Lecture 3: Inverted File Index（倒排索引）

主线：term → documents（词项定位文档）→ scalable indexing（可扩展建索引）→ relevance evaluation（相关性评价）。 Search trees/hash tables locate dictionary entries; posting lists identify matching documents.

Course overview（课程信息） · HW3: Document Distance（文档距离）

1. Structure and Boolean retrieval（结构与布尔检索）

1.1 From scanning to an inverted index（从扫描到倒排）

A search engine has three broad stages: crawl（爬取）→ index（整理并建索引）→ query service（检索服务）. A spider / crawler（爬虫） collects pages; indexing prepares them for repeated queries.

Approach	Representation	Cost / limitation
Scan every page（逐页扫描）	Search raw text for each query.	Query work scales with the whole collection's text size.
Term-document incidence matrix（词项—文档关联矩阵）	\(A[t,d]=1\) iff document \(d\) contains term \(t\).	Boolean operations are easy, but a dense \(V\times N\) matrix wastes space on zeros.
Inverted file index（倒排文件索引）	Dictionary entry → list of documents containing the term.	Store nonzero associations; process relevant postings rather than every page.

\(V\) = vocabulary size（词汇表规模）, \(N\) = document count（文档数）. Incidence rows can be bitsets（位集）; sparsity motivates postings, but a frequent term can also suit a bitmap representation.

Index（索引）： a mechanism for locating a term in text. Forward index（正排索引）： document → terms. Inverted index（倒排索引）： term → documents/occurrences. Posting list（倒排列表／出现记录表）： the records associated with one term; pointers can be document IDs, offsets or positions, not necessarily memory addresses or linked-list nodes.

WK3 T/F4: “Inverted file contains a list of pointers to all occurrences of a term in the text.” True, by the definition; a document-only index stores document references, and a positional index adds exact occurrence locations.

1.2 Course example, positions and frequencies（课件例子、位置与频率）

Document	Original text
D1	Gold silver truck
D2	Shipment of gold damaged in a fire
D3	Delivery of silver arrived in a silver truck
D4	Shipment of gold arrived in a truck

For this example only: case folding（大小写归一化）, no stemming or stop filtering; positions start at 1 in the original token sequence.

Term（词项）	Incidence D1–D4	\(df\)	Positional postings: docID: positions
a	0111	3	2:6; 3:6; 4:6
arrived	0011	2	3:4; 4:4
damaged	0100	1	2:4
delivery	0010	1	3:1
fire	0100	1	2:7
gold	1101	3	1:1; 2:3; 4:3
of	0111	3	2:2; 3:2; 4:2
in	0111	3	2:5; 3:5; 4:5
shipment	0101	2	2:1; 4:1
silver	1010	2	1:2; 3:3,7
truck	1011	3	1:3; 3:8; 4:7

• Document frequency, \(df(t)\)（文档频率）： number of documents containing \(t\); posting-list length when one record is stored per document.
• Term frequency, \(tf(t,d)\)（词频）： occurrence count of \(t\) within document \(d\). For silver, \(df=2\), but \(tf(\text{silver},D3)=2\) and total occurrences = 3.
• The slide's “Times” uses \(df\), not total occurrence count. “低频更重要”和“先处理短列表”中的频率主要指 \(df\)。

Discussion: how to print matching sentences and highlight words? Keep positions（词位置）and access to original text; character offsets/sentence boundaries（字符偏移／句界）help reconstruct snippets（摘要片段）and highlights（高亮）. A list of document IDs alone cannot locate the matched sentence. Preserve original positions when filtering tokens if phrase/proximity queries（短语／邻近查询）are required.

1.3 Query operations and processing order（查询运算与处理顺序）

• AND（与）： intersect document sets. silver AND truck: 1010 & 1011 = 1010 → D1, D3.
• OR（或）： union document sets; NOT（非） / exclusion: difference against a specified collection or candidate set.
• Phrase query（短语查询）: after finding common documents, require compatible positions. “silver truck” needs a truck at position \(p+1\) after silver; D3 matches at positions 7,8, not 3,8. Boolean AND alone does not enforce adjacency.
• Sorted lists of lengths \(a,b\) can be intersected by two pointers in \(O(a+b)\) time: advance the smaller docID, emit equal IDs, stop when either list ends. Skip pointers（跳跃指针）can bypass ranges that cannot match. Intersection reference.

