本文是Xapian检索过程的分析,本文内容中源码比较多。检索过程,总的来说就是拉取倒排链,取得合法doc,然后做打分排序的过程。
1 理论分析
1.1 检索语法
面对不同的检索业务,我们会有多种检索需求,譬如:要求A term和B term都在Doc中出现;要求A term或者B term任意在Doc中出现;要求A term或者B term任意在Doc出现,并且C term不出现…...,用符号表示:
A & B
A || B
(A || B) & ~C
( A & ( B || C ) ) || D
…
以上的种种检索需求,复杂繁多,每一个检索需求都单独实现一份代码,是不现实的,需要有一种简单、高效、可扩展的检索语法来支持他们。
1.2 检索过程
首先是根据业务需求,组装检索语句,然后调用检索内核提供的API获取检索结果。
检索内核的实现,以xapian为例:首先根据用户组装的检索语句形成query-tree(query检索树),然后将query-tree转换为postlist-tree(倒排链树),最后获取postlist-tree运算后的结果。在获取postlist-tree和最后的计算过程中,穿插着相关性公式(如:BM25)的运算。
1.3 相关性
计算query跟doc相关性方式有好几种,
(1) 布尔模型(Boolean Model)
判断用户的term在不在文档中出现,如果出现了则认为文档跟用户需求相关,否则认为不相关。
优点:简单;
缺点:结果是二元的,只有YES 或者 NO, 多条结果之间没有先后顺序;
(2)向量空间模型(Vector Space Model)
将query和doc都向量化,计算query跟doc的余弦值,这个值就是query跟doc的相似性打分。这里将查询跟文档的内容相似性替换相关性。
这个模型对长文本比较抑制。
consine公式:向量点积 / 向量长度相乘。
那么,怎么向量化?每一纬的值,给多少合适?
词频因子(TF):某个单词在文档中出现的次数;一般取log做平滑,避免直接使用词频导致出现1次和出现10次的term权重差异过大。 常见公式: Wtf = 1 + log(TF). 常量1是为了避免TF=1时,log(TF) = 0,导致W变成0。
变体公式: Wtf = a + (1 - a) * TF/Max(TF),其中a是调节因子,取值0.4或者0.5,TF表示词频,Max(TF)表示文档中出现次数最多的单词对应的词频数目。这个变种有利于抑制长文本,使得不同长度文档的词频因子具有可比性。
逆文档频率因子(IDF):包含有某个词的文档数量的倒数。如果一个词在所有文档中都出现,那么这个词对文档的区分度贡献不高,不是那么重要,反之,则说明这个词很重要。
公式: IDFk = log(N/nk), N代表文档集合总共有多少个文档;nk代表词在多少个文档中出现过。
TF*IDF框架:
Weight = TF * IDF
(3)概率检索模型
BIM模型的公式,由四个部分组成,这四个部分可以理解为:
1、含有某term的doc在相关集合中出现的次数,正面因素;
2、不含有某term的doc在相关集合中出现的次数,负面因素;
3、含有某term的doc在不相关集合中出现的次数,负面因素;
4、不含有某term的doc在不相关集合中不出现的次数,正面因素。
BM25公式,三部分:1、BIM模型,等价于IDF;2、term在文档中的权重(doc-tf);3、term在query中的权重(query-tf);
N,表示索引中总的文档数,
Ni,表示索引中包含有term的文档数,也就是df,
fi,表示term在文档中出现的次数,
qfi,表示term在query中出现的次数,
dl,表示文档长度,
avdl,表示平均文档长度
BM25F
考虑到不同的域,对第二部分的平均长度、调节因子,需要根据不同的域设置不同的值,并且需要一个跟域相关联的权重值。
相关性部分资料参考:《这就是搜索引擎》
2 源码分析
2.1 主要类
下面以xapian为例,介绍一般检索过程,因涉及源码众多,部分枝节策略不一一细说。首先,这里列出,涉及到的主要类,从这里也可以一窥xapian在检索上的设计思路。
绿色背景的块是用户看到的,蓝色背景是其底层涉及到的。
Enquire::Internal,Enquire的内部实现,Xapian的设计风格都是包一层壳,功能实际的实现放在Internal中;
BM25Weight,Xapian默认使用的相关性打分类;
Weight::Internal,打分需要用到的基础信息,譬如:索引库文档量、索引库总的term长度、query里的 term的tf、df数据…;
MultiMatch,检索的实现类;
LocalSubMatch,本地子索引库操作的封装。 xapian支持远程索引库,也支持一个索引库拆分成多个子索引库;
QueryOptimiser ,从Query-Tree构建PostList-Tree时的帮助类,主要记录了一些子索引库相关的信息,譬如:LocalSubMatch的引用、索引库DataBase的引用…;
QueryOr、QueryBranch、QueryTerm ,这系列是Query Tree上的一个个类;
PostList、LeafPostList,PostList-Tree上的一个个类;
InMemoryPostList,内存索引库的PostList封装;
OrContext,记录在Query-Tree转PostList-Tree过程中的PostList上下文信息,包括:QueryOptimiser对象指针、临时存放的PostList指针;
2.2 检索过程
2.2.1 用户demo代码
Xapian::Query term_one = Xapian::Query("T世界"); Xapian::Query term_two = Xapian::Query("T比赛"); Xapian::Query query = Xapian::Query(Xapian::Query::OP_OR, term_one, term_two); // query组装 std::cout << "query=" << query.get_description() << std::endl; Xapian::Enquire enquire(db); enquire.set_query(query); Xapian::MSet result = enquire.get_mset(0, 10); // 执行检索,获取结果 std::cout << "find results count=" << result.get_matches_estimated() << std::endl; for (auto it = result.begin(); it != result.end(); ++it) { Xapian::Document doc = it.get_document(); std::string data = doc.get_data(); double doc_score_weight = it.get_weight(); int doc_score_percent = it.get_percent(); std::cout << "doc=" << data << ",weight=" << doc_score_weight << ",percent=" << doc_score_percent << std::endl; }
