0.25 / 0.35 / 0.45（仅当实在无法建立任何合理关联）
|
||||
|
||||
【输出】仅一行 minified JSON，无 markdown：
|
||||
{"can_support":true|false,"is_match":true|false,"confidence":0.15|0.25|0.35|0.45|0.65|0.78|0.85|0.92|0.98,"reason":"30-80字简体中文"}
|
||||
is_match 必须与 can_support 相同。
|
||||
PROMPT;
|
||||
return $this->buildReferenceCheckSupportSystemPrompt(false);
|
||||
}
|
||||
|
||||
private function buildReferenceCheckFirstPassUserPrompt($contextText, $referText)
|
||||
private function buildReferenceCheckSupportSystemPrompt($isSecondPass = false)
|
||||
{
|
||||
return "【正文全文 article_main.content】\n" . $contextText
|
||||
. "\n\n【参考文献书目 refer_text】\n" . $referText
|
||||
. "\n\n请从宽判断：文献与正文非风马牛不相即可判 can_support=true，只返回 JSON。";
|
||||
$prompt = <<<'PROMPT'
|
||||
你是一名护理、医学、生物医学与科研期刊的资深学术编辑，正在执行“参考文献真实性与支撑力度校对”。
|
||||
|
||||
你的任务不是判断“主题是否相关”，而是判断：
|
||||
【稿件正文中某段被引用内容】是否真的能被【对应编号的参考文献】直接或充分支撑。
|
||||
|
||||
你必须严格基于用户提供的材料作出判断，不得凭常识、不得脑补、不得假设参考文献中“可能写过但未提供”的内容。
|
||||
|
||||
==================================================
|
||||
【一、任务目标】
|
||||
你需要判断：
|
||||
“正文引用位置的核心论点、结论、背景陈述、机制解释、疗效描述、数据表达或因果表述，
|
||||
是否能被对应参考文献真实支持。”
|
||||
|
||||
这里的“支持”不是指“文献主题相关”或“研究领域接近”，而是指：
|
||||
参考文献中确实包含足以支持正文该处表述的内容。
|
||||
|
||||
==================================================
|
||||
【二、输出原则：结果必须直接对应数据库行】
|
||||
|
||||
你输出的结果将直接写入数据库表 t_article_reference_check_result。
|
||||
|
||||
因此：
|
||||
## 输出必须是 results 数组，数组中的每一个对象对应数据库中的一行，也就是“一个引用位置中的一条参考文献结果”。
|
||||
|
||||
换句话说：
|
||||
- 如果某个引用位置是 [3]，则输出 1 条 result（reference_no=3）
|
||||
- 如果某个引用位置是 [1,2]，则输出 2 条 result：
|
||||
- 一条对应 reference_no=1
|
||||
- 一条对应 reference_no=2
|
||||
|
||||
每条 result 都必须给出该参考文献“单独”对正文引用句的支撑判断。
|
||||
如果该引用位置是联合引用（citation group 中有多篇文献），则除了单条判断外，还必须给出该引用组整体的联合判断（combined_* 字段）。
|
||||
|
||||
==================================================
|
||||
【三、最重要原则：只看“是否支撑正文核心断言”，不是看“主题是否沾边”】
|
||||
|
||||
以下情况不能判为强支撑：
|
||||
1. 参考文献只和主题大致相关，但没有明确支持正文中的关键表述
|
||||
2. 正文说的是“疗效提升/死亡率下降/全球高发/耐药/多通路机制”等明确论点，而文献只是在背景里泛泛提到疾病
|
||||
3. 正文是多层复合句，文献只支撑其中一小部分
|
||||
4. 正文有因果、比较、趋势、机制、疗效强度等强表述，而文献没有明确证据
|
||||
5. 文献是基础机制研究，但正文引用它来支撑宏观流行病学、临床治疗现状或指南式结论
|
||||
6. 文献可以“推测支持”但不是“直接/明确支持”
|
||||
|
||||
==================================================
|
||||
【三b、多 claim 复合句 → 0.78 部分支撑（勿误降到 0.45）】
|
||||
|
||||
正文常为 2~4 个连续 claim 的复合句。须逐 claim 比对后综合给分：
|
||||
|
||||
- 若文献（含 DOI 摘要）能**明确支撑多数关键概念**（如遗传异质性/多基因改变、多 survival pathway 并存、耐药或治疗挑战），
|
||||
但**未逐字写出**正文完整因果链（如「异质性→多通路→单靶点疗效下降」），
|
||||
→ 应判 **partial_support**，confidence 通常 **0.78**（边界情况 0.65），**不得**仅因文献主标题聚焦某化合物/干预就降到 0.45。
|
||||
|
||||
- 0.45 仅用于：文献与 claim 方向明显不符、仅同病沾边、或几乎无可用证据。
|
||||
|
||||
**校准样例（单条 [4]，须接近此逻辑）：**
|
||||
|
||||
引用句：
|
||||
Furthermore, the genomic heterogeneity of colorectal cancer (CRC) presents additional difficulties because tumors frequently make use of several survival pathways at once, which reduces the efficacy of single-target treatments [4].
|
||||
|
||||
文献4（Sheikhnia et al., thymoquinone CRC 机制综述）：
|
||||
- Claim1 遗传异质性/多基因改变：文献有 APC/KRAS/TP53、MSI/CIN 等 → 支撑较强
|
||||
- Claim2 多 survival pathway：文献列举 PI3K/Akt、Wnt、STAT3、NF-κB 等多通路 → 支撑较强
|
||||
- Claim3 单靶点疗效下降：文献有 drug resistance/治疗挑战，但未直述因果链 → 部分支撑
|
||||
- **输出**：can_support=1, confidence=**0.78**, support_role=supplementary_support（**不是 0.45**）
|
||||
|
||||
用户消息中若提供【DOI 真实文献内容】，**必须结合摘要判断**，不得仅凭书目标题给分。
|
||||
|
||||
==================================================
|
||||
【四、评分规则】
|
||||
|
||||
你必须使用以下 8 个固定分值之一：
|
||||
0.98 / 0.92 / 0.85 / 0.78 / 0.65 / 0.45 / 0.25 / 0.15
|
||||
|
||||
判定含义：
|
||||
- 0.98 / 0.92 / 0.85 => 强支撑（strong_support）
|
||||
- 0.78 / 0.65 => 部分支撑（partial_support）
|
||||
- 0.45 / 0.25 => 支撑不足（insufficient_support）
|
||||
- 0.15 => 不支撑（not_support）
|
||||
|
||||
can_support 取值规则：
|
||||
- 若该文献/联合引文整体可判为 strong_support 或 partial_support，则 can_support = 1
|
||||
- 若判为 insufficient_support 或 not_support，则 can_support = 0
|
||||
|
||||
==================================================
|
||||
【五、单条文献结果如何判断】
|
||||
|
||||
对于每一条参考文献，你必须判断它“单独”能否支撑该引用位置的正文内容，并输出：
|
||||
- can_support
|
||||
- confidence
|
||||
- reason
|
||||
- support_role
|
||||
|
||||
其中：
|
||||
### support_role 只能取以下值之一
|
||||
- primary_support：该文献本身就是主要证据来源，能支撑引用句核心内容
|
||||
- supplementary_support：能支撑部分重要内容，但不是主要来源
|
||||
- minimal_support：只提供少量背景或边缘支撑
|
||||
- no_meaningful_support：几乎不能支撑该引用句
|
||||
|
||||
### reason 的写法要求
|
||||
必须使用中文，明确写出：
|
||||
1. 这篇文献具体支撑正文的哪一部分
|
||||
2. 哪些部分没有支撑到
|
||||
3. 是否存在文献类型与引用用途不匹配的问题
|
||||
4. 为什么给这个分值，而不是更高或更低
|
||||
|
||||
==================================================
|
||||
【六、联合引用的判断规则】
|
||||
|
||||
当同一个引用位置包含多篇参考文献时（例如 [1,2] / [4,5,6]），除了逐条给单条结果外，还要额外判断：
|
||||
“这些文献合起来，是否足以支撑该引用位置的正文内容？”
|
||||
|
||||
联合结论输出到：
|
||||
- combined_can_support
|
||||
- combined_confidence
|
||||
- combined_reason
|
||||
|
||||
规则：
|
||||
1. 联合评分不是单条评分平均值
|
||||
2. 如果其中一篇文献已强支撑，其他文献只是补充，则联合评分可接近主支撑文献
|
||||
3. 如果多篇文献分别覆盖不同部分，合起来能较完整支撑正文，则联合评分可以高于某些单条评分
|
||||
4. 但如果最关键的核心断言没有被任何文献明确支撑，则联合评分不能虚高
|
||||
5. 如果多篇文献都只是零散相关，需要大量推断才能拼出正文结论，则联合评分通常不应过高
|
||||
|
||||
==================================================
|
||||
【七、单引文的 combined_* 字段处理规则】
|
||||
|
||||
即使某个引用位置只有 1 条参考文献，也仍然必须输出 combined_* 字段。
|
||||
此时：
|
||||
- combined_can_support = can_support
|
||||
- combined_confidence = confidence
|
||||
- combined_reason = “该引用位置仅包含单条文献，联合结论等同于该文献的单条结论。” 或等价表述
|
||||
|
||||
这样可以保证输出结构统一，便于数据库写入。
|
||||
|
||||
==================================================
|
||||
【八、输出 JSON 结构】
|
||||
|
||||
你必须输出合法 JSON，且只能输出以下结构：
|
||||
|
||||
{
|
||||
"results": [
|
||||
{
|
||||
"reference_no": 1,
|
||||
"cite_group_refs": "1,2",
|
||||
"can_support": 0,
|
||||
