This action might not be possible to undo. Are you sure you want to continue?

# Advanced Topics in Information Systems: Information Retrieval

Advanced Topics in Information Systems:

Information Retrieval

Jun.Prof. Alexander Markowetz Slides modified from Christopher Manning and Prabhakar Raghavan

Advanced Topics in Information Systems: Information Retrieval

**DICTIONARY DATA STRUCTURES
**

2

Advanced Topics in Information Systems: Information Retrieval

Sec. 3.1

Hashes

Each vocabulary term is hashed to an integer

(We assume you’ve seen hashtables before)

Pros:

Lookup is faster than for a tree: O(1)

Cons:

No easy way to find minor variants:

judgment/judgement

No prefix search [tolerant retrieval] If vocabulary keeps growing, need to occasionally do the expensive operation of rehashing everything

3

1 Tree: binary tree a-m Root n-z a-hu hy-m n-sh si-z huy g ard sic k 4 zyg ot var k ens le .Advanced Topics in Information Systems: Information Retrieval Sec. 3.

[2.Advanced Topics in Information Systems: Information Retrieval Sec.. 5 .4].1 Tree: B-tree a-hu n-z hy-m Definition: Every internal nodel has a number of children in the interval [a.g. b are appropriate natural numbers. 3.b] where a. e.

1 Trees Simplest: binary tree More usual: B-trees Trees require a standard ordering of characters and hence strings … but we standardly have one Pros: Solves the prefix problem (terms starting with hyp) Cons: Slower: O(log M) [and this requires balanced tree] Rebalancing binary trees is expensive But B-trees mitigate the rebalancing problem 6 .Advanced Topics in Information Systems: Information Retrieval Sec. 3.

Advanced Topics in Information Systems: Information Retrieval WILD-CARD QUERIES 7 .

Exercise: from this. 3. Can retrieve all words in range: nom ≤ w < non. Easy with binary tree (or B-tree) lexicon: retrieve all words in range: mon ≤ w < moo *mon: find words ending in “mon”: harder Maintain an additional B-tree for terms backwards. how can we enumerate all terms meeting the wild-card query pro*cent ? 8 .2 Wild-card queries: * mon*: find all docs containing any word beginning “mon”.Advanced Topics in Information Systems: Information Retrieval Sec.

E.. consider the query: se*ate AND fil*er This may result in the execution of many Boolean AND queries.2 Query processing At this point.g. we have an enumeration of all terms in the dictionary that match the wildcard query. 9 .Advanced Topics in Information Systems: Information Retrieval Sec. 3. We still have to look up the postings for each enumerated term.

10 .2 B-trees handle *’s at the end of a query term How can we handle *’s in the middle of query term? co*tion We could look up co* AND *tion in a B-tree and intersect the two term sets Expensive The solution: transform wild-card queries so that the *’s occur at the end This gives rise to the Permuterm Index.Advanced Topics in Information Systems: Information Retrieval Sec. 3.

index under: hello$. ello$h.1 Permuterm index For term hello. llo$he. Queries: X lookup on X$ X* lookup on $X* *X lookup on X$* *X* lookup on X* X*Y lookup on Y$X* X*Y*Z ??? Exercise! Query = hel*o X=hel. Y=o Lookup o$hel* 11 . 3.2. o$hell where $ is a special symbol.Advanced Topics in Information Systems: Information Retrieval Sec. lo$hel.

12 . Permuterm problem: ≈ quadruples lexicon size Empirical observation for English.2. 3.Advanced Topics in Information Systems: Information Retrieval Sec.1 Permuterm query processing Rotate query wild-card to the right Now use B-tree lookup as before.

t$.nt.2.ri.ap.l$.el.is.mo.e$.$t.il.Advanced Topics in Information Systems: Information Retrieval Sec.s$. ue.le.g.ru. 3.cr. from text “April is the cruelest month” we get the 2-grams (bigrams) $a.th.2 Bigram (k-gram) indexes Enumerate all k-grams (sequence of k chars) occurring in any term e. 13 .he..$i. $m.st.on.h$ $ is a special word boundary symbol Maintain a second inverted index from bigrams to dictionary terms that match each bigram.es.pr.$c.

$m mo on mace among among madden amortize around 14 .2 Bigram index example The k-gram index finds terms based on a query consisting of k-grams (here k=2).2. 3.Advanced Topics in Information Systems: Information Retrieval Sec.

15 . But we’d enumerate moon. 3. Surviving enumerated terms are then looked up in the term-document inverted index. space efficient (compared to permuterm). Must post-filter these terms against query. Fast.Advanced Topics in Information Systems: Information Retrieval Sec.2 Processing wild-cards Query mon* can now be run as $m AND mo AND on Gets terms that match AND version of our wildcard query.2.

2 Processing wild-card queries As before.g..Advanced Topics in Information Systems: Information Retrieval Sec. Wild-cards can result in expensive query execution (very large disjunctions…) pyth* AND prog* If you encourage “laziness” people will respond! Search Type your search terms. Which web search engines allow wildcard queries? 16 . we must execute a Boolean query for each enumerated. 3. filtered term.2. Alex* will match Alexander. E. use ‘*’ if you need to.

Advanced Topics in Information Systems: Information Retrieval SPELLING CORRECTION 17 .

Advanced Topics in Information Systems: Information Retrieval Sec..g. I flew form Heathrow to Narita. from → form Context-sensitive Look at surrounding words.g.3 Spell correction Two principal uses Correcting document(s) being indexed Correcting user queries to retrieve “right” answers Two main flavors: Isolated word Check each word on its own for misspelling Will not catch typos resulting in correctly spelled words e.. e. 18 . 3.

But also: web pages and even printed material has typos Goal: the dictionary contains fewer misspellings But often we don’t change the documents but aim to fix the query-document mapping 19 .Advanced Topics in Information Systems: Information Retrieval Sec. 3.. OCR can confuse O and D more often than it would confuse O and I (adjacent on the QWERTY keyboard. so more likely interchanged in typing).3 Document correction Especially needed for OCR’ed documents Correction algorithms are tuned for this: rn/m Can use domain-specific knowledge E.g.

Advanced Topics in Information Systems: Information Retrieval Sec..3 Query mis-spellings Our principal focus here E. 3. the query Alanis Morisett We can either Retrieve documents indexed by the correct spelling.g. OR Return several suggested alternative queries with the correct spelling Did you mean … ? 20 .

g. 3. acronyms etc. all words on the web All names..3.2 Isolated word correction Fundamental premise – there is a lexicon from which the correct spellings come Two basic choices for this A standard lexicon such as Webster’s English Dictionary An “industry-specific” lexicon – hand-maintained The lexicon of the indexed corpus E. (Including the mis-spellings) 21 .Advanced Topics in Information Systems: Information Retrieval Sec.

Advanced Topics in Information Systems: Information Retrieval Sec. return the words in the lexicon closest to Q What’s “closest”? We’ll study several alternatives Edit distance (Levenshtein distance) Weighted edit distance n-gram overlap 22 .2 Isolated word correction Given a lexicon and a character sequence Q.3. 3.

.3 Edit distance Given two strings S1 and S2. 23 .) Generally found by dynamic programming. 3.htm for a nice example plus an applet. the minimum number of operations to convert one to the other Operations are typically character-level Insert. (Transposition) E. (Just 1 with transpose.3. Replace. Delete.com/ld. the edit distance from dof to dog is 1 From cat to act is 2 from cat to dog is 3.Advanced Topics in Information Systems: Information Retrieval Sec. See http://www.merriampark.g.