2.2.2 query组装的实现
只有一个类——Query,通过构造函数重载,提供了一切需要的功能。
eg:
Query::Query(const string & term, Xapian::termcount wqf, Xapian::termpos pos) : internal(new Xapian::Internal::QueryTerm(term, wqf, pos)) { LOGCALL_CTOR(API, "Query", term | wqf | pos); }
Query(op op_, const Xapian::Query & a, const Xapian::Query & b) { init(op_, 2); bool positional = (op_ == OP_NEAR || op_ == OP_PHRASE); add_subquery(positional, a); add_subquery(positional, b); done(); }
/* 根据OP,生成对应的Query派生类,譬如:or的生成 QueryOr类,含有两个子query,这个QueryOr类对象作为Query的internal成员存在;
在组合多个query时,直接添加到vector中;
如果最后发现vector是空的则将internal设置为NULL,或者=1,则将internal设置为子query的internal,这样子可以避免不必要的vector嵌套,如:[xxquery],单个元素没必要放在vector中。*/
...
检索树的组织没有做特别的设计,譬如:用vector来存储OR的元素。
2.2.3 检索的实现
(1)检索函数入口:
MSet Enquire::Internal::get_mset(Xapian::doccount first, Xapian::doccount maxitems, Xapian::doccount check_at_least, const RSet *rset, const MatchDecider *mdecider) const { LOGCALL(MATCH, MSet, "Enquire::Internal::get_mset", first | maxitems | check_at_least | rset | mdecider); if (percent_cutoff && (sort_by == VAL || sort_by == VAL_REL)) { throw Xapian::UnimplementedError("Use of a percentage cutoff while sorting primary by value isn't currently supported"); } if (weight == 0) { weight = new BM25Weight; // 如果外界没有指定打分策略,采用BM25Weight } Xapian::doccount first_orig = first; { Xapian::doccount docs = db.get_doccount(); first = min(first, docs); maxitems = min(maxitems, docs - first); check_at_least = min(check_at_least, docs); check_at_least = max(check_at_least, first + maxitems); } AutoPtr<Xapian::Weight::Internal> stats(new Xapian::Weight::Internal); // 用于记录打分用的全局信息
// MultiMatch对象的初始化,会执行检索的初始化工作,譬如:填充stats对象,
::MultiMatch match(db, query, qlen, rset, collapse_max, collapse_key, percent_cutoff, weight_cutoff, order, sort_key, sort_by, sort_value_forward, time_limit, *(stats.get()), weight, spies, (sorter.get() != NULL), (mdecider != NULL)); // Run query and put results into supplied Xapian::MSet object. MSet retval; match.get_mset(first, maxitems, check_at_least, retval, *(stats.get()), mdecider, sorter.get()); // 检索 if (first_orig != first && retval.internal.get()) { retval.internal->firstitem = first_orig; } Assert(weight->name() != "bool" || retval.get_max_possible() == 0); // The Xapian::MSet needs to have a pointer to ourselves, so that it can // retrieve the documents. This is set here explicitly to avoid having // to pass it into the matcher, which gets messy particularly in the // networked case. retval.internal->enquire = this; if (!retval.internal->stats) { retval.internal->stats = stats.release(); } RETURN(retval); }
(2)检索之前的准备工作,在 MultiMatch 对象构造的时候做,prepare_sub_matches:
static void prepare_sub_matches(vector<intrusive_ptr<SubMatch> > & leaves, Xapian::Weight::Internal & stats) { LOGCALL_STATIC_VOID(MATCH, "prepare_sub_matches", leaves | stats); // We use a vector<bool> to track which SubMatches we're already prepared. vector<bool> prepared; prepared.resize(leaves.size(), false); size_t unprepared = leaves.size(); bool nowait = true; while (unprepared) { for (size_t leaf = 0; leaf < leaves.size(); ++leaf) { if (prepared[leaf]) continue; SubMatch * submatch = leaves[leaf].get(); if (!submatch || submatch->prepare_match(nowait, stats)) { prepared[leaf] = true; --unprepared; } } // Use blocking IO on subsequent passes, so that we don't go into // a tight loop. nowait = false; } }