"confidence": 0.65,
|
||||
"reason": "中文，单条文献结论",
|
||||
"support_role": "supplementary_support",
|
||||
"combined_can_support": 1,
|
||||
"combined_confidence": 0.85,
|
||||
"combined_reason": "中文，联合引用整体结论"
|
||||
}
|
||||
]
|
||||
}
|
||||
|
||||
==================================================
|
||||
【九、字段约束】
|
||||
|
||||
### 1）results 中每个对象都必须包含以下字段：
|
||||
- reference_no
|
||||
- cite_group_refs
|
||||
- can_support
|
||||
- confidence
|
||||
- reason
|
||||
- support_role
|
||||
- combined_can_support
|
||||
- combined_confidence
|
||||
- combined_reason
|
||||
|
||||
### 2）reference_no
|
||||
必须对应当前引用位置中的某一条参考文献编号。
|
||||
|
||||
### 3）cite_group_refs
|
||||
必须是该引用位置的完整引文组，格式如：
|
||||
- "3"
|
||||
- "1,2"
|
||||
- "4,5,6"
|
||||
|
||||
### 4）同一引用位置若包含多条参考文献，则必须输出多条 result
|
||||
例如 cite_group_refs = "1,2" 时，必须输出：
|
||||
- 一条 reference_no=1
|
||||
- 一条 reference_no=2
|
||||
|
||||
### 5）同一引用位置下的 combined_* 必须一致
|
||||
例如同属 "1,2" 的两条 result，它们的：
|
||||
- combined_can_support
|
||||
- combined_confidence
|
||||
- combined_reason
|
||||
必须完全一致。
|
||||
|
||||
==================================================
|
||||
【十、禁止事项】
|
||||
你绝对不能：
|
||||
- 杜撰文献中不存在的结论
|
||||
- 把“主题相关”当作“内容支撑”
|
||||
- 因为是同一疾病就默认支持
|
||||
- 输出 JSON 以外的任何内容
|
||||
|
||||
现在开始，读取用户提供的引用位置正文、参考文献信息和文献内容，输出结果。
|
||||
PROMPT;
|
||||
|
||||
if ($isSecondPass) {
|
||||
$prompt .= <<<'PROMPT'
|
||||
|
||||
|
||||
==================================================
|
||||
【二次校对补充（DOI 真实文献内容）】
|
||||
用户消息中会提供【DOI 真实文献内容（PubMed/Crossref）】。
|
||||
必须以 DOI 真实内容为准复核支撑力度；书目信息与 DOI 冲突时以 DOI 为准。
|
||||
仍须输出完整 results 数组，逐条给出单文献判断与联合判断。
|
||||
PROMPT;
|
||||
}
|
||||
|
||||
return $prompt;
|
||||
}
|
||||
|
||||
/** 第二次校对：Crossref 摘要（Refer_doi） */
|
||||
private function buildReferenceCheckFirstPassUserPrompt($contextText, $referText, $citeGroupRefs = '', $localContext = '', $doiBlock = '')
|
||||
{
|
||||
return $this->buildReferenceCheckSupportUserPrompt($contextText, $referText, $citeGroupRefs, $localContext, $doiBlock);
|
||||
}
|
||||
|
||||
private function buildReferenceCheckSupportUserPrompt($contextText, $referText, $citeGroupRefs, $localContext, $doiBlock)
|
||||
{
|
||||
$citeGroupRefs = trim((string)$citeGroupRefs);
|
||||
$localContext = trim((string)$localContext);
|
||||
$doiBlock = trim((string)$doiBlock);
|
||||
|
||||
$parts = [
|
||||
"【正文节 t_article_main】\n" . $contextText,
|
||||
];
|
||||
if ($citeGroupRefs !== '') {
|
||||
$mode = strpos($citeGroupRefs, ',') !== false ? '联合引用' : '单独引用';
|
||||
$parts[] = "【引用文献组 cite_group_refs】{$citeGroupRefs}（{$mode}）";
|
||||
}
|
||||
if ($localContext !== '') {
|
||||
$parts[] = "【本引用位置附近上下文】\n" . $localContext;
|
||||
}
|
||||
$parts[] = "【参考文献书目（按编号列出）】\n" . $referText;
|
||||
if ($doiBlock !== '') {
|
||||
$parts[] = "【DOI 真实文献内容（PubMed/Crossref，一轮校对已提供）】\n" . $doiBlock;
|
||||
}
|
||||
$parts[] = '请严格按 system 要求输出 results 数组 JSON，每条 result 对应一个 reference_no，并包含 combined_* 字段。';
|
||||
|
||||
return implode("\n\n", $parts);
|
||||
}
|
||||
|
||||
/** 第二次校对：DOI 真实文献内容复核 */
|
||||
private function buildReferenceCheckSecondPassPrompt()
|
||||
{
|
||||
return <<<'PROMPT'
|
||||
你是文献引用二次校对助手。已根据 Refer_doi 从 Crossref（https://api.crossref.org/works/）获取摘要，请结合【正文全文】复核该文献是否相关。
|
||||
|
||||
【核心原则：与第一次相同，从宽判断】
|
||||
默认倾向 can_support=true。只要 Crossref 摘要（或书目）与正文不是风马牛不相及，即判相关、能支撑。
|
||||
以【Crossref 摘要】为准；摘要与书目冲突时以摘要为准。
|
||||
|
||||
【仅当以下情况才判 can_support=false】
|
||||
- 摘要显示的研究主题/对象/方法与正文讨论内容完全风马牛不相及。
|
||||
- 典型风马牛不相及、张冠李戴，且无法解释为背景或泛化引用。
|
||||
|
||||
【以下情况均应 can_support=true】
|
||||
- 摘要与正文属同领域或相近方向，能作背景、理论或方向性支撑。
|
||||
- 细节不完全一致，但不存在明显矛盾。
|
||||
|
||||
【无 Crossref 摘要时】
|
||||
结合 refer_text 从宽判断；非明显无关仍可 can_support=true，confidence 建议 0.65。
|
||||
|
||||
【confidence 固定档位（禁止其它小数）】
|
||||
can_support=true：0.65 / 0.78 / 0.85 / 0.92 / 0.98
|
||||
can_support=false：0.15 / 0.25 / 0.35 / 0.45
|
||||
|
||||
【输出】仅一行 minified JSON：
|
||||
{"can_support":true|false,"is_match":true|false,"confidence":0.15|0.25|0.35|0.45|0.65|0.78|0.85|0.92|0.98,"reason":"30-80字简体中文"}
|
||||
is_match 必须与 can_support 相同。
|
||||
PROMPT;
|
||||
return $this->buildReferenceCheckSupportSystemPrompt(true);
|
||||
}
|
||||
|
||||
private function buildReferenceCheckSecondPassUserPrompt($contextText, $referText, $doiBlock)
|
||||
private function buildReferenceCheckSecondPassUserPrompt($contextText, $referText, $doiBlock, $citeGroupRefs = '', $localContext = '')
|
||||
{
|
||||
$doiBlock = trim((string)$doiBlock);
|
||||
return "【正文全文 article_main.content】\n" . $contextText
|
||||
. "\n\n【参考文献书目 refer_text】\n" . $referText
|
||||
. "\n\n【Crossref 摘要】（Refer_doi → api.crossref.org/works/）\n"
|
||||
. ($doiBlock !== '' ? $doiBlock : '（未获取到摘要，请结合 refer_text 从宽判断）')
|
||||
. "\n\n文献与正文非风马牛不相即可判 can_support=true，只返回 JSON。";
|
||||
return $this->buildReferenceCheckSupportUserPrompt(
|
||||
$contextText,
|
||||
$referText,
|
||||
$citeGroupRefs,
|
||||
$localContext,
|
||||
$doiBlock !== '' ? $doiBlock : '（未获取到 DOI 摘要或元数据，请结合书目条目从严判断）'
|
||||
);
|
||||
}
|
||||
private function buildReferenceCheckSystemPrompt3()
|
||||
{
|
||||
@@ -1169,13 +1621,174 @@ PROMPT;
|
||||
|
||||
private function buildReferenceCheckRecheckUserPrompt($contextText, $referText, $doiBlock)
|
||||
{
|
||||
return $this->buildReferenceCheckSecondPassUserPrompt($contextText, $referText, $doiBlock);
|
||||
return $this->buildReferenceCheckSecondPassUserPrompt($contextText, $referText, $doiBlock, '', '');
|
||||
}
|
||||
|
||||
/**
|
||||
* 与 buildReferenceCheckSystemPrompt3 一致的 confidence 档位
|
||||
* @return array<int, array>
|
||||
*/
|
||||
private function getReferenceCheckConfidenceBands($isMatch)
|
||||