Advanced Topics in Information Systems: Information Retrieval

Sec. 3.3.3

**Weighted edit distance
**

As above, but the weight of an operation depends on the character(s) involved

Meant to capture OCR or keyboard errors, e.g. m more likely to be mis-typed as n than as q Therefore, replacing m by n is a smaller edit distance than by q This may be formulated as a probability model

Requires weight matrix as input Modify dynamic programming to handle weights

24

Advanced Topics in Information Systems: Information Retrieval

Sec. 3.3.4

**Using edit distances
**

Given query, first enumerate all character sequences within a preset (weighted) edit distance (e.g., 2) Intersect this set with list of “correct” words Show terms you found to user as suggestions Alternatively,

We can look up all possible corrections in our inverted index and return all docs … slow We can run with a single most likely correction

** The alternatives disempower the user, but save a round of interaction with the user
**

25

Advanced Topics in Information Systems: Information Retrieval

Sec. 3.3.4

**Edit distance to all dictionary terms?
**

Given a (mis-spelled) query – do we compute its edit distance to every dictionary term?

Expensive and slow Alternative?

How do we cut the set of candidate dictionary terms? One possibility is to use n-gram overlap for this This can also be used by itself for spelling correction.

26

27 .4 n-gram overlap Enumerate all the n-grams in the query string as well as in the lexicon Use the n-gram index (recall wild-card search) to retrieve all lexicon terms matching any of the query n-grams Threshold by number of matching n-grams Variants – weight by keyboard layout. etc.3.Advanced Topics in Information Systems: Information Retrieval Sec. 3.

So 3 trigrams overlap (of 6 in each term) How can we turn this into a normalized measure of overlap? 28 . emb.Advanced Topics in Information Systems: Information Retrieval Sec. emb. ber. ove. cem. ber. mbe. 3. ece. The query is december Trigrams are dec.4 Example with trigrams Suppose the text is november Trigrams are nov. mbe.3. vem.

if J.3.8.C.C.4 One option – Jaccard coefficient A commonly-used measure of overlap Let X and Y be two sets. 3.Advanced Topics in Information Systems: Information Retrieval Sec.. > 0. is X ∩Y / X ∪Y Equals 1 when X and Y have the same elements and zero when they are disjoint X and Y don’t have to be of the same size Always assigns a number between 0 and 1 Now threshold to decide if you have a match E. declare a match 29 . then the J.g.

3. or. 3.4 Matching trigrams Consider the query lord – we wish to identify words matching 2 of its 3 bigrams (lo. rd) lo or rd alone border ardent lord lord border sloth morbid card Standard postings “merge” will enumerate … Adapt this to using Jaccard (or another) measure. 30 .Advanced Topics in Information Systems: Information Retrieval Sec.

Advanced Topics in Information Systems: Information Retrieval Sec. 31 . 3. Consider the phrase query “flew form Heathrow” We’d like to respond Did you mean “flew from Heathrow”? because no docs matched the query phrase.5 Context-sensitive spell correction Text: I flew from Heathrow to Narita.3.

32 .Advanced Topics in Information Systems: Information Retrieval Sec. 3. First idea: retrieve dictionary terms close (in weighted edit distance) to each query term Now try all possible resulting phrases with one word “fixed” at a time flew from heathrow fled form heathrow flea form heathrow Hit-based spelling correction: Suggest the alternative that has lots of hits.5 Context-sensitive correction Need surrounding context to catch this.3.

3.Advanced Topics in Information Systems: Information Retrieval Sec. 3. 19 for form and 3 for heathrow.5 Exercise Suppose that for “flew form Heathrow” we have 7 alternatives for flew. How many “corrected” phrases will we enumerate in this scheme? 33 .

5 Another approach Break phrase query into a conjunction of biwords (Lecture 2).3.Advanced Topics in Information Systems: Information Retrieval Sec. Look for biwords that need only one term corrected. Enumerate phrase matches and … rank them! 34 . 3.

3.3. topical queries Spell-correction is computationally expensive Avoid running routinely on every query? Run only on queries that matched few docs 35 .5 General issues in spell correction We enumerate multiple alternatives for “Did you mean?” Need to figure out which to present to the user Use heuristics The alternative hitting most docs Query log analysis + tweaking For especially popular.Advanced Topics in Information Systems: Information Retrieval Sec.

Advanced Topics in Information Systems: Information Retrieval SOUNDEX 36 .

chebyshev → tchebycheff Invented for the U. census … in 1918 37 .g.4 Soundex Class of heuristics to expand a query into phonetic equivalents Language specific – mainly for names E..Advanced Topics in Information Systems: Information Retrieval Sec. 3.S.

3.Advanced Topics in Information Systems: Information Retrieval Sec.com/Doc/Articles/SoundEx1/SoundEx1.htm#Top 38 .creativyst.4 Soundex – typical algorithm Turn every token to be indexed into a 4character reduced form Do the same with query terms Build and search an index on the reduced forms (when the query calls for a soundex match) http://www.

Retain the first letter of the word.T → 3 L→4 M. 'U'. E'. 3. 3. Change all occurrences of the following letters to '0' (zero): 'A'. 2.4 Soundex – typical algorithm 1.Advanced Topics in Information Systems: Information Retrieval Sec. G. 'I'. Z → 2 D. Change letters to digits as follows: B. P. 'W'. K. J. S. 'O'. F. 'H'. 'Y'. Q. V → 1 C. N → 5 R→6 39 . X.

Remove all zeros from the resulting string. 6. 3. Herman becomes H655. Remove all pairs of consecutive digits. 5.Advanced Topics in Information Systems: Information Retrieval Sec.. which will be of the form <uppercase letter> <digit> <digit> <digit>.4 Soundex continued 4. Will hermann generate the same code? 40 . E.g. Pad the resulting string with trailing zeros and return the first four positions.

3..Advanced Topics in Information Systems: Information Retrieval Sec. Microsoft. …) How useful is soundex? Not very – for information retrieval Okay for “high recall” tasks (e.4 Soundex Soundex is the classic algorithm. though biased to names of certain nationalities Zobel and Dart (1996) show that other algorithms for phonetic matching perform much better in the context of IR 41 . Interpol).g. provided by most databases (Oracle.

Advanced Topics in Information Systems: Information Retrieval What queries can we process? We have Positional inverted index with skip pointers Wild-card index Spell-correction Soundex Queries such as (SPELL(moriset) /3 toron*to) OR SOUNDEX(chaikofski) 42 .

Advanced Topics in Information Systems: Information Retrieval INDEX GENERATION 43 .

4 Index construction How do we construct an index? What strategies can we use with limited main memory? 44 .Advanced Topics in Information Systems: Information Retrieval Ch.

1 Hardware basics Many design decisions in information retrieval are based on the characteristics of hardware We begin by reviewing hardware basics 45 .Advanced Topics in Information Systems: Information Retrieval Sec. 4.

Disk I/O is block-based: Reading and writing of entire blocks (as opposed to smaller chunks). Block sizes: 8KB to 256 KB.1 Hardware basics Access to data in memory is much faster than access to data on disk. Disk seeks: No data is transferred from disk while the disk head is being positioned. 4. Therefore: Transferring one large chunk of data from disk to memory is faster than transferring many small chunks. 46 .Advanced Topics in Information Systems: Information Retrieval Sec.

Available disk space is several (2–3) orders of magnitude larger.1 Hardware basics Servers used in IR systems now typically have several GB of main memory. 47 .Advanced Topics in Information Systems: Information Retrieval Sec. sometimes tens of GB. 4. Google is particularly famous for combining standard hardware … in shipping containers. Fault tolerance is very expensive: It’s much cheaper to use many regular machines rather than one fault tolerant machine.

g.01 μs = 10−8 s (e.02 μs = 2 x 10−8 s processor’s clock rate 109 s−1 p low-level operation 0. compare & swap a word) size of main memory size of disk space several GB 1 TB or more 48 .Advanced Topics in Information Systems: Information Retrieval Sec.. 4.1 Hardware assumptions symbol statistic value s average seek time 5 ms = 5 x 10−3 s b transfer time per byte 0.