bool LocalSubMatch::prepare_match(bool nowait, Xapian::Weight::Internal & total_stats) { LOGCALL(MATCH, bool, "LocalSubMatch::prepare_match", nowait | total_stats); (void)nowait; Assert(db); total_stats.accumulate_stats(*db, rset); RETURN(true); }
void Weight::Internal::accumulate_stats(const Xapian::Database::Internal &subdb, const Xapian::RSet &rset) { #ifdef XAPIAN_ASSERTIONS Assert(!finalised); ++subdbs; #endif total_length += subdb.get_total_length(); collection_size += subdb.get_doccount(); rset_size += rset.size(); total_term_count += subdb.get_doccount() * subdb.get_total_length(); Xapian::TermIterator t; for (t = query.get_unique_terms_begin(); t != Xapian::TermIterator(); ++t) { const string & term = *t; Xapian::doccount sub_tf; Xapian::termcount sub_cf; subdb.get_freqs(term, &sub_tf, &sub_cf); TermFreqs & tf = termfreqs[term]; tf.termfreq += sub_tf; tf.collfreq += sub_cf; } const set<Xapian::docid> & items(rset.internal->get_items()); set<Xapian::docid>::const_iterator d; for (d = items.begin(); d != items.end(); ++d) { Xapian::docid did = *d; Assert(did); // The query is likely to contain far fewer terms than the documents, // and we can skip the document's termlist, so look for each query term // in the document. AutoPtr<TermList> tl(subdb.open_term_list(did)); map<string, TermFreqs>::iterator i; for (i = termfreqs.begin(); i != termfreqs.end(); ++i) { const string & term = i->first; TermList * ret = tl->skip_to(term); Assert(ret == NULL); (void)ret; if (tl->at_end()) { break; } if (term == tl->get_termname()) { ++i->second.reltermfreq; } } } }
prepare_sub_matches(): BM25计算之前的准备工作
Weight::Internal::accumulate_stats:
total_length:db的总文档长度加和;
collection_size:db的总文档数量;
total_term_count: 存疑,变量名是term计数,实际上是总文档长度加和 * 总文档数量;
termfreqs: term的tf信息(term在多少个doc中出现)和cf信息(term在索引集合中出现的次数);
query中涉及到的所有term,都获取到它们的TF、IDF信息;
极致的压缩:VectorTermList,把几个string存储的term压缩存储到一个块内存中。如果使用vector来存储,则会增加30Byte每一个term。
(3)打开倒排链,构造postlist-tree:
打开倒排链和检索放在一个800行的超大函数里面:
void MultiMatch::get_mset(Xapian::doccount first, Xapian::doccount maxitems, Xapian::doccount check_at_least, Xapian::MSet & mset, Xapian::Weight::Internal & stats, const Xapian::MatchDecider *mdecider, const Xapian::KeyMaker *sorter) { ........ }
打开倒排链的过程,函数多层嵌套非常深入,这也是检索树解析-->重建过程:
PostList * LocalSubMatch::get_postlist(MultiMatch * matcher, Xapian::termcount * total_subqs_ptr) { LOGCALL(MATCH, PostList *, "LocalSubMatch::get_postlist", matcher | total_subqs_ptr); if (query.empty()) { RETURN(new EmptyPostList); // MatchNothing } // Build the postlist tree for the query. This calls // LocalSubMatch::open_post_list() for each term in the query. PostList * pl; { QueryOptimiser opt(*db, *this, matcher); pl = query.internal->postlist(&opt, 1.0); *total_subqs_ptr = opt.get_total_subqs(); } AutoPtr<Xapian::Weight> extra_wt(wt_factory->clone()); // Only uses term-independent stats. extra_wt->init_(*stats, qlen); if (extra_wt->get_maxextra() != 0.0) { // There's a term-independent weight contribution, so we combine the // postlist tree with an ExtraWeightPostList which adds in this // contribution. pl = new ExtraWeightPostList(pl, extra_wt.release(), matcher); } RETURN(pl); }