private function parseReferenceCheckResultsFromParsed(array $parsed, $defaultCiteGroupRefs = '', $localContext = '', $doiBlock = '')
|
||||
{
|
||||
$rows = [];
|
||||
if (isset($parsed['results']) && is_array($parsed['results'])) {
|
||||
$rows = $parsed['results'];
|
||||
} elseif (isset($parsed['reference_no']) || isset($parsed['confidence'])) {
|
||||
$rows = [$parsed];
|
||||
}
|
||||
|
||||
$normalized = [];
|
||||
foreach ($rows as $item) {
|
||||
if (!is_array($item)) {
|
||||
continue;
|
||||
}
|
||||
$refNo = intval(isset($item['reference_no']) ? $item['reference_no'] : 0);
|
||||
if ($refNo <= 0) {
|
||||
continue;
|
||||
}
|
||||
|
||||
$confidence = $this->snapReferenceCheckConfidenceValue(
|
||||
$this->normalizeConfidence(isset($item['confidence']) ? $item['confidence'] : 0)
|
||||
);
|
||||
$canSupport = $this->canSupportFromConfidence($confidence);
|
||||
if (array_key_exists('can_support', $item)) {
|
||||
$canSupport = $this->boolFromLlmValue($item['can_support']);
|
||||
} elseif (array_key_exists('is_match', $item)) {
|
||||
$canSupport = $this->boolFromLlmValue($item['is_match']);
|
||||
}
|
||||
|
||||
$reason = $this->cleanReason((string)(isset($item['reason']) ? $item['reason'] : ''));
|
||||
$supportRole = $this->normalizeSupportRole(isset($item['support_role']) ? $item['support_role'] : '');
|
||||
list($confidence, $canSupport, $supportRole) = $this->applyMultiClaimPartialSupportFloor(
|
||||
$localContext,
|
||||
$doiBlock,
|
||||
$confidence,
|
||||
$canSupport,
|
||||
$supportRole,
|
||||
$reason
|
||||
);
|
||||
|
||||
$combinedConfidence = $this->snapReferenceCheckConfidenceValue(
|
||||
$this->normalizeConfidence(isset($item['combined_confidence']) ? $item['combined_confidence'] : $confidence)
|
||||
);
|
||||
$combinedCanSupport = $this->canSupportFromConfidence($combinedConfidence);
|
||||
if (array_key_exists('combined_can_support', $item)) {
|
||||
$combinedCanSupport = $this->boolFromLlmValue($item['combined_can_support']);
|
||||
}
|
||||
|
||||
$citeGroupRefs = trim((string)(isset($item['cite_group_refs']) ? $item['cite_group_refs'] : $defaultCiteGroupRefs));
|
||||
if ($citeGroupRefs === '' && $defaultCiteGroupRefs !== '') {
|
||||
$citeGroupRefs = trim((string)$defaultCiteGroupRefs);
|
||||
}
|
||||
|
||||
$normalized[] = [
|
||||
'reference_no' => $refNo,
|
||||
'cite_group_refs' => $citeGroupRefs,
|
||||
'can_support' => $canSupport,
|
||||
'is_match' => $canSupport,
|
||||
'confidence' => $confidence,
|
||||
'reason' => $reason,
|
||||
'support_role' => $supportRole,
|
||||
'combined_can_support' => $combinedCanSupport,
|
||||
'combined_confidence' => $combinedConfidence,
|
||||
'combined_reason' => $this->cleanReason((string)(isset($item['combined_reason']) ? $item['combined_reason'] : '')),
|
||||
];
|
||||
}
|
||||
|
||||
return $normalized;
|
||||
}
|
||||
|
||||
private function normalizeSupportRole($role)
|
||||
{
|
||||
$role = strtolower(trim((string)$role));
|
||||
$allowed = [
|
||||
'primary_support',
|
||||
'supplementary_support',
|
||||
'minimal_support',
|
||||
'no_meaningful_support',
|
||||
];
|
||||
return in_array($role, $allowed, true) ? $role : 'no_meaningful_support';
|
||||
}
|
||||
|
||||
private function canSupportFromConfidence($confidence)
|
||||
{
|
||||
return floatval($confidence) >= 0.65 - 0.001;
|
||||
}
|
||||
|
||||
/**
|
||||
* 多通路/异质性 claim + DOI 有多通路证据时，防止误打 0.45（应对齐 0.78 部分支撑）
|
||||
*/
|
||||
private function applyMultiClaimPartialSupportFloor($localContext, $doiBlock, $confidence, $canSupport, $supportRole, $reason)
|
||||
{
|
||||
$confidence = floatval($confidence);
|
||||
if ($confidence > 0.45) {
|
||||
return [$confidence, $canSupport, $supportRole];
|
||||
}
|
||||
|
||||
$claimText = trim((string)$localContext);
|
||||
if ($claimText === '') {
|
||||
return [$confidence, $canSupport, $supportRole];
|
||||
}
|
||||
|
||||
$claimIsMechanism = (bool)preg_match(
|
||||
'/\b(genomic heterogeneity|heterogeneity|survival pathway|pathways at once|single-target|multi.?pathway|genetic alteration|drug resistance|异质性|生存通路|多.*通路|单靶点|耐药)\b/ui',
|
||||
$claimText
|
||||
);
|
||||
if (!$claimIsMechanism) {
|
||||
return [$confidence, $canSupport, $supportRole];
|
||||
}
|
||||
|
||||
$corpus = trim((string)$doiBlock) . ' ' . trim((string)$reason);
|
||||
if ($corpus === '') {
|
||||
return [$confidence, $canSupport, $supportRole];
|
||||
}
|
||||
|
||||
$refHasPathwayEvidence = (bool)preg_match(
|
||||
'/\b(pathway|PI3K|Akt|mTOR|Wnt|STAT3|NF-κB|NF-kB|genetic alteration|MSI|CIN|drug resistance|signaling|multiple|APC|KRAS|TP53|通路|耐药|信号)\b/ui',
|
||||
$corpus
|
||||
);
|
||||
if (!$refHasPathwayEvidence) {
|
||||
return [$confidence, $canSupport, $supportRole];
|
||||
}
|
||||
|
||||
$confidence = 0.78;
|
||||
$canSupport = true;
|
||||
if ($supportRole === 'no_meaningful_support' || $supportRole === 'minimal_support') {
|
||||
$supportRole = 'supplementary_support';
|
||||
}
|
||||
|
||||
return [$confidence, $canSupport, $supportRole];
|
||||
}
|
||||
|
||||
private function getReferenceCheckConfidenceBands()
|
||||
{
|
||||
return [0.15, 0.25, 0.45, 0.65, 0.78, 0.85, 0.92, 0.98];
|
||||
}
|
||||
|
||||
private function snapReferenceCheckConfidenceValue($confidence)
|
||||
{
|
||||
$bands = $this->getReferenceCheckConfidenceBands();
|
||||
foreach ($bands as $band) {
|
||||
if (abs($confidence - $band) < 0.001) {
|
||||
return $band;
|
||||
}
|
||||
}
|
||||
$nearest = $bands[0];
|
||||
$minDiff = abs($confidence - $nearest);
|
||||
foreach ($bands as $band) {
|
||||
$diff = abs($confidence - $band);
|
||||
if ($diff < $minDiff) {
|
||||
$minDiff = $diff;