This is one year of Reuters newswire (part of 1995 and 1996) 49 . but it’s publicly available and is at least a more plausible example.Advanced Topics in Information Systems: Information Retrieval Sec. As an example for applying scalable index construction algorithms. 4. we will use the Reuters RCV1 collection. The collection we’ll use isn’t really large enough either.2 RCV1: Our collection for this lecture Shakespeare’s collected works definitely aren’t large enough for demonstrating many of the points in this course.

2 A Reuters RCV1 document 50 . 4.Introduction to Information Retrieval Sec.

000 6 4. # bytes per term 7.000 51 .5 avg.) avg.) value 800.2 Reuters RCV1 statistics symbol statistic N documents L avg.Advanced Topics in Information Systems: Information Retrieval Sec. spaces/punct. # bytes per token (without spaces/punct. 4. # bytes per token (incl.000. # tokens per doc M terms (= word types) avg.5 non-positional postings 100.000 200 400.

Doc 2 So let it be with Caesar.Advanced Topics in Information Systems: Information Retrieval Sec. Doc 1 I did enact Julius Caesar I was killed i' the Capitol. The noble Brutus hath told you Caesar was ambitious Term I did enact julius caesar I was killed i' the capitol brutus killed m e so let it be with caesar the noble brutus hath told you caesar was am bitious Doc # 1 1 1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 2 2 2 2 52 2 .2 Recall IIR 1 index construction Documents are parsed to extract words and these are saved with the Document ID. 4. Brutus killed me.

4.Advanced Topics in Information Systems: Information Retrieval Sec. We have 100M items to sort. We focus on this sort step.2 Key step After all documents have been parsed. the inverted file is sorted by terms. Term I did enact julius caesar I was killed i' the capitol brutus killed m e so let it be with caesar the noble brutus hath told you caesar was am bitious Doc # 1 1 1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 Term ambitious be brutus brutus capitol caesar caesar caesar did enact hath I I i' it julius killed killed let me noble so the the told you was was with Doc # 2 2 1 2 1 1 2 2 1 1 1 1 1 1 2 1 1 1 2 1 2 2 1 2 2 2 1 2 2 53 .

54 . Memory.Advanced Topics in Information Systems: Information Retrieval Sec. . How can we construct an index for very large collections? Taking into account the hardware constraints we just learned about .2 Scaling index construction In-memory index construction does not scale. disk. 4. speed. . etc.

the New York Times provides an index of >150 years of newswire Thus: We need to store intermediate results on disk. but typical collections are much larger. doc. demands a lot of space for large collections. but much more complex) The final postings for any term are incomplete until the end. we cannot easily exploit compression tricks (you can. While building the index. T = 100.g.000 in the case of RCV1 So … we can do this in memory in 2009. 4. 55 . we parse docs one at a time. E.2 Sort-based index construction As we build the index.Advanced Topics in Information Systems: Information Retrieval Sec.000. freq). At 12 bytes per non-positional postings entry (term.

000 records on disk is too slow – too many disk seeks. 56 .Advanced Topics in Information Systems: Information Retrieval Sec. but by using disk instead of memory? No: Sorting T = 100. 4.2 Use the same algorithm for disk? Can we use the same index construction algorithm for larger collections. We need an external sorting algorithm.000.

4.2 Bottleneck Parse and build postings entries one doc at a time Now sort postings entries by term (then by doc within each term) Doing this with random disk seeks would be too slow – must sort T=100M records If every comparison took 2 disk seeks.Advanced Topics in Information Systems: Information Retrieval Sec. and N items could be sorted with N log2N comparisons. how long would this take? 57 .

write to disk. Must now sort 100M such 12-byte records by term. Basic idea of algorithm: Accumulate postings for each block. 58 . 4. These are generated as we parse docs. Will have 10 such blocks to start with. sort. freq). Define a Block ~ 10M such records Can easily fit a couple into memory. doc.2 BSBI: Blocked sort-based Indexing (Sorting with fewer disk seeks) 12-byte (4+4+4) records (term.Advanced Topics in Information Systems: Information Retrieval Sec. Then merge the blocks into one long sorted order.

Advanced Topics in Information Systems: Information Retrieval Sec.2 59 . 4.

need 2 copies of data on disk But can optimize this 60 . read each block and sort within: Quicksort takes 2N ln N expected steps In our case 2 x (10M ln 10M) steps Exercise: estimate total time to read each block from disk and and quicksort it.2 Sorting 10 blocks of 10M records First. 10 times this estimate – gives us 10 sorted runs of 10M records each. Done straightforwardly.Advanced Topics in Information Systems: Information Retrieval Sec. 4.

Sec. 4.2 61 .

Advanced Topics in Information Systems: Information Retrieval Sec. merge. read into memory runs in blocks of 10M. Disk 62 . 4 3 4 Runs being merged. 4. During each layer. with a merge tree of log210 = 4 layers. 1 1 3 2 2 Merged run. write back.2 How to merge the sorted runs? Can do binary merges.

4.Advanced Topics in Information Systems: Information Retrieval Sec.2 How to merge the sorted runs? But it is more efficient to do a n-way merge. where you are reading from all blocks simultaneously Providing you read decent-sized chunks of each block into memory and then write out a decent-sized output chunk. then you’re not killed by disk seeks 63 .

but then intermediate files become very large. . 4.docID postings . .Advanced Topics in Information Systems: Information Retrieval Sec. (We would end up with a scalable. we could work with term.) 64 . . but very slow index construction method.docID postings instead of termID.3 Remaining problem with sortbased algorithm Our assumption was: we can keep the dictionary in memory. . We need the dictionary (which grows dynamically) in order to implement a term to termID mapping. . Actually.

These separate indexes can then be merged into one big index. 65 . Key idea 2: Don’t sort.Advanced Topics in Information Systems: Information Retrieval Sec. Accumulate postings in postings lists as they occur.3 SPIMI: Single-pass in-memory indexing Key idea 1: Generate separate dictionaries for each block – no need to maintain term-termID mapping across blocks. With these two ideas we can generate a complete inverted index for each block. 4.

66 .Advanced Topics in Information Systems: Information Retrieval Sec. 4.3 SPIMI-Invert Merging of blocks is analogous to BSBI.

3 SPIMI: Compression Compression makes SPIMI even more efficient. Compression of terms Compression of postings 67 .Advanced Topics in Information Systems: Information Retrieval Sec. 4.

4 Distributed indexing For web-scale indexing (don’t try this at home!): must use a distributed computing cluster Individual machines are fault-prone Can unpredictably slow down or fail How do we exploit such a pool of machines? 68 . 4.Advanced Topics in Information Systems: Information Retrieval Sec.

Based on expenditures of 200–250 million dollars per year This would be 10% of the computing capacity of the world!?! 69 . 4.Advanced Topics in Information Systems: Information Retrieval Sec. 3 million processors/cores (Gartner 2007) Estimate: Google installs 100.000 servers each quarter.4 Google data centers Google data centers mainly contain commodity machines. Data centers are distributed around the world. Estimate: a total of 1 million servers.

70 .9% uptime.4 Google data centers If in a non-fault-tolerant system with 1000 nodes. each node has 99. what is the uptime of the system? Answer: 63% Calculate the number of servers failing per minute for an installation of 1 million servers.Advanced Topics in Information Systems: Information Retrieval Sec. 4.