Discussion: why keep frequency? For a conjunctive query, start with the smallest \(df\) to reduce intermediate candidates（中间候选集）. truck AND silver AND delivery: lengths 3,2,1; start with delivery's [D3], then check the other two → D3. Frequency also informs ranking; list length and within-document frequency serve different roles.

2. Index construction and modules（构建与模块）

2.1 Generator pipeline（索引生成流程）

For each document D:
    parse tokens; normalize them into terms
    apply the selected stop filter
    for each retained term T and original position p:
        find T in the dictionary; insert it if absent
        get T's posting list
        create/update D's posting: frequency + positions
write the inverted index to disk

Repeated occurrences in a document update that document's posting; do not accidentally increase \(df\) on every occurrence. Building an index and scoring a query are different stages.

Module	Responsibility（职责）
Token analyzer（分词／词元分析器） + stop filter（停用词过滤器）	Parse text, normalize forms, decide retained terms.
Vocabulary scanner（词汇表查找器）	Find the dictionary entry for a term.
Vocabulary inserter（词汇表插入器）	Create an entry/posting list for a new term.
Memory management（内存管理）	Buffer index data, flush to disk, merge partial indexes.

Quiz 3.1 / WK3 MC2: which is NOT an index-building step?

• A. Read strings and parse words.
• B. Use stemming and stop filtering to obtain terms.
• C. Check the dictionary and insert missing terms.
• D. Get each term's posting list and calculate precision.

Answer: D. Getting/updating postings is part of construction; precision requires query results and relevance judgments（相关性标注）, not just the index.

2.2 Tokenization, stemming and stop words（分词、词干提取与停用词）

Tokenization（分词）： identify token boundaries. English whitespace/punctuation are useful cues; Chinese needs segmentation（词语切分）because boundaries are not explicit. The offline name-search example illustrates that matching the characters “陈越” inside an unrelated phrase is not equivalent to identifying a person's name.

Stemming（词干提取）： map morphological variants（词形变体）to a common stem. process / processing / processes / processed → process is the course example. Apply the same normalization to indexed documents and queries（索引端与查询端规则一致）.

词干不一定是字典中的完整词。 The slide's say / says / said / saying → say expresses the desired grouping; regular suffix stripping alone need not recognize said. Lemmatization（词形还原）uses vocabulary/morphology; Porter stemming is a suffix-based heuristic, not a general irregular-verb dictionary. Stemming versus lemmatization.

Stop words（停用词）： very common words, e.g. a, the, it, often have little discriminative value（区分能力）and huge postings. Filtering saves space/work; it is a configurable policy, not a claim that those words never matter. In the offline example, a dictionary article specifically explaining “a” should remain retrievable. Context, phrase support and the application determine retention; “remove” here means exclude from indexing, not destroy source documents.

Quiz 3.2 Q3 / WK3 T/F1: “Word stemming eliminates commonly used words.” False: that describes stop filtering. HW3 explicitly retains stop words; do not reuse the lecture's optional filter blindly.

2.3 Dictionary access（词典访问）

Method	Strength	Limitation / qualification
B/B+ trees（多路搜索树）	Ordered traversal and efficient term ranges; page-oriented layout.	Exact lookup traverses the tree; constants depend on layout.
Tries（字典树／前缀树）	Branch on characters; natural prefix access.	Memory/layout and alphabet size matter.
Hashing（哈希）	Expected fast exact-word access with suitable hashing/load control.	No lexical order; range scan generally needs examining many entries or an extra ordered index.

For a large disk-resident dictionary, the course favors page-oriented B/B+ trees over binary AVL/RBT layouts. Its “RBT is 10–20% faster than AVL with many updates” is a workload-dependent observation, not an asymptotic guarantee; layout, caching and scale determine practical speed.

MOOC 3.2.2 discussion: hashing versus search trees? Faster for a single word in the course's expected-cost model, but expensive for range search. Hash lookup still processes the key and can suffer collisions; “faster” is not a worst-case or universal timing guarantee.

Source	Statement	Answer
Quiz 3.2 Q1 / WK3 T/F2	B+ dictionary range searches are expensive.	False: ordered keys/leaf links support efficient range access.
Quiz 3.2 Q2	Hashing is faster than search trees for term access.	True in the course's ideal/expected exact-lookup model, with the qualifications above.
HW3 T/F4	Hash-based dictionary range searches are expensive.	True without an additional ordered structure.