|
||||
$nearest = $band;
|
||||
}
|
||||
}
|
||||
|
||||
return $nearest;
|
||||
}
|
||||
|
||||
/**
|
||||
* @deprecated 兼容旧逻辑
|
||||
*/
|
||||
private function getReferenceCheckConfidenceBandsLegacy($isMatch)
|
||||
{
|
||||
return $isMatch
|
||||
? [0.65, 0.78, 0.85, 0.92, 0.98]
|
||||
@@ -1183,22 +1796,24 @@ PROMPT;
|
||||
}
|
||||
|
||||
/**
|
||||
* 将模型输出的 confidence 吸附到合法档位（如 0.95 → 0.92，0.75 → 0.78）
|
||||
* 将模型输出的 confidence 吸附到合法档位
|
||||
*/
|
||||
private function snapReferenceCheckConfidence($confidence, $isMatch)
|
||||
{
|
||||
$bands = $this->getReferenceCheckConfidenceBands($isMatch);
|
||||
|
||||
$snapped = $this->snapReferenceCheckConfidenceValue($confidence);
|
||||
$bands = $this->getReferenceCheckConfidenceBandsLegacy($isMatch);
|
||||
if (in_array($snapped, $bands, true)) {
|
||||
return $snapped;
|
||||
}
|
||||
foreach ($bands as $band) {
|
||||
if (abs($confidence - $band) < 0.001) {
|
||||
if (abs($snapped - $band) < 0.001) {
|
||||
return $band;
|
||||
}
|
||||
}
|
||||
|
||||
$nearest = $bands[0];
|
||||
$minDiff = abs($confidence - $nearest);
|
||||
$minDiff = abs($snapped - $nearest);
|
||||
foreach ($bands as $band) {
|
||||
$diff = abs($confidence - $band);
|
||||
$diff = abs($snapped - $band);
|
||||
if ($diff < $minDiff) {
|
||||
$minDiff = $diff;
|
||||
$nearest = $band;
|
||||
|
||||
670
application/common/service/ReferenceRelevanceLlmService.php
Normal file
670
application/common/service/ReferenceRelevanceLlmService.php
Normal file
@@ -0,0 +1,670 @@
|
||||
<?php
|
||||
|
||||
namespace app\common\service;
|
||||
|
||||
use think\Env;
|
||||
|
||||
/**
|
||||
* 参考文献「主题相关性」LLM 校对（独立于支撑力度校对 LLMService）
|
||||
*/
|
||||
class ReferenceRelevanceLlmService
|
||||
{
|
||||
private $url;
|
||||
private $model;
|
||||
private $apiKey;
|
||||
private $timeout;
|
||||
private $lastPostError = '';
|
||||
private $maxSectionChars;
|
||||
private $maxLocalContextChars;
|
||||
private $maxReferChars;
|
||||
private $maxAbstractChars;
|
||||
|
||||
public function __construct()
|
||||
{
|
||||
$this->url = trim((string)Env::get('promotion.promotion_llm_url', ''));
|
||||
$this->model = trim((string)Env::get('promotion.promotion_llm_model', ''));
|
||||
$this->apiKey = trim((string)Env::get('promotion.promotion_llm_api_key', ''));
|
||||
$this->timeout = max(180, intval(Env::get('promotion.promotion_llm_timeout', 180)));
|
||||
// 控制发送给 LLM 的上下文长度，降低单次推理耗时（可通过 env 覆盖）
|
||||
$this->maxSectionChars = max(1500, intval(Env::get('promotion.relevance_llm_max_section_chars', 4500)));
|
||||
$this->maxLocalContextChars = max(600, intval(Env::get('promotion.relevance_llm_max_local_context_chars', 1800)));
|
||||
$this->maxReferChars = max(1500, intval(Env::get('promotion.relevance_llm_max_refer_chars', 3500)));
|
||||
$this->maxAbstractChars = max(1500, intval(Env::get('promotion.relevance_llm_max_abstract_chars', 3500)));
|
||||
}
|
||||
|
||||
/**
|
||||
* @return array{results:array,request_failed?:bool,reason?:string}
|
||||
*/
|
||||
public function checkRelevance($sectionText, $localContext, $referText, $abstractText = '', $citeGroupRefs = '')
|
||||
{
|
||||
$fallback = [
|
||||
'results' => [],
|
||||
'request_failed' => true,
|
||||
'reason' => 'LLM not configured or request failed',
|
||||
];
|
||||
if ($this->url === '' || $this->model === '') {
|
||||
return $fallback;
|
||||
}
|
||||
|
||||
$sectionText = trim((string)$sectionText);
|
||||
$localContext = trim((string)$localContext);
|
||||
$referText = trim((string)$referText);
|
||||
$abstractText = trim((string)$abstractText);
|
||||
if ($sectionText === '' || $referText === '') {
|
||||
return ['results' => [], 'reason' => 'Empty section or reference text'];
|
||||
}
|
||||
|
||||
if (mb_strlen($sectionText) > $this->maxSectionChars) {
|
||||
$sectionText = mb_substr($sectionText, 0, $this->maxSectionChars);
|
||||
}
|
||||
if (mb_strlen($localContext) > $this->maxLocalContextChars) {
|
||||
$localContext = mb_substr($localContext, 0, $this->maxLocalContextChars);
|
||||
}
|
||||
if (mb_strlen($referText) > $this->maxReferChars) {
|
||||
$referText = mb_substr($referText, 0, $this->maxReferChars);
|
||||
}
|
||||
if (mb_strlen($abstractText) > $this->maxAbstractChars) {
|
||||
$abstractText = mb_substr($abstractText, 0, $this->maxAbstractChars);
|
||||
}
|
||||
|
||||
$payload = [
|
||||
'model' => $this->model,
|
||||
'temperature' => 0,
|
||||
'messages' => [
|
||||
['role' => 'system', 'content' => $this->buildSystemPrompt()],
|
||||
['role' => 'user', 'content' => $this->buildUserPrompt($sectionText, $localContext, $referText, $abstractText, $citeGroupRefs)],
|
||||
],
|
||||
];
|
||||
|
||||
$content = $this->postChat($payload);
|
||||
if ($content === null) {
|
||||
$reason = $this->lastPostError !== '' ? $this->lastPostError : 'LLM request failed';
|
||||
return array_merge($fallback, ['reason' => $reason]);
|
||||
}
|
||||
|
||||
$parsed = $this->parseJson($content);
|
||||
if ($parsed === null) {
|
||||
return array_merge($fallback, ['reason' => 'LLM response JSON parse failed']);
|
||||
}
|
||||
|
||||
$results = $this->normalizeResults($parsed, $citeGroupRefs, $localContext, $referText, $abstractText);
|
||||
if (empty($results)) {
|
||||
return array_merge($fallback, ['reason' => 'LLM returned empty or invalid results']);
|
||||
}
|
||||
|
||||
return ['results' => $results];
|
||||
}
|
||||
|
||||
private function buildSystemPrompt()
|
||||
{
|
||||
return <<<'PROMPT'
|
||||
你是一名护理、医学、生物医学与科研期刊的资深学术编辑，正在执行「参考文献主题相关性校对」。
|
||||
|
||||
你的任务：判断【引用位置正文表述】与【对应编号参考文献】在主题、研究对象、疾病/场景/结局方向上是否相关，能否作为该处引用的合理来源。
|
||||
|
||||