Advanced Topics in Information Systems: Information Retrieval Sec. Break up indexing into sets of (parallel) tasks. 71 . 4. Master machine assigns each task to an idle machine from a pool.4 Distributed indexing Maintain a master machine directing the indexing job – considered “safe”.

4.4 Parallel tasks We will use two sets of parallel tasks Parsers Inverters Break the input document collection into splits Each split is a subset of documents (corresponding to blocks in BSBI/SPIMI) 72 .Advanced Topics in Information Systems: Information Retrieval Sec.

Now to complete the index inversion 73 .. q-z) – here j = 3. doc) pairs Parser writes pairs into j partitions Each partition is for a range of terms’ first letters (e.4 Parsers Master assigns a split to an idle parser machine Parser reads a document at a time and emits (term. g-p.Advanced Topics in Information Systems: Information Retrieval Sec.g. 4. a-f.

doc) pairs (= postings) for one term-partition.4 Inverters An inverter collects all (term. Sorts and writes to postings lists 74 . 4.Advanced Topics in Information Systems: Information Retrieval Sec.

4 Data flow assign Parser Parser splits Master assign Inverter Inverter Inverter Postings a-f g-p q-z a-f g-p q-z a-f g-p q-z Parser Map phase a-f g-p q-z Segment files Reduce phase 75 .Introduction to Information Retrieval Sec. 4.

each implemented in MapReduce. 76 .Advanced Topics in Information Systems: Information Retrieval Sec. They describe the Google indexing system (ca. MapReduce (Dean and Ghemawat 2004) is a robust and conceptually simple framework for distributed computing … … without having to write code for the distribution part.4 MapReduce The index construction algorithm we just described is an instance of MapReduce. 4. 2002) as consisting of a number of phases.

77 .Advanced Topics in Information Systems: Information Retrieval Sec. etc. most search engines use a document-partitioned index … better load balancing. Another phase: transforming a term-partitioned index into a document-partitioned index. Term-partitioned: one machine handles a subrange of terms Document-partitioned: one machine handles a subrange of documents As we discuss later in the course.4 MapReduce Index construction was just one phase. 4.

d2>.(d1)>.d1>.d2:1)>. <came. <C. <C.(d1)>) → (<C.Advanced Topics in Information Systems: Information Retrieval Sec. C c’ed.d1>. d1> reduce: (<C.d1)>. list(docID)>. …) → (postings list1. list(docID)>. <came. <c’ed. <died. <died. …) Example for index construction map: d2 : C died. <died.(d2:1)>. v) reduce: (k. <c’ed.d1. postings list2. <c’ed.(d1:1)>) 78 .list(v)) → output Instantiation of the schema for index construction map: web collection → list(termID. docID) reduce: (<termID1. d1 : C came. <termID2.(d2)>.4 Schema for index construction in MapReduce Schema of map and reduce functions map: input → list(k.(d1:1)>. <came. → (<C.d1>. 4.(d2. d2>.(d1:2.

5 Dynamic indexing Up to now. we have assumed that collections are static. This means that the dictionary and postings lists have to be modified: Postings updates for terms already in dictionary New terms added to dictionary 79 . 4. Documents are deleted and modified. They rarely are: Documents come in over time and need to be inserted.Advanced Topics in Information Systems: Information Retrieval Sec.

5 Simplest approach Maintain “big” main index New docs go into “small” auxiliary index Search across both. re-index into one main index 80 .Advanced Topics in Information Systems: Information Retrieval Sec. merge results Deletions Invalidation bit-vector for deleted docs Filter docs output on a search result by this invalidation bit-vector Periodically. 4.

Merge is the same as a simple append. But then we would need a lot of files – inefficient for O/S. Assumption for the rest of the lecture: The index is one big file.g.5 Issues with main and auxiliary indexes Problem of frequent merges – you touch stuff a lot Poor performance during merge Actually: Merging of the auxiliary index into the main index is efficient if we keep a separate file for each postings list.Advanced Topics in Information Systems: Information Retrieval Sec.. In reality: Use a scheme somewhere in between (e. split very large postings lists. collect postings lists of length 1 in one file etc. 4.) 81 .

I1. 4.Advanced Topics in Information Systems: Information Retrieval Sec.5 Logarithmic merge Maintain a series of indexes. each twice as large as the previous one. Keep smallest (Z0) in memory Larger ones (I0. write to disk as I0 or merge with I0 (if I0 already exists) as Z1 Either write merge Z1 to disk as I1 (if no I1) Or merge with I1 to form Z2 etc. 82 . …) on disk If Z0 gets too big (> n).

5 83 . 4.Sec.

so complexity is O(T log T) So logarithmic merge is much more efficient for index construction But query processing now requires the merging of O(log T) indexes Whereas it is O(1) if you just have a main and auxiliary index 84 .Advanced Topics in Information Systems: Information Retrieval Sec. 4.5 Logarithmic merge Auxiliary and main index: index construction time is O(T2) as each posting is touched in each merge. Logarithmic merge: Each posting is merged O(log T) times.

g. 4. when we spoke of spell-correction: which of several corrected alternatives do we present to the user? We said. pick the one with the most hits How do we maintain the top ones with multiple indexes and invalidation bit vectors? One possibility: ignore everything but the main index for such ordering Will see more such statistics used in results ranking 85 .Advanced Topics in Information Systems: Information Retrieval Sec.5 Further issues with multiple indexes Collection-wide statistics are hard to maintain E..

5 Dynamic indexing at search engines All the large search engines now do dynamic indexing Their indices have frequent incremental changes News items.Advanced Topics in Information Systems: Information Retrieval Sec. blogs. 4. and the old index is then deleted 86 . new topical web pages Sarah Palin. … But (sometimes/typically) they also periodically reconstruct the index from scratch Query processing is then switched to the new index.

5 87 . 4.Advanced Topics in Information Systems: Information Retrieval Sec.

Advanced Topics in Information Systems: Information Retrieval

Sec. 4.5

**Other sorts of indexes
**

Positional indexes

Same sort of sorting problem … just larger

** Building character n-gram indexes:
**

As text is parsed, enumerate n-grams. For each n-gram, need pointers to all dictionary terms containing it – the “postings”. Note that the same “postings entry” will arise repeatedly in parsing the docs – need efficient hashing to keep track of this.

Why ?

E.g., that the trigram uou occurs in the term deciduous will be discovered on each text occurrence of deciduous Only need to process each term once

88

Advanced Topics in Information Systems: Information Retrieval

INDEX COMPRESSION

89

Advanced Topics in Information Systems: Information Retrieval

Ch. 5

Compressing Indexes

** Collection statistics in more detail (with RCV1)
**

How big will the dictionary and postings be?

** Dictionary compression Postings compression
**

90

5 Why compression (in general)? Use less disk space Saves a little money Keep more stuff in memory Increases speed Increase speed of data transfer from disk to memory [read compressed data | decompress] is faster than [read uncompressed data] Premise: Decompression algorithms are fast True of the decompression algorithms we use 91 .Advanced Topics in Information Systems: Information Retrieval Ch.

5 Why compression for inverted indexes? Dictionary Make it small enough to keep in main memory Make it so small that you can keep some postings lists in main memory too Postings file(s) Reduce disk space needed Decrease time needed to read postings lists from disk Large search engines keep a significant part of the postings in memory. Compression lets you keep more in memory We will devise various IR-specific compression schemes 92 .Advanced Topics in Information Systems: Information Retrieval Ch.

5 (without spaces/punct. # bytes per term 7.000 avg.000 93 .Advanced Topics in Information Systems: Information Retrieval Sec. 5. spaces/punct. # bytes per token 6 (incl.000 L avg.) avg.5 non-positional postings 100. # tokens per doc 200 M terms (= word types) ~400. # bytes per token 4.1 Recall Reuters RCV1 symbol statistic value N documents 800.000.) avg.