Misspelling suggestions（拼写纠错）need candidate generation and similarity checking. Ordered/prefix structures help generate candidates; 字典序相邻不等于编辑距离最小. Hashing can also be combined with separate correction indexes. The offline narrow-domain example motivates tuning hashing for a small vocabulary, not assuming every dictionary must use a B+ tree.

2.4 Memory-limited construction（内存受限构建）

Build an in-memory block index（分块索引）; when capacity is near its limit, sort/write it to disk, free buffers, and continue. Flush the final nonempty block（别漏最后未满的一块）. Merge sorted blocks by term and docID; combine duplicate document postings, frequencies and positions as appropriate.

blocks = []
for each parsed occurrence:
    if the next update would exceed memory: flush current block
    update the in-memory dictionary/postings
flush the remaining block
merge all sorted blocks into the final index

Sorted order permits sequential merging; blocking bounds working memory instead of requiring the whole index in RAM. Terms appearing across blocks must map to the same final term; do not treat unrelated block-local IDs as globally unique.

3. Scaling, updates and compression（扩展、更新与压缩）

3.1 Distributed indexing（分布式索引）

A cluster node（集群节点）stores an index of part of the collection. Two partition directions:

Strategy	Local contents	Query / scaling tradeoff
Term-partitioned（按词项划分）	A subset of dictionary terms, each with its complete postings.	Route a single term to its owner; multi-term queries may span nodes. Hot terms/list sizes cause skew（负载倾斜）and difficult rebalancing.
Document-partitioned（按文档划分）	A subset of documents and postings restricted to those docIDs. The lecture models a full term dictionary at every node.	Broadcast to document shards（文档分片）, then merge/rank results; add shards for new documents.

Real systems may combine both. A shard can omit absent terms physically while retaining the logical vocabulary; a “node” here means a machine/service, not a tree node.

Quiz 3.3 Q1 / WK3 T/F13 / WK4 T/F8: each document-partitioned node contains a subset of documents within a specified index range. True: “index” means document ID range（文档编号区间） here.

HW3 T/F1: document partitioning stores all documents containing terms in a certain term range at each node. False: this describes term partitioning's allocation by vocabulary, not by docID.

3.2 Dynamic indexing（动态索引）

New documents require updates to existing postings and insertion of new terms; documents can also change or be deleted. Rewriting large sorted postings on every arrival is expensive.

Maintain a large main index（主索引） and small auxiliary index（辅助／增量索引）. Write incoming changes to the auxiliary index; queries consult both, merge results and filter deleted documents. Periodically merge the auxiliary into the main index in the background and publish a consistent new version.

Offline discussion: when to re-index and how to delete? Balance auxiliary size/query overhead against merge cost and freshness（新鲜度）. A deletion bitmap/tombstone（删除位图／墓碑标记）can hide obsolete docIDs immediately; later merging physically removes their postings. These are implementation options beyond the slide's open questions.

辅助索引不是查询缓存。 It contains new index data; a cache stores prior results or frequently used data. Searching only the recent auxiliary can be a deliberate approximation, but complete retrieval must also cover the main index. Distributed arrival order and updates to existing IDs can require insertion in the middle of a posting list even when newly assigned IDs are monotone.

3.3 Dictionary and posting compression（词典与倒排列表压缩）

Dictionary: fixed-width string slots waste space when term lengths differ. Pack term bytes into one concatenated string and store start offsets plus lengths/boundaries（起始偏移与长度／边界）. Filtering optional stop terms reduces entries but is separate from lossless encoding（无损编码）.

Gap / delta encoding（差分编码）： for sorted docIDs \(d_1<\cdots<d_p\), store

\[ g_1=d_1,\qquad g_i=d_i-d_{i-1}\ (i>1),\qquad d_i=\sum_{j=1}^{i}g_j. \]

Example: 2,15,47,…,58879,58890,… → 2,13,32,…,11,…. Small gaps need fewer bits when combined with variable-length coding（变长编码）, such as variable-byte or gamma codes; storing every gap in the same fixed-width integer does not by itself save bytes. Rare-term gaps can still be large; “fewer than 20 bits” is an example observation, not a universal bound. Absolute restart IDs must remain representable. Postings compression.

3.4 Offline discussion: compression tradeoffs（压缩的代价与方案）

Question: what serious problems does gap encoding introduce, and how can we mitigate them?