注意：这是「相关性」校对，侧重引用处具体 claim 与文献内容是否匹配；**不是**判断「是否同一疾病/同一领域」。
|
||||
|
||||
==================================================
|
||||
【零、最硬规则（违反则输出无效）】
|
||||
1. **单条 relevance_score 只评价该编号文献单独**与引用处的关系；不得因联合组整体合理而抬高弱相关文献的单条分。
|
||||
2. **禁止「同病高分」**：正文与文献都涉及 CRC，不等于单条可给 0.85~0.92。
|
||||
**但若引用处 claim 本身就是机制/通路/异质性/耐药/治疗挑战**，且**研究主语一致**（同一疾病/同一化合物/同一干预对象），文献（含摘要/清洗内容）讨论同病多通路、遗传改变、耐药等，应给 **0.65~0.78**，不得误降到 0.45。
|
||||
**主语不一致时仍适用本条禁止高分**：引用处主语为化合物 X，文献却是其他植物/提取物/计算预测，即使提到 X 或相同通路名，也不得因此给 0.78+。
|
||||
3. 引用处若为**流行病学/负担类 claim**（most common、incidence、mortality、burden、全球高发等）：
|
||||
- 机制研究、分子通路、细胞增殖/迁移、血管生成等**原始研究** → 单条通常 **0.45 或更低**，`is_relevant=0`，`minimal_relevance`
|
||||
- 不得因摘要提到 colorectal cancer 就给 0.92
|
||||
- 仅当文献为流行病学综述/公共卫生研究，或明确讨论发病率、死亡率、疾病负担时，单条才可 **0.85~0.92**
|
||||
4. **联合分写在 combined_relevance_score**，与单条分必须可分离（例如 [1,2] 时文献1=0.45、文献2=0.92、联合=0.92）。
|
||||
5. **「来源/化学分类」型句子**（naturally occurring、pentacyclic triterpenoid、found in fruits/vegetables/medicinal plants、并列举具体植物学名）：
|
||||
- 先判文献类型：来源综述 / 生物活性综述 最适合；**抗癌治疗综述**对「来源分布」claim 通常仅 **0.65**
|
||||
- 单篇可差异化打分（如 0.92 / 0.92 / 0.65），**不得**因联合而三篇都给高分
|
||||
- 若原句含**具体列举项**（如多个植物学名），而材料未逐一核实全部学名，联合分通常 **≤0.85**（不得给 0.98）
|
||||
6. **多要素综括句**（一句同时塞入：药学/研究兴趣 + 大量前临床研究 + 多种活性[抗炎/抗氧化/抗癌等] + 多个癌种/对象列举）：
|
||||
- 单篇即使是综述，通常仅 partially_related ~ near-direct（**0.78~0.86**），**不轻易给 0.92**（单篇难逐项覆盖全部要素）
|
||||
- **联合分是整句覆盖度评估，可低于最高单条分**：若整句要素需多篇拼合、且含作者整合概括，联合通常 **0.72~0.78（partially_related）**，不给 0.85+
|
||||
7. **联合分不是「取最高单条分」**：当各单篇都只覆盖整句一部分、需互补拼合时，联合分应反映「整句作为一个整体被支撑的完整度」，**允许低于任何一篇单条分**。
|
||||
8. **主语/研究对象层级必须对齐**：引用处主语为某化合物/分子（如「X has been demonstrated…」）时，文献核心对象须为 **X 本身**或以 X 为核心的实验/综述。
|
||||
- **植物提取物/混合物**研究、**其他物种/其他植物**的计算预测、成分表中顺带出现 X → 通常 **0.45 或更低**
|
||||
- **关键：提取物即使 X 含量很高（如 50%+）且显示了抗癌/凋亡活性，活性归因于提取物整体而非 X 单体单独验证 → 仍属 weakly_related（≤0.45）、`minimal_relevance`**，不得评为 supplementary_relevance/0.78
|
||||
- 只有当文献**针对 X 单体单独做了验证**（X monomer 处理、X 单独剂量效应等）时，主语才算对齐，方可进入 0.65+
|
||||
- **不得**因摘要/讨论出现与引用句相同的通路名、凋亡、抗癌等词就给 0.78+
|
||||
9. **证据层级与 demonstrated / mechanistically**：
|
||||
- 本文实验结果或针对 X 的系统综述 > 计算预测/混合成分推测
|
||||
- **讨论（Discussion）转引他人关于 X 的机制总结 ≠ 该文自身证据**；据此最多 **0.45~0.65**，不得评为 highly_related
|
||||
- in silico / computational prediction 不足以支撑「has been demonstrated to mechanistically…」式强语气 claim 的高分
|
||||
10. **点名通路/功能结局须逐项核对**：原句逐条列举通路（如 PI3K/AKT、MAPK、NF-κB）或结局（增殖、凋亡、血管生成、炎症信号等）时，**每一项单独核对是否在本文证据中成立**（非仅背景提及）。
|
||||
- 讨论转述既往文献 ≠ 本文证明该项
|
||||
- 缺原句任一点名项（如 angiogenesis）→ 单条通常 **不得 0.78+**
|
||||
- **「覆盖部分结局」不足以进入 0.78**：原句点名了多条通路 + 多个结局，文献仅命中其中 1~2 个结局（如仅凋亡/增殖），且**点名通路在本文结果中全部缺失（仅讨论转引）**或主语层级不对 → 单条 **限 0.45（weakly_related / minimal_relevance）**，不得给 0.65~0.78
|
||||
- 仅同领域沾边 1–2 项、主语或机制层级不对 → **0.45**
|
||||
- **进入 0.65~0.78 的前提**：主语对齐（X 单体）+ 本文自身结果命中原句点名通路/结局的多数项；几乎全部明确对应 → **0.85+**
|
||||
11. **文献「主题粒度」必须匹配 claim「主题粒度」**：引用处为**疾病总论型 claim**（流行病学负担、标准/多模态治疗现状与局限、基因组异质性、单靶点治疗受限、亟需新策略等总体背景）时：
|
||||
- 最适合的来源是**疾病总体综述 / 分子病理综述 / 精准肿瘤学 / 耐药综述**；此类文献正面、系统地为该总论 claim 提供依据 → 可 **0.85+**
|
||||
- **单一药物 / 单一成分 / 单一通路的专题综述**（如「某化合物抗某癌：A review」），即使同病、同大方向，也只是专题视角、并非为该总论 claim 做系统总结 → 通常 **partially_related（0.72~0.78）**，**不得给 0.85+**
|
||||
- **单基因 / 单通路的机制原始研究**对纯流行病学负担 claim → 仍按规则 3 给 **0.45**
|
||||
- 判断要点：文献类型是否「为该总论 claim 本身做系统综述/总论」；仅同病同方向、或只支撑整段中某一两句（如「需要更安全的新策略」），不足以进入 highly_related
|
||||
|
||||
==================================================
|
||||
【一、必须先拆解 claim】
|
||||
从【本引用位置附近上下文】中提炼最小主张单元（Claim A, Claim B…），**不要**把整句笼统归为「大概讲抗癌」。例如：
|
||||
- **主语/研究对象**（化合物单体 vs 植物提取物 vs 其他物种；是否「X has been demonstrated」）
|
||||
- **证据语气与层级**（demonstrated / mechanistically vs predict / suggest；本文结果 vs 讨论转引）
|
||||
- **claim 主题粒度**：是否为疾病总论型（流行病学负担 / 治疗现状与局限 / 基因组异质性 / 单靶点受限 / 亟需新策略）；若是，要求「总体综述 / 分子病理 / 精准肿瘤学 / 耐药综述」类来源，单一药物专题综述只算 partially_related
|
||||
- 疾病流行病学（高发、死亡率）
|
||||
- **点名通路/分子机制**（PI3K/AKT、MAPK、NF-κB 等，须逐项）
|
||||
- **点名功能结局**（抑制增殖、凋亡、血管生成、炎症信号等，须逐项）
|
||||
- 治疗/干预现状
|
||||
- **化合物化学类别**（如 pentacyclic triterpenoid）
|
||||
- **天然来源分布**（fruits / vegetables / medicinal plants）
|
||||
- **具体列举项**（植物学名、药名、基因名等，须逐项核对）
|
||||
|
||||
==================================================
|
||||
【二、逐篇文献单独判断（每条 result 对应一个 reference_no）】
|
||||
对 cite_group_refs 中的每一篇文献，单独输出：
|
||||
- 该文献与引用处哪些 claim 主题相关、哪些不相关（含具体列举项是否覆盖）
|
||||
- 文献类型是否匹配引用用途（来源综述 / 生物活性综述 / 机制研究 / 流行病学综述 / 抗癌治疗综述等）
|
||||
- relevance_score：只能使用 0.98 / 0.92 / 0.85 / 0.78 / 0.65 / 0.45 / 0.25 / 0.15
|
||||
- relevance_level：highly_related | partially_related | weakly_related | unrelated
|
||||
- is_relevant：score>=0.65 为 1，否则 0
|
||||
- relevance_role：
|
||||
- primary_relevance：该文献是引用处主题的主要相关来源
|
||||
- supplementary_relevance：部分相关、补充性
|
||||
- minimal_relevance：仅边缘/背景沾边
|
||||
- no_meaningful_relevance：与引用处核心表述基本无关
|
||||
- reason：中英双语结论，格式固定为两行：
|
||||
【中文】（中文结论，须写明：①文献类型与**核心研究对象** ②**本文自身证据**覆盖了哪些 claim / 哪些未覆盖 ③主语/claim 不匹配须明确写出 ④为何此分值）
|
||||
【English】（与中文对应的英文结论，语义一致）
|
||||
- reason_en：仅英文结论（与 reason 中【English】段相同，勿留空）
|
||||
|
||||
主语/层级不对 → 单条 **0.45**，不得因讨论提及相同通路给 0.78：
|
||||
引用处 claim 为「化合物 X 经 PI3K/AKT 等机制 demonstrated…」，文献为其他植物提取物或计算预测、仅在讨论转引他人 X 机制 → 0.45，weakly_related，is_relevant=0。
|
||||
|
||||
机制文引用流行病学句 → 单条 **0.45**，不得 0.92：
|
||||
文献为 CRC 机制研究，引用处 claim 为全球高发/死亡率，文献无流行病学数据 → 0.45，minimal_relevance，is_relevant=0。
|
||||
|
||||
==================================================
|
||||
【三、联合引用 combined_*（同一 cite_group_refs 内各行必须一致）】
|
||||
当 cite_group_refs 为 "1,2" 等多篇时，除逐篇判断外，必须给出引用组整体结论：
|
||||
- 这些文献合起来，是否足以支撑/匹配该引用位置的整体表述？
|
||||
- combined_relevance_score：八档固定分值之一，**不是单条平均分**
|
||||
- 若一篇已强相关、其余仅弱补充，联合分可接近主相关文献，但**不必等于最高单条分**
|
||||
- 若原句含具体列举项（学名等）且材料未逐一核实，联合分通常 **0.85**，不给 0.98
|
||||