PostingIterator::Internal * QueryOr::postlist(QueryOptimiser * qopt, double factor) const { LOGCALL(QUERY, PostingIterator::Internal *, "QueryOr::postlist", qopt | factor); OrContext ctx(qopt, subqueries.size()); do_or_like(ctx, qopt, factor); RETURN(ctx.postlist()); }
void QueryBranch::do_or_like(OrContext& ctx, QueryOptimiser * qopt, double factor, Xapian::termcount elite_set_size, size_t first) const { LOGCALL_VOID(MATCH, "QueryBranch::do_or_like", ctx | qopt | factor | elite_set_size); // FIXME: we could optimise by merging OP_ELITE_SET and OP_OR like we do // for AND-like operations. // OP_SYNONYM with a single subquery is only simplified by // QuerySynonym::done() if the single subquery is a term or MatchAll. Assert(subqueries.size() >= 2 || get_op() == Query::OP_SYNONYM); vector<PostList *> postlists; postlists.reserve(subqueries.size() - first); QueryVector::const_iterator q; for (q = subqueries.begin() + first; q != subqueries.end(); ++q) { // MatchNothing subqueries should have been removed by done(). Assert((*q).internal.get()); (*q).internal->postlist_sub_or_like(ctx, qopt, factor); } if (elite_set_size && elite_set_size < subqueries.size()) { ctx.select_elite_set(elite_set_size, subqueries.size()); // FIXME: not right! } }
...
LeafPostList * LocalSubMatch::open_post_list(const string& term, Xapian::termcount wqf, double factor, bool need_positions, bool in_synonym, QueryOptimiser * qopt, bool lazy_weight) { LOGCALL(MATCH, LeafPostList *, "LocalSubMatch::open_post_list", term | wqf | factor | need_positions | qopt | lazy_weight); bool weighted = (factor != 0.0 && !term.empty()); LeafPostList * pl = NULL; if (!term.empty() && !need_positions) { if ((!weighted && !in_synonym) || !wt_factory->get_sumpart_needs_wdf_()) { Xapian::doccount sub_tf; db->get_freqs(term, &sub_tf, NULL); if (sub_tf == db->get_doccount()) { // If we're not going to use the wdf or term positions, and the // term indexes all documents, we can replace it with the // MatchAll postlist, which is especially efficient if there // are no gaps in the docids. pl = db->open_post_list(string()); // Set the term name so the postlist looks up the correct term // frequencies - this is necessary if the weighting scheme // needs collection frequency or reltermfreq (termfreq would be // correct anyway since it's just the collection size in this // case). pl->set_term(term); } } } if (!pl) { const LeafPostList * hint = qopt->get_hint_postlist(); if (hint) pl = hint->open_nearby_postlist(term); if (!pl) pl = db->open_post_list(term); qopt->set_hint_postlist(pl); } if (lazy_weight) { // Term came from a wildcard, but we may already have that term in the // query anyway, so check before accumulating its TermFreqs. map<string, TermFreqs>::iterator i = stats->termfreqs.find(term); if (i == stats->termfreqs.end()) { Xapian::doccount sub_tf; Xapian::termcount