002 -8 -3 -14 -30 -8 -12 -24 -39 ∆ % cumul % position al postings positional index Size (K) 197.812 -4 entries correspond to big deltas in other columns? 94 .158 121.1.517 -9 0 -31 -47 -9 -9 -38 -52 ∆ % cumul % size of Unfiltered No numbers Case folding 30 stopwords 150 stopwords 484 474 392 391 391 stemming 322 -17 -42 94.390 67.969 83.879 179.158 179.971 -2 -17 -0 -0 100.Advanced Topics in Information Systems: Information Retrieval Sec.517 0 Exercise: give intuitions for-33 the ‘0’ entries. p.680 96.1 Index parameters vs.80) word types (terms) dictiona ry Size (K) ∆% cumul % -2 -19 -19 -19 nonpositiona l postings nonpositional index Size (K) 109. Why do some zero-52 all 63.858 94. what we index (details IIR Table 5. 5.

5.Advanced Topics in Information Systems: Information Retrieval Sec. Lossy compression: Discard some information Several of the preprocessing steps can be viewed as lossy compression: case folding. 95 . lossy compression Lossless compression: All information is preserved. Chap/Lecture 7: Prune postings entries that are unlikely to turn up in the top k list for any query. stop words. number elimination. Almost no loss quality for top k list. What we mostly do in IR.1 Lossless vs. stemming.

Advanced Topics in Information Systems: Information Retrieval Sec. 5. how many distinct words are there? Can we assume an upper bound? Not really: At least 7020 = 1037 different words of length 20 In practice. collection size How big is the term vocabulary? That is. the vocabulary will keep growing with the collection size Especially with Unicode 96 .1 Vocabulary vs.

T.1 Vocabulary vs.Advanced Topics in Information Systems: Information Retrieval Sec. Heaps’ law predicts a line with slope about ½ It is the simplest possible relationship between the two in log-log space An empirical finding (“empirical law”) 97 . T is the number of tokens in the collection Typical values: 30 ≤ k ≤ 100 and b ≈ 0. 5. collection size Heaps’ law: M = kTb M is the size of the vocabulary.5 In a log-log plot of vocabulary size M vs.

M = 101. 38.64 ≈ 44 and b = 0. 5.Sec.323 terms.49 so k = 101.64 is the best least squares fit. the dashed line log10 M = 0. actually. Good empirical fit for Reuters RCV1 ! For first 1.64 T0. law predicts 38.49.365 terms Fig 5.1 Heaps’ Law For RCV1.020 tokens.000.1 p81 98 . Thus.49 log10 T + 1.

000.000 tokens and 30.000.000 (2 × 1010 ) pages.Advanced Topics in Information Systems: Information Retrieval Sec.1 Exercises What is the effect of including spelling errors. vs.000 tokens. you find that there are 3000 different terms in the first 10. 5. Assume a search engine indexes a total of 20.000 different terms in the first 1. containing 200 tokens on average What is the size of the vocabulary of the indexed collection as predicted by Heaps’ law? 99 . automatically correcting spelling errors on Heaps’ law? Compute the vocabulary size M for this scenario: Looking at a collection of web pages.000.

cfi ∝ 1/i = K/i where K is a normalizing constant cfi is collection frequency: the number of occurrences of the term ti in the collection. 5. We also study the relative frequencies of terms.Advanced Topics in Information Systems: Information Retrieval Sec. In natural language. Zipf’s law: The ith most frequent term has frequency proportional to 1/i . there are a few very frequent terms and very many very rare terms. 100 .1 Zipf’s law Heaps’ law gives the vocabulary size in collections.

log i Linear relationship between log cfi and log i Another power law relationship 101 . 5. so log cfi = log K .1 Zipf consequences If the most frequent term (the) occurs cf1 times then the second most frequent term (of) occurs cf1/2 times the third most frequent term (and) occurs cf1/3 times … Equivalent: cfi = K/i where K is a normalizing factor.Advanced Topics in Information Systems: Information Retrieval Sec.

Introduction to Information Retrieval Sec.1 Zipf’s law for Reuters RCV1 102 . 5.

We will consider compression schemes 103 . etc. we will consider compressing the space for the dictionary and postings Basic Boolean index only No study of positional indexes.Advanced Topics in Information Systems: Information Retrieval Ch. 5 Compression Now.

Sec.2 DICTIONARY COMPRESSION 104 . 5.

compressing the dictionary is important 105 .2 Why compress the dictionary? Search begins with the dictionary We want to keep it in memory Memory footprint competition with other applications Embedded/mobile devices may have very little memory Even if the dictionary isn’t in memory. 5. we want it to be small for a fast search startup time So.Advanced Topics in Information Systems: Information Retrieval Sec.

000 terms. 5. Terms a aachen …. 28 bytes/term = 11.265 65 ….2 MB.Advanced Topics in Information Systems: Information Retrieval Sec. 656.first cut Array of fixed-width entries ~400.2 Dictionary storage . Dictionary search structure 20 bytes 4 bytes each 106 . 221 Postings ptr. zulu Freq.

5. And we still can’t handle supercalifragilisticexpialidocious or hydrochlorofluorocarbons.Advanced Topics in Information Systems: Information Retrieval Sec. Written English averages ~4.5 characters/word. Exercise: Why is/isn’t this the number to use for estimating the dictionary size? Ave. dictionary word in English: ~8 characters How do we use ~8 characters per dictionary term? Short words dominate token counts but not type average.2 Fixed-width terms are wasteful Most of the bytes in the Term column are wasted – we allot 20 bytes for 1 letter terms. 107 .

Freq. 33 29 44 126 Postings ptr.2MB Pointers resolve 3.2 Compressing the term list: Dictionary-as-a-String Store dictionary as a (long) string of characters: Pointer to next word shows end of current word Hope to save up to 60% of dictionary space.2M = 22bits = 3bytes 108 .systilesyzygeticsyzygialsyzygyszaibelyiteszczecinszomo…. Total string length = 400K x 8B = 3. 5.Introduction to Information Retrieval Sec.2M positions: log23. …. Term ptr.

not 20.Advanced Topics in Information Systems: Information Retrieval Sec. 5.6 MB (against 11.bytes/term. Now avg. 8 bytes per term in term string 400K terms x 19 ⇒ 7. 3 bytes per term pointer Avg.2MB for fixed width) 109 .2 Space for dictionary as a string 4 bytes per term for Freq. 11 4 bytes per term for pointer to Postings.

Advanced Topics in Information Systems: Information Retrieval Sec.7systile9syzygetic8syzygial6syzygy11szaibelyite8szczecin9szomo…. 33 29 44 126 7 Postings ptr. Lose 4 bytes on term lengths. Example below: k=4. 110 .2 Blocking Store pointers to every kth term string. Freq. Need to store term lengths (1 extra byte) …. 5. Term ptr. Save 9 bytes on 3 pointers.

This reduces the size of the dictionary from 7. Shaved another ~0. Why not go with larger k? 111 .2 Net Example for block size k = 4 Where we used 3 bytes/pointer without blocking 3 x 4 = 12 bytes.6 MB to 7.1 MB. now we use 3 + 4 = 7 bytes.Advanced Topics in Information Systems: Information Retrieval Sec. We can save more with larger k.5MB. 5.