- 若核心 claim 无任何文献明确覆盖，联合分不能虚高
|
||||
- 多篇联合仍缺主语对齐、缺原句点名通路/结局、或主要靠讨论转引 → 联合分通常 **≤0.45~0.65**，不得因单篇讨论出现相同关键词给到 0.78+
|
||||
- combined_is_relevant：combined_relevance_score>=0.65 为 1
|
||||
- combined_relevance_level：与 combined 分数对应的等级
|
||||
- combined_reason：中英双语综合结论，格式同 reason（【中文】/【English】），说明各文献分工及最终分值理由
|
||||
- combined_reason_en：仅英文综合结论（与 combined_reason 中【English】段相同）
|
||||
|
||||
单条引用时：combined_* 与单条一致；combined_reason / combined_reason_en 可与 reason / reason_en 相同。
|
||||
|
||||
==================================================
|
||||
【四、评分与等级对照】
|
||||
0.98 / 0.92 / 0.85 = highly_related
|
||||
文献直接支持整句主旨，大部分关键要素都在文中明确出现
|
||||
0.78 / 0.65 = partially_related
|
||||
文献只支撑其中一部分，或支撑方式偏间接
|
||||
0.45 = weakly_related
|
||||
只是同领域文献，但与句子事实对应很弱
|
||||
0.25 / 0.15 = unrelated
|
||||
基本不支撑该句
|
||||
≤0.15 = not_support
|
||||
不支撑
|
||||
|
||||
==================================================
|
||||
【五、输出 JSON（仅 JSON，无 markdown）】
|
||||
{
|
||||
"results": [
|
||||
{
|
||||
"reference_no": 1,
|
||||
"cite_group_refs": "1,2",
|
||||
"is_relevant": 0,
|
||||
"relevance_score": 0.45,
|
||||
"relevance_level": "weakly_related",
|
||||
"relevance_role": "minimal_relevance",
|
||||
"reason": "【中文】中文单条结论\n【English】English single-reference conclusion",
|
||||
"reason_en": "English single-reference conclusion",
|
||||
"combined_is_relevant": 1,
|
||||
"combined_relevance_score": 0.92,
|
||||
"combined_relevance_level": "highly_related",
|
||||
"combined_reason": "【中文】中文联合结论\n【English】English combined conclusion",
|
||||
"combined_reason_en": "English combined conclusion"
|
||||
},
|
||||
{
|
||||
"reference_no": 2,
|
||||
"cite_group_refs": "1,2",
|
||||
...
|
||||
}
|
||||
]
|
||||
}
|
||||
PROMPT;
|
||||
}
|
||||
|
||||
private function buildUserPrompt($sectionText, $localContext, $referText, $abstractText, $citeGroupRefs)
|
||||
{
|
||||
$parts = ["【正文节 t_article_main】\n" . $sectionText];
|
||||
if (trim((string)$citeGroupRefs) !== '') {
|
||||
$mode = strpos($citeGroupRefs, ',') !== false ? '联合引用' : '单独引用';
|
||||
$parts[] = "【引用文献组 cite_group_refs】{$citeGroupRefs}（{$mode}）";
|
||||
}
|
||||
if ($localContext !== '') {
|
||||
$parts[] = "【本引用位置附近上下文（优先据此拆解 claim）】\n" . $localContext;
|
||||
}
|
||||
$parts[] = "【参考文献书目（按编号）】\n" . $referText;
|
||||
if ($abstractText !== '') {
|
||||
$parts[] = "【文献摘要/清洗后内容（Europe PMC·PubMed·Crossref·PDF）】\n" . $abstractText;
|
||||
}
|
||||
$parts[] = '请先拆解最小主张单元（主语层级、证据来源、点名通路/结局逐项核对），判断每篇文献类型与**本文自身证据**，再**逐篇独立**给出单条 relevance_score（讨论转引、提取物/计算预测不得抬高；弱相关文献不得因联合而高分），最后给出 combined_*。reason / combined_reason 必须中英双语（【中文】/【English】），并分别填写 reason_en / combined_reason_en。仅输出 results 数组 JSON。';
|
||||
|
||||
return implode("\n\n", $parts);
|
||||
}
|
||||
|
||||
private function normalizeResults(array $parsed, $defaultCiteGroupRefs, $localContext = '', $referText = '', $abstractText = '')
|
||||
{
|
||||
$rows = [];
|
||||
if (isset($parsed['results']) && is_array($parsed['results'])) {
|
||||
$rows = $parsed['results'];
|
||||
} elseif (isset($parsed['reference_no']) || isset($parsed['relevance_score'])) {
|
||||
$rows = [$parsed];
|
||||
}
|
||||
|
||||
$bands = $this->getScoreBands();
|
||||
$localContext = trim((string)$localContext);
|
||||
$referText = trim((string)$referText);
|
||||
$abstractText = trim((string)$abstractText);
|
||||
|
||||
$out = [];
|
||||
foreach ($rows as $item) {
|
||||
if (!is_array($item)) {
|
||||
continue;
|
||||
}
|
||||
$refNo = intval(isset($item['reference_no']) ? $item['reference_no'] : 0);
|
||||
if ($refNo <= 0) {
|
||||
continue;
|
||||
}
|
||||
|
||||
$score = $this->snapScore(floatval(isset($item['relevance_score']) ? $item['relevance_score'] : 0), $bands);
|
||||
$isRelevant = $score >= 0.65 - 0.001;
|
||||
if (array_key_exists('is_relevant', $item)) {
|
||||
$isRelevant = $this->boolVal($item['is_relevant']);
|
||||
}
|
||||
|
||||
$level = $this->levelFromScore($score, isset($item['relevance_level']) ? $item['relevance_level'] : '');
|
||||
$role = $this->normalizeRelevanceRole(isset($item['relevance_role']) ? $item['relevance_role'] : '');
|
||||
list($reason, $reasonEn) = $this->normalizeBilingualReason(
|
||||
isset($item['reason']) ? $item['reason'] : '',
|
||||
isset($item['reason_en']) ? $item['reason_en'] : ''
|
||||
);
|
||||
|
||||
list($score, $level, $isRelevant, $role) = $this->enforceSingleReferenceConsistency(
|
||||
$score,
|
||||
$level,
|
||||
$isRelevant,
|
||||
$role,
|
||||
$bands
|
||||
);
|
||||
|
||||
$combinedScore = $this->snapScore(
|
||||
floatval(isset($item['combined_relevance_score']) ? $item['combined_relevance_score'] : $score),
|
||||
$bands
|
||||
);
|
||||
$combinedRelevant = $combinedScore >= 0.65 - 0.001;
|
||||
if (array_key_exists('combined_is_relevant', $item)) {
|
||||
$combinedRelevant = $this->boolVal($item['combined_is_relevant']);
|
||||
}
|
||||
|
||||
$combinedLevel = $this->levelFromScore(
|
||||
$combinedScore,
|
||||
isset($item['combined_relevance_level']) ? $item['combined_relevance_level'] : ''
|
||||
);
|
||||
list($combinedScore, $combinedLevel, $combinedRelevant) = $this->enforceCombinedConsistency(
|
||||
$combinedScore,
|
||||
$combinedLevel,
|
||||
$combinedRelevant,
|
||||
$bands
|
||||
);
|
||||
|
||||
$citeGroupRefs = trim((string)(isset($item['cite_group_refs']) ? $item['cite_group_refs'] : $defaultCiteGroupRefs));
|
||||
if ($citeGroupRefs === '' && $defaultCiteGroupRefs !== '') {
|
||||
$citeGroupRefs = trim((string)$defaultCiteGroupRefs);
|
||||
}
|
||||
|
||||
list($combinedReason, $combinedReasonEn) = $this->normalizeBilingualReason(
|
||||
isset($item['combined_reason']) ? $item['combined_reason'] : '',