Problem	Idea / limitation
Random access（随机访问）needs prefix decoding/accumulation	Store absolute restart points（重启点） every block; an offset directory locates the compressed block. With \(B\) entries/block, decode at most \(O(B)\) entries after choosing the block.
A wrong gap shifts later decoded IDs; corrupted variable-length boundaries can desynchronize decoding	Independent blocks/checksums limit error spread.
Mid-list insertion/deletion plus packed variable-length storage causes movement	Buffer updates in the auxiliary index; rebuild compressed blocks during merging.
Dense skip metadata can erase space savings	Tune block/skip density to query and update patterns.

关键区分：数学差值更新 ≠ 整段编码移动。 Insert \(x\) between \(a<b\): replace gap \(b-a\) by \(x-a\) and \(b-x\); all later adjacent-ID gaps are unchanged if IDs are stable. Deleting \(x\) merges these two gaps. Packed bytes may shift; fixed “every fourth entry” skip/restart positions may need rebuilding. The instructor clarified this distinction in the post-class discussion. Renumbering IDs is a separate source of widespread changes.

Other class proposals:

• Block-base offsets（块基值偏移）： store one absolute base and each entry's offset from it. Direct decoding within a located fixed-width block is easy, but offsets may exceed adjacent gaps; layout/codec determine access cost.
• Multilevel skips / skip lists（多级跳跃／跳表）： express lanes reduce traversal. Storing \(\log p\) jump fields at every posting costs \(O(p\log p)\) fields, plus their bit widths; it is not \(O(\log p)\) total space. Sparse/randomized levels reduce overhead. Exact fixed-position skips and probabilistic skip lists have different update behavior; expected bounds are not per-operation worst-case guarantees.
• Hot-term policy（热词策略）： use smaller blocks, more restart points, or uncompressed/bitmap representations for frequently accessed/updated terms; tune total byte cost as well as operation count.
• “取对数存大 ID” cannot be lossless just by storing a rounded exponent: many IDs share that exponent. Preserve all identifying bits or a reversible encoding; compression must reconstruct exact docIDs.

列表顺序不可混用：gap compression and merge intersection assume docID order; the dictionary is in term order; displayed hits are in score/rank order. Ranking postings by score may make docID differences negative and invalidate this positive-gap codec. Keep ranked access as a separate representation/processing step when needed.

3.5 Thresholding（阈值截断）

Type	Course strategy	Consequence
Document thresholding（文档截断）	Rank documents by weight; retain only top \(k\).	Can lose relevant results. Truncating individual term lists before AND does not preserve exact Boolean results.
Query thresholding（查询截断）	Sort query terms by increasing \(df\); use only an initial fraction.	Saves work by retaining discriminative terms, but changes the original query.

Example: ten query terms → try 20%, 40%, 80%; compare answer sets. Offline incremental example: 10,000 candidates → 5,000 after another term → little change after the next; stop when change is small enough. 结果稳定只是启发式停止条件，不保证完整查询的答案或相关性。 For AND, dropping terms broadens the candidate set; do not confuse this with merely reordering all terms, which preserves the exact answer.

MOOC 3.3.4 Q1: given high- and low-frequency words, which is generally more important? Choices A high / B low / C unrelated / D unknown. Answer: B under the course's discriminativeness heuristic: low document frequency distinguishes fewer documents. WK3 T/F9: very long posting lists are generally less important than short ones — True in this sense; rarity alone does not prove semantic relevance.

Quiz 3.3 Q2 / WK3 T/F3: “Thresholding for query retrieves top \(k\) documents by weight.” False: that is document thresholding; query thresholding selects terms. 中文是“阈值”，不是“阀值”。

4. Evaluation（评价）

4.1 System performance versus relevance（系统性能与相关性）

Measure	Meaning
Indexing throughput（建索引吞吐量）	Documents/hour.
Query latency（查询延迟）	Response time as a function of index size, with comparable query/workload conditions.
Query-language expressiveness（查询语言表达能力）	Complex needs, e.g. apple NOT company, phrases, combinations; also speed on complex queries.
User happiness（用户满意度）	In this lecture, efficient correct retrieval plus relevant results.

Data retrieval（数据检索）: establish correctness of exact matching, then measure response time/index space. Information retrieval, IR（信息检索）: also evaluate relevance（相关性）to the information need. red and apple occurring together can mean a red balloon outside Apple Inc., not a red fruit.

HW3 T/F2: measuring answer-set relevance is important for data retrieval performance — False in the course's terminology; that extra criterion distinguishes IR. Do not read this as “relevance is unimportant to search engines.”