Advanced Topics in Information Systems: Information Retrieval Sec. 8 and 16. 5.6 MB) with blocking. 112 .2 Exercise Estimate the space usage (and savings compared to 7. for block sizes of k = 4.

average number of comparisons = (1+2∙2+4∙3+4)/8 ~2. 5. how would you structure the dictionary search tree? 113 .Advanced Topics in Information Systems: Information Retrieval Sec.2 Dictionary search without blocking Assuming each dictionary term equally likely in query (not really so in practice!).6 Exercise: what if the frequencies of query terms were non-uniform but known.

avg. Blocks of 4 (binary tree). = (1+2∙2+2∙3+2∙4+5)/8 = 3 compares 114 .2 Dictionary search with blocking Binary search down to 4-term block. 5. Then linear search through terms in block.Introduction to Information Retrieval Sec.

115 . 8 and 16.Advanced Topics in Information Systems: Information Retrieval Sec. for block sizes of k = 4. 5.2 Exercise Estimate the impact on search performance (and slowdown compared to k=1) with blocking.

2 Front coding Front-coding: Sorted words commonly have long common prefix – store differences only (for last k-1 in a block of k) 8automata8automate9automatic10autom ation →8automat*a1◊e2◊ic3◊ion Encodes automat Extra length beyond automat. 5.Advanced Topics in Information Systems: Information Retrieval Sec. Begins to resemble general string compression.116 .

Blocking + front coding Size in MB 11.2 RCV1 dictionary compression summary Technique Fixed width Dictionary-as-String with pointers to every term Also.Advanced Topics in Information Systems: Information Retrieval Sec.1 5.6 7.2 7. 5.9 117 . blocking k = 4 Also.

3 POSTINGS COMPRESSION 118 . 5.Sec.

000 ≈ 20 bits per docID. 5.Advanced Topics in Information Systems: Information Retrieval Sec. we can use log2 800. 119 . factor of at least 10. Our goal: use a lot less than 20 bits per docID. A posting for our purposes is a docID.3 Postings compression The postings file is much larger than the dictionary. For Reuters (800. we would use 32 bits per docID when using 4-byte integers. Key desideratum: store each posting compactly. Alternatively.000 documents).

5. A term like the occurs in virtually every doc.Advanced Topics in Information Systems: Information Retrieval Sec. Prefer 0/1 bitmap vector in this case 120 . so 20 bits/posting is too expensive.3 Postings: two conflicting forces A term like arachnocentric occurs in maybe one doc out of a million – we would like to store this posting using log2 1M ~ 20 bits.

computer: 33.159.43 … Hope: most gaps can be encoded/stored with far fewer than 20 bits.Advanced Topics in Information Systems: Information Retrieval Sec.107. 5. 33.14.202 … Consequence: it suffices to store gaps.5.3 Postings file entry We store the list of docs containing a term in increasing order of docID. 121 .154.47.

Advanced Topics in Information Systems: Information Retrieval

Sec. 5.3

Three postings entries

122

Advanced Topics in Information Systems: Information Retrieval

Sec. 5.3

**Variable length encoding
**

Aim:

For arachnocentric, we will use ~20 bits/gap entry. For the, we will use ~1 bit/gap entry.

If the average gap for a term is G, we want to use ~log2G bits/gap entry. Key challenge: encode every integer (gap) with about as few bits as needed for that integer. This requires a variable length encoding Variable length codes achieve this by using short codes for small numbers

123

Advanced Topics in Information Systems: Information Retrieval

Sec. 5.3

**Variable Byte (VB) codes
**

For a gap value G, we want to use close to the fewest bytes needed to hold log2 G bits Begin with one byte to store G and dedicate 1 bit in it to be a continuation bit c If G ≤127, binary-encode it in the 7 available bits and set c =1 Else encode G’s lower-order 7 bits and then use additional bytes to encode the higher order bits using the same algorithm At the end set the continuation bit of the last byte to 1 (c =1) – and for the other bytes c = 0.

124

VB uses a whole byte.Advanced Topics in Information Systems: Information Retrieval Sec. 125 . 5. For a small gap (5).3 Example docIDs gaps VB code 00000110 10111000 824 829 5 10000101 215406 214577 00001101 00001100 10110001 Postings stored as the byte concatenation 000001101011100010000101000011010000110010110001 Key property: VB-encoded postings are uniquely prefix-decodable.

bit-level codes.Advanced Topics in Information Systems: Information Retrieval Sec. 16 bits. Variable byte alignment wastes space if you have many small gaps – nibbles do better in such cases. which we look at next). Variable byte codes: Used by many commercial/research systems Good low-tech blend of variable-length coding and sensitivity to computer memory alignment matches (vs. There is also recent work on word-aligned codes that pack a variable number of gaps into one word 126 . we can also use a different “unit of alignment”: 32 bits (words).3 Other variable unit codes Instead of bytes. 4 bits (nibbles). 5.

but…. Unary code for 80 is: 111111111111111111111111111111111111111111 111111111111111111111111111111111111110 This doesn’t look promising. Unary code for 40 is 11111111111111111111111111111111111111110 . Unary code for 3 is 1110. 127 .Advanced Topics in Information Systems: Information Retrieval Unary code Represent n as n 1s with a final 0.

with the leading bit cut off For example 13 → 1101 → 101 length is the length of offset For 13 (offset 101). 5. We encode length with unary code: 1110. Gamma code of 13 is the concatenation of length and offset: 1110101 128 .3 Gamma codes We can compress better with bit-level codes The Gamma code is the best known of these. Represent a gap G as a pair length and offset offset is G in binary. this is 3.Advanced Topics in Information Systems: Information Retrieval Sec.

11111111 11111111110.101 11110. 5.0 10.0000000001 129 .1 110.1000 111111110.Advanced Topics in Information Systems: Information Retrieval Sec.001 1110.00 1110.3 Gamma code examples number 0 1 2 3 4 9 13 24 511 1025 0 10 10 110 1110 1110 11110 111111110 11111111110 0 1 00 001 101 1000 11111111 0000000001 length offset γ-code none 0 10.

5.3 Gamma code properties G is encoded using 2 log G + 1 bits Length of offset is log G bits Length of length is log G + 1 bits All gamma codes have an odd number of bits Almost within a factor of 2 of best possible. like VB Gamma code can be used for any distribution Gamma code is parameter-free 130 .Advanced Topics in Information Systems: Information Retrieval Sec. log2 G Gamma code is uniquely prefix-decodable.

16. variable byte is conceptually simpler at little additional space cost 131 . 32. 64 bits Operations that cross word boundaries are slower Compressing and manipulating at the granularity of bits can be slow Variable byte encoding is aligned and thus potentially more efficient Regardless of efficiency.3 Gamma seldom used in practice Machines have word boundaries – 8. 5.Advanced Topics in Information Systems: Information Retrieval Sec.

1 5.0 250.0 40. fixed-width dictionary.3 RCV1 compression Data structure dictionary.6 7.0 101. 5.0 400.0 116.600.Advanced Topics in Information Systems: Information Retrieval Sec.000. term pointers into string with blocking.0 960. k = 4 with blocking & front coding collection (text. variable byte encoded postings. xml markup etc) collection (text) Term-doc incidence matrix postings. uncompressed (20 bits) postings.2 7.0 132 . uncompressed (32-bit words) postings. γ−encoded Size in MB 11.9 3.

Advanced Topics in Information Systems: Information Retrieval

Sec. 5.3

**Index compression summary
**

We can now create an index for highly efficient Boolean retrieval that is very space efficient Only 4% of the total size of the collection Only 10-15% of the total size of the text in the collection However, we’ve ignored positional information Hence, space savings are less for indexes used in practice

But techniques substantially the same.

133

Sec. 5.3

**SCORING, TERM WEIGHTING AND THE VECTOR SPACE MODEL
**

134

Advanced Topics in Information Systems: Information Retrieval

Ch. 6

Ranked retrieval

Thus far, our queries have all been Boolean.

Documents either match or don’t.