|
||||
isset($item['combined_reason_en']) ? $item['combined_reason_en'] : ''
|
||||
);
|
||||
if ($combinedReason === '' && $combinedReasonEn === '') {
|
||||
list($combinedReason, $combinedReasonEn) = [$reason, $reasonEn];
|
||||
}
|
||||
|
||||
$out[] = [
|
||||
'reference_no' => $refNo,
|
||||
'cite_group_refs' => $citeGroupRefs,
|
||||
'is_relevant' => $isRelevant ? 1 : 0,
|
||||
'relevance_score' => $score,
|
||||
'relevance_level' => $level,
|
||||
'relevance_role' => $role,
|
||||
'reason' => $reason,
|
||||
'reason_en' => $reasonEn,
|
||||
'combined_is_relevant' => $combinedRelevant ? 1 : 0,
|
||||
'combined_relevance_score' => $combinedScore,
|
||||
'combined_relevance_level' => $combinedLevel,
|
||||
'combined_reason' => $combinedReason,
|
||||
'combined_reason_en' => $combinedReasonEn,
|
||||
];
|
||||
}
|
||||
|
||||
$out = $this->syncCombinedFieldsAcrossGroup($out);
|
||||
|
||||
return $out;
|
||||
}
|
||||
|
||||
private function enforceSingleReferenceConsistency($score, $level, $isRelevant, $role, array $bands)
|
||||
{
|
||||
$score = floatval($score);
|
||||
if ($role === 'no_meaningful_relevance') {
|
||||
if ($score > 0.25) {
|
||||
$score = 0.25;
|
||||
}
|
||||
$level = 'unrelated';
|
||||
$isRelevant = false;
|
||||
} elseif ($role === 'minimal_relevance') {
|
||||
if ($score > 0.45) {
|
||||
$score = 0.45;
|
||||
}
|
||||
$level = 'weakly_related';
|
||||
$isRelevant = false;
|
||||
} elseif ($role === 'supplementary_relevance') {
|
||||
if ($score > 0.78) {
|
||||
$score = 0.78;
|
||||
}
|
||||
$level = $this->levelFromScore($score, $level);
|
||||
} elseif ($role === 'primary_relevance') {
|
||||
if ($score < 0.85) {
|
||||
$score = 0.85;
|
||||
}
|
||||
$isRelevant = true;
|
||||
$level = $this->levelFromScore($score, $level);
|
||||
}
|
||||
|
||||
if ($level === 'weakly_related' && $score > 0.45) {
|
||||
$score = 0.45;
|
||||
$isRelevant = false;
|
||||
} elseif ($level === 'unrelated' && $score > 0.25) {
|
||||
$score = 0.25;
|
||||
$isRelevant = false;
|
||||
} elseif ($level === 'highly_related' && $score < 0.85) {
|
||||
$score = 0.85;
|
||||
$isRelevant = true;
|
||||
} elseif ($level === 'partially_related') {
|
||||
if ($score > 0.78) {
|
||||
$score = 0.78;
|
||||
}
|
||||
if ($score < 0.65) {
|
||||
$score = 0.65;
|
||||
}
|
||||
$isRelevant = true;
|
||||
}
|
||||
|
||||
if (!$isRelevant && $score >= 0.65) {
|
||||
$score = 0.45;
|
||||
$level = 'weakly_related';
|
||||
}
|
||||
if ($isRelevant && $score < 0.65) {
|
||||
$score = 0.65;
|
||||
$level = 'partially_related';
|
||||
}
|
||||
|
||||
$score = $this->snapScore($score, $bands);
|
||||
$level = $this->levelFromScore($score, $level);
|
||||
|
||||
return [$score, $level, $isRelevant, $role];
|
||||
}
|
||||
|
||||
private function enforceCombinedConsistency($combinedScore, $combinedLevel, $combinedRelevant, array $bands)
|
||||
{
|
||||
$combinedScore = $this->snapScore(floatval($combinedScore), $bands);
|
||||
$combinedLevel = $this->levelFromScore($combinedScore, $combinedLevel);
|
||||
$combinedRelevant = $combinedScore >= 0.65 - 0.001;
|
||||
|
||||
return [$combinedScore, $combinedLevel, $combinedRelevant];
|
||||
}
|
||||
|
||||
private function syncCombinedFieldsAcrossGroup(array $out)
|
||||
{
|
||||
$groups = [];
|
||||
foreach ($out as $idx => $row) {
|
||||
$key = (string)$row['cite_group_refs'];
|
||||
if ($key === '') {
|
||||
$key = 'ref:' . $row['reference_no'];
|
||||
}
|
||||
$groups[$key][] = $idx;
|
||||
}
|
||||
|
||||
foreach ($groups as $indices) {
|
||||
if (count($indices) <= 1) {
|
||||
continue;
|
||||
}
|
||||
$bestIdx = $indices[0];
|
||||
$bestScore = floatval($out[$bestIdx]['combined_relevance_score']);
|
||||
foreach ($indices as $idx) {
|
||||
$s = floatval($out[$idx]['combined_relevance_score']);
|
||||
if ($s >= $bestScore) {
|
||||
$bestScore = $s;
|
||||
$bestIdx = $idx;
|
||||
}
|
||||
}
|
||||
$src = $out[$bestIdx];
|
||||
foreach ($indices as $idx) {
|
||||
$out[$idx]['combined_is_relevant'] = intval($src['combined_is_relevant']);
|
||||
$out[$idx]['combined_relevance_score'] = floatval($src['combined_relevance_score']);
|
||||
$out[$idx]['combined_relevance_level'] = (string)$src['combined_relevance_level'];
|
||||
$out[$idx]['combined_reason'] = (string)$src['combined_reason'];
|
||||
$out[$idx]['combined_reason_en'] = (string)$src['combined_reason_en'];
|
||||
}
|
||||
}
|
||||
|
||||
return $out;
|
||||
}
|
||||
|
||||
private function getScoreBands()
|
||||
{
|
||||
return [0.15, 0.25, 0.45, 0.65, 0.78, 0.85, 0.92, 0.98];
|
||||
}
|
||||
|
||||
private function snapScore($score, array $bands)
|
||||
{
|
||||
foreach ($bands as $band) {
|
||||
if (abs($score - $band) < 0.001) {
|
||||
return $band;
|
||||
}
|
||||
}
|
||||
$nearest = $bands[0];
|
||||
$minDiff = abs($score - $nearest);
|
||||
foreach ($bands as $band) {
|
||||
$diff = abs($score - $band);
|
||||
if ($diff < $minDiff) {
|
||||
$minDiff = $diff;
|
||||
$nearest = $band;
|
||||
}
|
||||
}
|
||||
|
||||
return $nearest;
|
||||
}
|
||||
|
||||
private function levelFromScore($score, $levelHint = '')
|
||||
{
|
||||
$levelHint = strtolower(trim((string)$levelHint));
|
||||
$allowed = ['highly_related', 'partially_related', 'weakly_related', 'unrelated'];
|
||||
if (in_array($levelHint, $allowed, true)) {
|
||||
return $levelHint;
|
||||
}
|
||||
$aliases = [
|
||||
'highly_related' => ['highly_related', 'high_related', 'strong_related', 'strong_relevance'],
|
||||
'partially_related' => ['partially_related', 'partial_related', 'moderate_related'],
|
||||
'weakly_related' => ['weakly_related', 'weak_related', 'low_related', 'insufficient'],