HW3 MC2: which is NOT among the stated search-engine measures? A indexing speed; B search speed; C interface friendliness; D answer relevance. Answer: C for this lecture's list. Interface usability can matter in broader product evaluation; the question's scope is the stated technical measures.

4.2 Precision and recall（查准率与查全率）

Evaluation needs a benchmark collection（基准文档集）, a query suite（查询集）, and binary relevance judgments（相关／不相关标注）for each query–document pair.

	Relevant（相关）	Irrelevant（不相关）
Retrieved（检出）	\(TP=RR\)	\(FP=IR\)
Not retrieved（未检出）	\(FN=RN\)	\(TN=IN\)

\[ P=\frac{TP}{TP+FP}\quad\text{(precision, 查准率／精确率)},\qquad R=\frac{TP}{TP+FN}\quad\text{(recall, 查全率／召回率)}. \]

P 问“返回的有多准”，R 问“该找的找全了吗”。 Precision is not accuracy（准确率，分类中通常指 \((TP+TN)/\text{total}\)）. Empty denominators need an explicit evaluation convention; do not silently divide by zero.

Behavior	Precision	Recall
Return one relevant document while many relevant ones exist	100%	Low
Return the entire collection	Relevant fraction of collection	100% if relevant documents exist
Return exactly the relevant set	100%	100%

The slide's precision–recall curve illustrates a tradeoff, not an identity forcing one metric to decrease whenever the other increases.

WK3 T/F5: “Precision measures the quality of all retrieved documents.” True: it measures the relevant fraction of that returned set.

Quiz 3.4 Q2 / WK3 T/F12: high precision, low recall means most relevant documents are returned but many irrelevant ones are included. False: that describes high recall, low precision.

HW3 MC1 / WK4 MC4: high precision and low recall mean which?

• A. Most relevant documents retrieved, but too many irrelevant ones returned.
• B. Most retrieved documents relevant, but many relevant documents missed.
• C. Most relevant documents retrieved, but benchmark too small.
• D. Most retrieved documents relevant, but benchmark too small.

Answer: B. These metrics do not establish benchmark adequacy.

4.3 Priorities and improving relevance（指标取舍与相关性改进）

Offline discussion: when is precision or recall more important? Define the positive class（正类） and compare the costs of false positives（误报） and false negatives（漏报）.

• Prioritize precision when unwanted returned/flagged items are costly: avoid blocking legitimate mail; show a small, useful set of recommendations rather than many irrelevant ones.
• Prioritize recall when omissions are costly: evidence gathering, exhaustive code/paper search, initial screening of suspicious items. HW3 T/F3: precision is more important than recall for airport explosive detection — False in the question's scenario; detecting dangerous positives is the priority.
• A two-stage pipeline（两阶段流程）can seek broad recall first, then verify/rank candidates to improve precision. WK4 T/F7: recall is more important than precision for cheating detection in an Online Monitor System — True in the stated initial-alert scenario. Broad detection is followed by human verification; an automatic alert is not a final misconduct judgment. The class also contrasts broad initial screening with later confirmation: priority depends on the stage and error costs.

Shopping/paper-search debate: information density or corpus size alone does not choose the metric. A comprehensive literature search may emphasize recall; top recommendations may emphasize precision. Personalization（个性化／用户画像）can improve relevance for a particular user, with evaluation tied to that need.

MOOC 3.4.2 discussion: how to improve relevance? The course names PageRank（网页链接重要性排序） and Semantic Web（语义网）. Add context/meaning and user preferences to distinguish senses such as fruit/company. PageRank is an importance signal, not proof of query relevance; keyword overlap alone does not capture meaning. Evaluate improvements against the same relevance benchmark.

5. Integrated exercises（综合题）

5.1 Spam detection（垃圾邮件检测）· Quiz 3.4 Q1

Dataset: 7,981 legitimate and 2,019 spam emails. A flags 200 legitimate + 1,800 spam; B flags 160 legitimate + 1,500 spam. If keeping important legitimate mail safe is the priority, which choice is correct?

A precision, A better; B recall, B better; C precision, B better; D recall, A better.