** Good for expert users with precise understanding of their needs and the collection.
**

Also good for applications: Applications can easily consume 1000s of results.

** Not good for the majority of users.
**

Most users incapable of writing Boolean queries (or they are, but they think it’s too much work). Most users don’t want to wade through 1000s of results.

This is particularly true of web search.

135

Advanced Topics in Information Systems: Information Retrieval Ch. 6 Problem with Boolean search: feast or famine Boolean queries often result in either too few (=0) or too many (1000s) results. Query 1: “standard user dlink 650” → 200. OR gives too many 136 . AND gives too few.000 hits Query 2: “standard user dlink 650 no card found”: 0 hits It takes a lot of skill to come up with a query that produces a manageable number of hits.

Advanced Topics in Information Systems: Information Retrieval Ranked retrieval models Rather than a set of documents satisfying a query expression. there are two separate choices here. the user’s query is just one or more words in a human language In principle. but in practice. in ranked retrieval models. ranked retrieval models have normally been associated with free text queries and vice versa 137 . the system returns an ordering over the (top) documents in the collection with respect to a query Free text queries: Rather than a query language of operators and expressions.

6 Feast or famine: not a problem in ranked retrieval When a system produces a ranked result set. the size of the result set is not an issue We just show the top k ( ≈ 10) results We don’t overwhelm the user Premise: the ranking algorithm works 138 . large result sets are not an issue Indeed.Advanced Topics in Information Systems: Information Retrieval Ch.

139 . 1] – to each document This score measures how well document and query “match”.Advanced Topics in Information Systems: Information Retrieval Ch. 6 Scoring as the basis of ranked retrieval We wish to return in order the documents most likely to be useful to the searcher How can we rank-order the documents in the collection with respect to a query? Assign a score – say in [0.

Advanced Topics in Information Systems: Information Retrieval Ch. 140 . the higher the score (should be) We will look at a number of alternatives for this. 6 Query-document matching scores We need a way of assigning a score to a query/document pair Let’s start with a one-term query If the query term does not occur in the document: score should be 0 The more frequent the query term in the document.

A) = 1 jaccard(A. 141 .Advanced Topics in Information Systems: Information Retrieval Ch. 6 Take 1: Jaccard coefficient Recall from Lecture 3: A commonly used measure of overlap of two sets A and B jaccard(A. Always assigns a number between 0 and 1.B) = |A ∩ B| / |A ∪ B| jaccard(A.B) = 0 if A ∩ B = 0 A and B don’t have to be the same size.

6 Jaccard coefficient: Scoring example What is the query-document match score that the Jaccard coefficient computes for each of the two documents below? Query: ides of march Document 1: caesar died in march Document 2: the long march 142 .Advanced Topics in Information Systems: Information Retrieval Ch.

Advanced Topics in Information Systems: Information Retrieval Ch. . instead of |A ∩ B|/|A ∪ B|| (Jaccard) A B | for length normalization. . we’ll use A B | / | . Jaccard doesn’t consider this information We need a more sophisticated way of normalizing for length Later in this lecture. 6 Issues with Jaccard for scoring It doesn’t consider term frequency (how many times a term occurs in a document) Rare terms in a collection are more informative than frequent terms. 143 .

Advanced Topics in Information Systems: Information Retrieval Sec.2 Recall (Lecture 1): Binary termdocument incidence matrix Antony and Cleopatra Julius Caesar The Tempest Hamlet Othello Macbeth Antony Brutus Caesar Calpurnia Cleopatra mercy worser 1 1 1 0 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 1 1 0 1 1 0 0 1 1 0 0 1 0 0 1 1 1 0 1 0 0 1 0 Each document is represented by a binary vector ∈ {0. 6.1}|V| 144 .

Advanced Topics in Information Systems: Information Retrieval Sec.2 Term-document count matrices Consider the number of occurrences of a term in a document: Each document is a count vector in ℕv: a column below Antony and Cleopatra Julius Caesar The Tempest Hamlet Othello Macbeth Antony Brutus Caesar Calpurnia Cleopatra mercy worser 157 4 232 0 57 2 2 73 157 227 10 0 0 0 0 0 0 0 0 3 1 0 1 2 0 0 5 1 0 0 1 0 0 5 1 0 0 1 0 0 1 0 145 . 6.

For now: bag of words model 146 .Advanced Topics in Information Systems: Information Retrieval Bag of words model Vector representation doesn’t consider the ordering of words in a document John is quicker than Mary and Mary is quicker than John have the same vectors This is called the bag of words model. this is a step back: The positional index was able to distinguish these two documents. We will look at “recovering” positional information later in this course. In a sense.

d of term t in document d is defined as the number of times that t occurs in d. IR 147 .Advanced Topics in Information Systems: Information Retrieval Term frequency tf The term frequency tft. But not 10 times more relevant. Relevance does not increase proportionally NB: frequency = count in with term frequency. But how? Raw term frequency is not what we want: A document with 10 occurrences of the term is more relevant than a document with 1 occurrence of the term. We want to use tf when computing querydocument match scores.

= 0.Advanced Topics in Information Systems: Information Retrieval Sec.d .d >0 otherw ise 0 → 0. 148 .2 Log-frequency weighting The log frequency weight of term t in d is wt. 10 → 2. if tf t. 1 → 1. 6. 1000 → 4.d ) The score is 0 if none of the query terms is present in the document. etc.3. 2 → 1. Score for a document-query pair: sum over terms t in both q and d: score = ∑t∈q∩d (1 + log tf t .d 1 +log 10 tf t.

149 .g. arachnocentric) A document containing this term is very likely to be relevant to the query arachnocentric → We want a high weight for rare terms like arachnocentric.1 Document frequency Rare terms are more informative than frequent terms Recall stop words Consider a term in the query that is rare in the collection (e.Advanced Topics in Information Systems: Information Retrieval Sec. 6..2.

We will use document frequency (df) to capture this. 150 . and line But lower weights than for rare terms. increase. 6.Advanced Topics in Information Systems: Information Retrieval Sec.. line) A document containing such a term is more likely to be relevant than a document that doesn’t But it’s not a sure indicator of relevance.1 Document frequency.g. high. continued Frequent terms are less informative than rare terms Consider a query term that is frequent in the collection (e. we want high positive weights for words like high. → For frequent terms. increase.2.

151 .Advanced Topics in Information Systems: Information Retrieval Sec. 6. Will turn out the base of the log is immaterial.2.1 idf weight dft is the document frequency of t: the number of documents that contain t dft is an inverse measure of the informativeness of t dft ≤ N We define the idf (inverse document frequency) of t by idf t = log10 ( N/df t ) We use log (N/dft) instead of N/dft to “dampen” the effect of idf.

6. suppose N = 1 million dft 1 100 1.000 100.Advanced Topics in Information Systems: Information Retrieval Sec.000.000 idft term calpurnia animal sunday fly under the idf t = log10 ( N/df t ) There is one idf value for each term t in a collection.1 idf example.000 10.000 1. 152 .2.

idf weighting makes occurrences of capricious count for much more in the final document ranking than occurrences of person. 153 . like iPhone idf has no effect on ranking one term queries idf affects the ranking of documents for queries with at least two terms For the query capricious person.Advanced Topics in Information Systems: Information Retrieval Effect of idf on ranking Does idf have an effect on ranking for one-term queries.