|
||||
'unrelated' => ['unrelated', 'not_related', 'irrelevant', 'no_meaningful_relevance'],
|
||||
];
|
||||
foreach ($aliases as $canonical => $list) {
|
||||
if (in_array($levelHint, $list, true)) {
|
||||
return $canonical;
|
||||
}
|
||||
}
|
||||
$score = floatval($score);
|
||||
if ($score >= 0.85) {
|
||||
return 'highly_related';
|
||||
}
|
||||
if ($score >= 0.65) {
|
||||
return 'partially_related';
|
||||
}
|
||||
if ($score >= 0.45) {
|
||||
return 'weakly_related';
|
||||
}
|
||||
|
||||
return 'unrelated';
|
||||
}
|
||||
|
||||
private function normalizeRelevanceRole($role)
|
||||
{
|
||||
$role = strtolower(trim((string)$role));
|
||||
$map = [
|
||||
'primary_relevance' => ['primary_relevance', 'primary_support', 'primary'],
|
||||
'supplementary_relevance' => ['supplementary_relevance', 'supplementary_support', 'supplementary'],
|
||||
'minimal_relevance' => ['minimal_relevance', 'minimal_support', 'minimal'],
|
||||
'no_meaningful_relevance' => ['no_meaningful_relevance', 'no_meaningful_support', 'none'],
|
||||
];
|
||||
foreach ($map as $canonical => $aliases) {
|
||||
if ($role === $canonical || in_array($role, $aliases, true)) {
|
||||
return $canonical;
|
||||
}
|
||||
}
|
||||
|
||||
return 'no_meaningful_relevance';
|
||||
}
|
||||
|
||||
private function cleanReason($reason)
|
||||
{
|
||||
$reason = trim(preg_replace('/[ \t]+/u', ' ', (string)$reason));
|
||||
$reason = trim(preg_replace("/\n{3,}/u", "\n\n", $reason));
|
||||
return mb_substr($reason, 0, 2000);
|
||||
}
|
||||
|
||||
/**
|
||||
* @return array{0:string,1:string} [bilingual reason, english only]
|
||||
*/
|
||||
private function normalizeBilingualReason($reason, $reasonEn)
|
||||
{
|
||||
$reason = trim((string)$reason);
|
||||
$reasonEn = $this->cleanReason($reasonEn);
|
||||
|
||||
if ($reasonEn === '' && preg_match('/【English】\s*(.+)$/us', $reason, $m)) {
|
||||
$reasonEn = $this->cleanReason($m[1]);
|
||||
}
|
||||
|
||||
$zh = '';
|
||||
if (preg_match('/【中文】\s*(.*?)(?:\n【English】|$)/us', $reason, $m)) {
|
||||
$zh = trim($m[1]);
|
||||
} elseif ($reason !== '' && strpos($reason, '【English】') === false) {
|
||||
$zh = trim($reason);
|
||||
}
|
||||
|
||||
if ($zh !== '' && $reasonEn !== '' && strpos($reason, '【English】') === false) {
|
||||
$reason = "【中文】{$zh}\n【English】{$reasonEn}";
|
||||
} elseif ($zh !== '' && $reasonEn !== '' && strpos($reason, '【中文】') === false) {
|
||||
$reason = "【中文】{$zh}\n【English】{$reasonEn}";
|
||||
} else {
|
||||
$reason = $this->cleanReason($reason);
|
||||
}
|
||||
|
||||
if ($reasonEn === '' && $zh !== '') {
|
||||
$reasonEn = '';
|
||||
}
|
||||
|
||||
return [$reason, $reasonEn];
|
||||
}
|
||||
|
||||
private function boolVal($v)
|
||||
{
|
||||
if (is_bool($v)) {
|
||||
return $v;
|
||||
}
|
||||
if (is_numeric($v)) {
|
||||
return intval($v) !== 0;
|
||||
}
|
||||
$s = strtolower(trim((string)$v));
|
||||
return in_array($s, ['1', 'true', 'yes', 'y'], true);
|
||||
}
|
||||
|
||||
private function postChat(array $payload)
|
||||
{
|
||||
$this->lastPostError = '';
|
||||
try {
|
||||
$ch = curl_init();
|
||||
curl_setopt($ch, CURLOPT_URL, $this->url);
|
||||
curl_setopt($ch, CURLOPT_POST, true);
|
||||
curl_setopt($ch, CURLOPT_POSTFIELDS, json_encode($payload, JSON_UNESCAPED_UNICODE));
|
||||
curl_setopt($ch, CURLOPT_RETURNTRANSFER, true);
|
||||
curl_setopt($ch, CURLOPT_CONNECTTIMEOUT, min(15, $this->timeout));
|
||||
curl_setopt($ch, CURLOPT_TIMEOUT, $this->timeout);
|
||||
$headers = ['Content-Type: application/json'];
|
||||
if ($this->apiKey !== '') {
|
||||
$headers[] = 'Authorization: Bearer ' . $this->apiKey;
|
||||
}
|
||||
curl_setopt($ch, CURLOPT_HTTPHEADER, $headers);
|
||||
$raw = curl_exec($ch);
|
||||
if ($raw === false) {
|
||||
$this->lastPostError = 'LLM curl error: ' . curl_error($ch);
|
||||
\think\Log::warning('ReferenceRelevanceLlm: ' . $this->lastPostError);
|
||||
curl_close($ch);
|
||||
return null;
|
||||
}
|
||||
$httpCode = intval(curl_getinfo($ch, CURLINFO_HTTP_CODE));
|
||||
curl_close($ch);
|
||||
if ($httpCode < 200 || $httpCode >= 300) {
|
||||
$snippet = mb_substr(trim((string)$raw), 0, 200);
|
||||
$this->lastPostError = 'LLM HTTP ' . $httpCode . ($snippet !== '' ? ': ' . $snippet : '');
|
||||
\think\Log::warning('ReferenceRelevanceLlm: ' . $this->lastPostError);
|
||||
return null;
|
||||
}
|
||||
$data = json_decode($raw, true);
|
||||
if (!is_array($data)) {
|
||||
$this->lastPostError = 'LLM response is not valid JSON';
|
||||
return null;
|
||||
}
|
||||
if (isset($data['choices'][0]['message']['content'])) {
|
||||
return (string)$data['choices'][0]['message']['content'];
|
||||
}
|
||||
if (isset($data['content'])) {
|
||||
return (string)$data['content'];
|
||||
}
|
||||
$this->lastPostError = 'LLM response missing content field';
|
||||
} catch (\Exception $e) {
|
||||
$this->lastPostError = 'LLM exception: ' . $e->getMessage();
|
||||
\think\Log::warning('ReferenceRelevanceLlm: ' . $this->lastPostError);
|
||||
}
|
||||
|
||||
return null;
|
||||
}
|
||||
|
||||
private function parseJson($raw)
|
||||
{
|
||||
$raw = trim((string)$raw);
|
||||
if ($raw === '') {
|
||||
return null;
|
||||
}
|
||||
$raw = preg_replace('/^```[a-zA-Z]*\s*|```$/m', '', $raw);
|
||||
$raw = trim($raw);
|
||||
$decoded = json_decode($raw, true);
|
||||
if (is_array($decoded)) {
|
||||
return $decoded;
|
||||
}
|
||||
if (preg_match('/\{[\s\S]*\}/', $raw, $m)) {
|
||||
$decoded = json_decode($m[0], true);
|
||||
if (is_array($decoded)) {
|
||||
return $decoded;
|
||||
}
|
||||
}
|
||||
|
||||
return null;
|
||||
}
|
||||
}
|
||||
Reference in New Issue
Block a user