sub_cf; db->get_freqs(term, &sub_tf, &sub_cf); stats->termfreqs.insert(make_pair(term, TermFreqs(sub_tf, 0, sub_cf))); } } if (weighted) { Xapian::Weight * wt = wt_factory->clone(); if (!lazy_weight) { wt->init_(*stats, qlen, term, wqf, factor); // BM25Weight::init()计算不涉及query跟doc相关性部分的打分(只跟term和query相关) stats->set_max_part(term, wt->get_maxpart()); } else { // Delay initialising the actual weight object, so that we can // gather stats for the terms lazily expanded from a wildcard // (needed for the remote database case). wt = new LazyWeight(pl, wt, stats, qlen, wqf, factor); } pl->set_termweight(wt); } RETURN(pl); }
weight的init:
void BM25Weight::init(double factor) { Xapian::doccount tf = get_termfreq(); double tw = 0; if (get_rset_size() != 0) { Xapian::doccount reltermfreq = get_reltermfreq(); // There can't be more relevant documents indexed by a term than there // are documents indexed by that term. AssertRel(reltermfreq,<=,tf); // There can't be more relevant documents indexed by a term than there // are relevant documents. AssertRel(reltermfreq,<=,get_rset_size()); Xapian::doccount reldocs_not_indexed = get_rset_size() - reltermfreq; // There can't be more relevant documents not indexed by a term than // there are documents not indexed by that term. AssertRel(reldocs_not_indexed,<=,get_collection_size() - tf); Xapian::doccount Q = get_collection_size() - reldocs_not_indexed; Xapian::doccount nonreldocs_indexed = tf - reltermfreq; double numerator = (reltermfreq + 0.5) * (Q - tf + 0.5); double denom = (reldocs_not_indexed + 0.5) * (nonreldocs_indexed + 0.5); tw = numerator / denom; } else { tw = (get_collection_size() - tf + 0.5) / (tf + 0.5); } AssertRel(tw,>,0); // The "official" formula can give a negative termweight in unusual cases // (without an RSet, when a term indexes more than half the documents in // the database). These negative weights aren't actually helpful, and it // is common for implementations to replace them with a small positive // weight or similar. // // Truncating to zero doesn't seem a great approach in practice as it // means that some terms in the query can have no effect at all on the // ranking, and that some results can have zero weight, both of which // are seem surprising. // // Xapian 1.0.x and earlier adjusted the termweight for any term indexing // more than a third of documents, which seems rather "intrusive". That's // what the code currently enabled does, but perhaps it would be better to // do something else. (FIXME) #if 0 if (rare(tw <= 1.0)) { termweight = 0; } else { termweight = log(tw) * factor; if (param_k3 != 0) { double wqf_double = get_wqf(); termweight *= (param_k3 + 1) * wqf_double / (param_k3 + wqf_double); } } #else if (tw < 2) tw = tw * 0.5 + 1; termweight = log(tw) * factor; if (param_k3 != 0) { double wqf_double = get_wqf(); termweight *= (param_k3 + 1) * wqf_double / (param_k3 + wqf_double); } #endif termweight *= (param_k1 + 1); LOGVALUE(WTCALC, termweight); if (param_k2 == 0 && (param_b == 0 || param_k1 == 0)) { // If k2 is 0, and either param_b or param_k1 is 0 then the document // length doesn't affect the weight. len_factor = 0; } else { len_factor = get_average_length(); // len_factor can be zero if all documents are empty (or the database // is empty!) if (len_factor != 0) len_factor = 1 / len_factor; } LOGVALUE(WTCALC, len_factor); }