Positive = spam, retrieved = flagged as spam:

System	\(TP\)	\(FP\)	\(FN\)	Precision	Recall
A	1800	200	219	\(1800/2000=90\%\)	\(1800/2019\approx89.15\%\)
B	1500	160	519	\(1500/1660\approx90.36\%\)	\(1500/2019\approx74.29\%\)

Answer: C. B has higher spam precision and also flags fewer legitimate emails in this dataset. A catches more spam. 若把普通邮件改成正类，指标名称与分母也必须重新解释，不能只凭“保护邮件”猜 recall。

WK4 MC2 (different dataset): 10,000 legitimate + 2,000 spam emails; A flags 300 legitimate + 1,600 spam, B flags 315 legitimate + 1,800 spam. Choices: A precision/A better; B recall/B better; C precision/B better; D recall/A better. Protecting important legitimate mail again selects C under the course's precision criterion:

System	Spam precision	Spam recall	Legitimate-mail false-positive rate
A	\(1600/1900\approx84.21\%\)	\(1600/2000=80\%\)	\(300/10000=3\%\)
B	\(1800/2115\approx85.11\%\)	\(1800/2000=90\%\)	\(315/10000=3.15\%\)

指标与代价要分开：B has higher precision, yet deletes 15 more legitimate messages on this fixed dataset. If “safe” means minimizing their absolute loss/false-positive rate alone, A is better. The accepted answer C follows the lecture's use of spam precision; it is not a universal guarantee of fewer false deletions. Set the operational objective explicitly rather than substituting one metric for another.

5.2 Recall calculation（查全率计算）· HW3 MC3

There are 28,000 documents:

	Relevant	Irrelevant
Retrieved	4000	12000
Not retrieved	8000	4000

Recall choices: A 14% / B 25% / C 33% / D 50%.

Answer: C, \(R=4000/(4000+8000)=1/3\approx33\%\). Precision is \(4000/(4000+12000)=25\%\); \(4000/28000\approx14\%\) is neither metric.

WK4 T/F2: with 100,000 documents, is precision 20%? True for the supplied table:

	Relevant	Irrelevant
Retrieved	10000	40000
Not retrieved	30000	20000

\(P=10000/(10000+40000)=20\%\); \(R=10000/(10000+30000)=25\%\). The whole-collection size is neither denominator.

5.3 Concept grouping（概念辨析）· WK3 MC4

Which group is not entirely related to the search-engine topics taught here?

• A. Inverted file index, stop words, precision.
• B. Word stemming, hashing, compression.
• C. Distributed index, backtracking, query.
• D. Posting list, thresholding, recall.

Answer: C. Backtracking（回溯搜索）is not one of this lesson's indexing/query modules; all other listed concepts are. This does not claim no search-related implementation could ever use backtracking.

6. Review and references（复习与来源）

复习抓手：正排／倒排方向；\(df\) 与 \(tf\)；位置与短语匹配；短列表优先；stemming 与 stop filtering；分块后最终 flush；词项／文档分片；主索引＋增量索引；差值局部更新与编码移动；docID／词项／排序分数三种顺序；文档／查询截断；P/R 分母、正类和误报／漏报成本。

Historical context（线下背景）： the lecture traces early FTP-directory discovery (Archie), Web crawling (Wanderer), manually curated portals (Yahoo/Sohu), Tianwang, and automated engines (Google/Baidu). The technical progression is collect → organize → serve, from manual organization to scalable automatic indexing/ranking. Inverted indexing predates Web search, including patent keyword retrieval. Slide traffic rankings, index-size estimates and company anecdotes are historical examples, not current measurements or algorithm guarantees.

Sources: all supplied ~/ADS-docs/3 MOOC merged/lesson exports and manifest (units 1320673534–1320673550), 111 video captures, the 21-page courseware PDF, offline lecture-record-17/course_content.md and 24 captures, plus WK3/HW3 archives. This page covers all 11 MOOC questions and all 10 current-topic WK3 questions; duplicated questions share an explanation. WK4 supplements from ~/ADS-docs/4/[Archive] ZJUADS_cy2026_QuizWK4.md and the offline review: all five inverted-index/evaluation questions are integrated here (two new datasets and three shared explanations). HW3's seven nonprogramming questions appear here; its programming task has a separate complete problem and solution.

• ZJU MOOC, supplied term 1488053496.
• Manning, Raghavan & Schütze, Introduction to Information Retrieval: Boolean/positional indexes, construction, compression and evaluation; supplementary clarification, not an assigned page range.
• The supplied slides name InvertedFileIndex.zip: The Google File System, Building an Inverted Index, Inverted Index Construction (ppt), Compression. That archive and its individual files are not present in the supplied directory; the notes cover the provided sources and do not imply those additional readings were inspected.

Four editable SVG figures are original diagrams. Status remains draft pending instructor/textbook review.