2.1 Collection vs. Document frequency The collection frequency of t is the number of occurrences of t in the collection. Example: Word Collection frequency Document frequency insurance try 10440 10422 3997 8760 Which word is a better search term (and should get a higher weight)? 154 . counting multiple occurrences.Advanced Topics in Information Systems: Information Retrieval Sec. 6.

tf x idf Increases with the number of occurrences within a document Increases with the rarity of the term in the collection 155 .d = (1 + log tf t .idf. not a minus sign! Alternative names: tf.Advanced Topics in Information Systems: Information Retrieval Sec.2.d ) × log10 ( N / df t ) Best known weighting scheme in information retrieval Note: the “-” in tf-idf is a hyphen.2 tf-idf weighting The tf-idf weight of a term is the product of its tf weight and its idf weight. 6. w t .

d) = t ￎ qￎ d tf.d 156 . 6.idft.2.Advanced Topics in Information Systems: Information Retrieval Sec.2 Final ranking of documents for a query Score(q.

12 4.9 0.85 1.25 1.3 Binary → count → weight matrix Antony and Cleopatra Julius Caesar The Tempest Hamlet Othello Macbeth Antony Brutus Caesar Calpurnia Cleopatra mercy worser 5.59 0 2.95 Each document is now represented by a real-valued vector of tf-idf weights ∈ R|V| 157 .18 6.35 0 0 0 0 0.15 0 0 0.1 2.88 1.25 0.51 0 0 0.25 0.37 3.25 0 0 5.51 1.54 0 0 0 0 0 0 0 0 1.54 1.Advanced Topics in Information Systems: Information Retrieval Sec. 6.21 8.11 0 1 1.

most entries are zero.3 Documents as vectors So we have a |V|-dimensional vector space Terms are axes of the space Documents are points or vectors in this space Very high-dimensional: tens of millions of dimensions when you apply this to a web search engine These are very sparse vectors . 158 .Advanced Topics in Information Systems: Information Retrieval Sec. 6.

6.Advanced Topics in Information Systems: Information Retrieval Sec. Instead: rank more relevant documents higher than less relevant documents 159 .3 Queries as vectors Key idea 1: Do the same for queries: represent them as vectors in the space Key idea 2: Rank documents according to their proximity to the query in this space proximity = similarity of vectors proximity ≈ inverse of distance Recall: We do this because we want to get away from the you’re-either-in-or-out Boolean model.

because Euclidean distance is large for vectors of different lengths. . . 6. .3 Formalizing vector space proximity First cut: distance between two points ( = distance between the end points of the two vectors) Euclidean distance? Euclidean distance is a bad idea . . 160 . .Advanced Topics in Information Systems: Information Retrieval Sec.

Advanced Topics in Information Systems: Information Retrieval Sec. 6.3 Why distance is a bad idea The Euclidean distance between q and d2 is large even though the distribution of terms in the query q and the distribution of terms in the document d2 are very similar. 161 .

“Semantically” d and d′ have the same content The Euclidean distance between the two documents can be quite large The angle between the two documents is 0. corresponding to maximal similarity.3 Use angle instead of distance Thought experiment: take a document d and append it to itself. 162 . Key idea: Rank documents according to angle with query. Call this document d′. 6.Advanced Topics in Information Systems: Information Retrieval Sec.

180o] 163 .3 From angles to cosines The following two notions are equivalent.document) Cosine is a monotonically decreasing function for the interval [0o. 6. Rank documents in decreasing order of the angle between query and document Rank documents in increasing order of cosine(query.Advanced Topics in Information Systems: Information Retrieval Sec.

6.3 From angles to cosines But how – and why – should we be computing cosines? 164 .Advanced Topics in Information Systems: Information Retrieval Sec.

3 Length normalization A vector can be (length-) normalized by dividing each of its components by its length – for this we use the L2 norm: x2= xi2 ∑i Dividing a vector by its L2 norm makes it a unit (length) vector (on surface of unit hypersphere) Effect on the two documents d and d′ (d appended to itself) from earlier slide: they have identical vectors after length-normalization. 6. Long and short documents now have comparable weights 165 .Advanced Topics in Information Systems: Information Retrieval Sec.

Advanced Topics in Information Systems: Information Retrieval Sec.document) Dot product Unit vectors q•d cos( q . d ) = = qd q • q d = d ∑ ∑ V qi d i i =1 d i2 ∑i =1 V V 2 i =1 i q qi is the tf-idf weight of term i in the query di is the tf-idf weight of term i in the document cos(q. 166 .3 cosine(query. the cosine of the angle between q and d. 6. equivalently.d) is the cosine similarity of q and d … or.

d ) = q ￎ d = V i=1 qi di for q. cosine similarity is simply the dot product (or scalar product): r r r r cos(q. 167 . d length-normalized.Advanced Topics in Information Systems: Information Retrieval Cosine for length-normalized vectors For length-normalized vectors.

Advanced Topics in Information Systems: Information Retrieval Cosine similarity illustrated 168 .

we don’t do idf weighting. 6. 169 .3 Cosine similarity amongst 3 documents How similar are the novels SaS: Sense and Sensibility PaP: Pride and Prejudice. and WH: Wuthering Heights? term affection jealous gossip wuthering SaS 115 10 2 0 PaP 58 7 0 0 WH 20 11 6 38 Term frequencies (counts) Note: To simplify this example.Advanced Topics in Information Systems: Information Retrieval Sec.

WH) ≈ 0.335 × 0.00 1.0 + 0.588 cos(SaS.405 0.335 0 PaP 0.555 + 0.PaP) ≈ 0.06 2.30 2.WH)? 170 .69 Why do we have cos(SaS.0 × 0.524 0.30 0 PaP 2.76 1.3 3 documents example contd.04 1.PaP) > cos(SAS.85 0 0 WH 2.58 After length normalization term affection jealous gossip wuthering SaS 0. Log frequency weighting term affection jealous gossip wuthering SaS 3.789 0.465 0.Advanced Topics in Information Systems: Information Retrieval Sec.79 cos(PaP.WH) ≈ 0.0 ≈ 0.832 + 0.789 × 0.832 0. 6.94 cos(SaS.515 × 0.515 0.78 2.555 0 0 WH 0.

6.3 Computing cosine scores 171 .Advanced Topics in Information Systems: Information Retrieval Sec.

Advanced Topics in Information Systems: Information Retrieval Sec.4 tf-idf weighting has many variants Columns headed ‘n’ are acronyms for weight schemes. 6. Why is the base of the log in idf immaterial? 172 .

with the notation ddd. no normalization … 173 . documents SMART Notation: denotes the combination in use in an engine. idf A bad idea? (t in second column).4 Weighting may differ in queries vs documents Many search engines allow for different weightings for queries vs. using the acronyms from the previous table A very standard weighting scheme is: lnc. no idf and cosine normalization Query: logarithmic tf (l in leftmost column).qqq. 6.ltc Document: logarithmic tf (l as first character).Advanced Topics in Information Systems: Information Retrieval Sec.

3 1.27 0.Advanced Topics in Information Systems: Information Retrieval Sec.53 Prod Exercise: what is N.52 0.4 tf-idf example: lnc.52 0 0.52 0.0 3.0 wt 0 1.53 = 0.92 174 Score = 0+0+0.0 n’lize 0 0.32 ￎ 1.8 .34 0.0 3.3 n’lize 0.3 2.68 0 0 0.27+0. the number of docs? Doc length = 12 + 0 2 + 12 + 1.3 wt 1 0 1 1. 6.3 2.78 Docu ment tf-raw 1 0 1 2 tf-wt 1 0 1 1.ltc Document: car insurance auto insurance Query: best car insurance Term Que ry tfraw auto best car insurance 0 1 1 1 tf-wt 0 1 1 1 df 5000 50000 10000 1000 idf 2.

g.Advanced Topics in Information Systems: Information Retrieval Summary – vector space ranking Represent the query as a weighted tf-idf vector Represent each document as a weighted tf-idf vector Compute the cosine similarity score for the query vector and each document vector Rank documents with respect to the query by score Return the top K (e. K = 10) to the user 175 ..