总的来说,这一阶段:
stats设置给LocalSubMatch对象;
获取倒排列表,根据query-tree构建postlist-tree;同时,clone一个Weight对象,计算BM25所需要的计算因子;平均文档长度,文档的最短长度,term最大的wdf(term在某doc中出现的次数);
计算BM25公式的idf部分:tw = (get_collection_size() - tf + 0.5) / (tf + 0.5); termweight = log(tw) * factor;
计算BM25公式的term在query中的权重部分:double wqf_double = get_wqf(); termweight *= (param_k3 + 1) * wqf_double / (param_k3 + wqf_double);
计算BM25公式的term跟doc相关程度的一部分参数: termweight *= (param_k1 + 1);
计算BM25公式的平均长度分之一:len_factor = 1 / len_factor;
计算maxpart() ,BM25算法,没有地方用这个值;
这就把BM25公式中,不跟具体doc相关的第一和第三部分计算完成。
构建postlist-tree,如果是And的语法,则使用PostList * AndContext::postlist() 生成postlist,然后把子postlist-tree销毁掉;
(4)最终召回排序
循环从postlist-tree拉取docid,然后计算BM25打分,
倒排与链求交过程:
PostList * MultiAndPostList::find_next_match(double w_min)
两个有序链表求交:
0、第一个链表pos往前走一步;
1、取出第一个链表的元素;
2、find_next_match() --> check_helper() 将第二链表的pos往前走,保证第二链表当前位置大于等于第一链表;
3、取出来第二链表的当前元素,跟第一链表原始做比较;
4、如果不匹配则让第一链表往前走;
注:拉链法。
主要代码如下:
/// 注:在调用这个函数之前会先调用next_helper函数,将第一条链表的位置向前移动一位(如果是首次调用则不移动),
/// find_next_match函数让plist[0]倒排链定位到合适的位置,当能定位到合适的位置(plist[0]和plist[1]有交集)则返回,
/// 否则说明没有交集,设置did=0后返回;调用者会通过判断did==0来确定当前链表交集是否已经做完;
PostList * MultiAndPostList::find_next_match(double w_min) { advanced_plist0: if (plist[0]->at_end()) { did = 0; return NULL; } did = plist[0]->get_docid(); for (size_t i = 1; i < n_kids; ++i) { bool valid; check_helper(i, did, w_min, valid); if (!valid) { next_helper(0, w_min); goto advanced_plist0; } if (plist[i]->at_end()) { did = 0; return NULL; } Xapian::docid new_did = plist[i]->get_docid(); if (new_did != did) {
/// 两条链表的pos元素不相等,只可能是因为plist[0].pos的元素比较小,需要向前移 skip_to_helper(0, new_did, w_min); goto advanced_plist0; } } return NULL; }
获取BM25打分:
double LeafPostList::get_weight() const { if (!weight) return 0; Xapian::termcount doclen = 0, unique_terms = 0; // Fetching the document length and number of unique terms is work we can // avoid if the weighting scheme doesn't use them. if (need_doclength) doclen = get_doclength(); if (need_unique_terms) unique_terms = get_unique_terms(); double sumpart = weight->get_sumpart(get_wdf(), doclen, unique_terms); // 这里对某个doc的最终BM25打分做了汇总,利用到了前面计算的第一和第三部分打分 AssertRel(sumpart, <=, weight->get_maxpart()); return sumpart; }
两个有序链表求并:
PostList * OrPostList::next(double w_min)
两个链表都取,在get_docid()时取最小did;如果其中一条倒排链已经取完,则用剩下的链替换之前两条链的owner。
percent是怎么计算的?
percent_scale = greatest_wt_subqs_matched / double(total_subqs);
percent_scale /= greatest_wt;
首先跟命中词个数占总搜索term个数有关系,然后,跟最大的匹配得分有关系,percent_scale会作为percent的因子:
double v = wt * percent_factor + 100.0 * DBL_EPSILON; // percent_scale就是percent_factor,v就是percent
从BM25打分的执行过程,可以想到,有部分BM25打分因子(第一部分idf因子、第二部分term-doc相关性因子)是不需要在线计算的,只需要离线计算后并存储在倒排中即可。
当前默认使用的BM25Weight打分策略,没有使用get_maxextra函数和get_sumextra函数。
percent更详细的介绍可以看这里:https://www.cnblogs.com/cswuyg/p/10552564.html
最终召回结果怎么做limit截断?
当用户只需要n条,而召回结果大于n条,在处理n+1条时,使用std::make_heap,构造堆(如果之前已经构造了,则不需要再构造,直接往堆里加元素),并弹出打分最小的doc,保证只有n条资源。另外,程序还记录了min_weight,当资源打分小于min_weight,则直接丢弃,不需要走构建堆的过程。(详细源码见 multimatch.cc,746行起)