Introduction to Information Retrieval

Introduction to
Information Retrieval
Chapter 4: Index Construction
By
Binita Kumari
Asst. Professor
CSE, ITER,SOADU
Introduction to Information Retrieval

Plan
 Last lecture:
 Dictionary data structures a-hu
hy-m
n-z

 Tolerant retrieval
 Wildcards
 Spell correction
$m mace madden
 Soundex
mo among amortize

 This time: on abandon among

 Index construction
Introduction to Information Retrieval Ch. 4

Index construction
 How do we construct an index?
 What strategies can we use with limited main
memory?
Introduction to Information Retrieval Sec. 4.1

Hardware basics
 Many design decisions in information retrieval are
based on the characteristics of hardware
 We begin by reviewing hardware basics
Introduction to Information Retrieval Sec. 4.1

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.
 Therefore: Transferring one large chunk of data from
disk to memory is faster than transferring many small
chunks.
 Disk I/O is block-based: Reading and writing of entire
blocks (as opposed to smaller chunks).
 Block sizes: 8KB to 256 KB.
Introduction to Information Retrieval Sec. 4.1

Hardware basics
 Servers used in IR systems now typically have several
GB of main memory, sometimes tens of GB.
 Available disk space is several (2–3) orders of
magnitude larger.
 Fault tolerance is very expensive: It’s much cheaper
to use many regular machines rather than one fault
tolerant machine.
Introduction to Information Retrieval Sec. 4.1

Hardware assumptions for this lecture

 symbol statistic value
 s average seek time 5 ms = 5 x 10−3 s
 b transfer time per byte 0.02 μs = 2 x 10−8 s
 processor’s clock rate 109 s−1
 p low-level operation 0.01 μs = 10−8 s
(e.g., compare & swap a word)
 size of main memory several GB
 size of disk space 1 TB or more
Introduction to Information Retrieval Sec. 4.2

RCV1: Our collection for this lecture

 Shakespeare’s collected works definitely aren’t large
enough for demonstrating many of the points in this
course.
 The collection we’ll use isn’t really large enough
either, but it’s publicly available and is at least a more
plausible example.
 As an example for applying scalable index
construction algorithms, we will use the Reuters
RCV1 collection.
 This is one year of Reuters newswire (part of 1995
and 1996)
Introduction to Information Retrieval Sec. 4.2

A Reuters RCV1 document

Introduction to Information Retrieval Sec. 4.2

Reuters RCV1 statistics

 symbol statistic value
 N documents 800,000
 L avg. # tokens per doc 200
 M terms (= word types) 400,000
 avg. # bytes per token 6
(incl. spaces/punct.)
 avg. # bytes per token 4.5
(without spaces/punct.)
 avg. # bytes per term 7.5
 non-positional postings 100,000,000
4.5 bytes per word token vs. 7.5 bytes per word type: why?
Introduction to Information Retrieval Sec. 4.2

Term Doc #

Recall IIR 1 index construction I

did
1
1
enact 1
julius 1

 Documents are parsed to extract words and these caesar

I
1
1
are saved with the Document ID. was 1
killed 1
i' 1
the 1
capitol 1
brutus 1
killed 1
me 1
Doc 1 Doc 2 so 2
let 2
it 2
be 2
I did enact Julius So let it be with with 2

Caesar I was killed Caesar. The noble

caesar
the
2
2
i' the Capitol; Brutus hath told you noble
brutus
2
2
Brutus killed me. Caesar was ambitious hath 2
told 2
you 2
caesar 2
was 2
ambitious 2
Introduction to Information Retrieval Sec. 4.2

Key step Term

I
did
Doc #
1
1
Term
ambitious
be
Doc #
2
2
enact 1 brutus 1
julius 1 brutus 2
 After all documents have been caesar
I
1
1
capitol
caesar
1
1
parsed, the inverted file is was
killed
1
1
caesar
caesar
2
2
sorted by terms. i' 1 did 1
the 1 enact 1
capitol 1 hath 1
brutus 1 I 1
killed 1 I 1

We focus on this sort step. me

so
1
2
i'
it
1
2

We have 100M items to sort. let

it
2
2
julius
killed
1
1
be 2 killed 1
with 2 let 2
caesar 2 me 1
the 2 noble 2
noble 2 so 2
brutus 2 the 1
hath 2 the 2
told 2 told 2
you 2 you 2
caesar 2 was 1
was 2 was 2
ambitious 2 with 2
Introduction to Information Retrieval Sec. 4.2

Scaling index construction

 In-memory index construction does not scale
 Can’t stuff entire collection into memory, sort, then write
back
 How can we construct an index for very large
collections?
 Taking into account the hardware constraints we just
learned about . . .
 Memory, disk, speed, etc.
Introduction to Information Retrieval Sec. 4.2

Sort-based index construction

 As we build the index, we parse docs one at a time.
 While building the index, we cannot easily exploit
compression tricks (you can, but much more complex)
 The final postings for any term are incomplete until the end.
 At 12 bytes per non-positional postings entry (term, doc,
freq), demands a lot of space for large collections.
 T = 100,000,000 in the case of RCV1
 So … we can do this in memory in 2009, but typical
collections are much larger. E.g., the New York Times
provides an index of >150 years of newswire
 Thus: We need to store intermediate results on disk.
Introduction to Information Retrieval Sec. 4.2

Sort using disk as “memory”?

 Can we use the same index construction algorithm
for larger collections, but by using disk instead of
memory?
 No: Sorting T = 100,000,000 records on disk is too
slow – too many disk seeks.
 We need an external sorting algorithm.
Introduction to Information Retrieval Sec. 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, and N items could be

sorted with N log2N comparisons, how long would this take?
Introduction to Information Retrieval Sec. 4.2

BSBI: Blocked sort-based Indexing

(Sorting with fewer disk seeks)
 12-byte (4+4+4) records (term, doc, freq).
 These are generated as we parse docs.
 Must now sort 100M such 12-byte records by term.
 Define a Block ~ 10M such records
 Can easily fit a couple into memory.
 Will have 10 such blocks to start with.
 Basic idea of algorithm:
 Accumulate postings for each block, sort, write to disk.
 Then merge the blocks into one long sorted order.
Introduction to Information Retrieval Sec. 4.2
Introduction to Information Retrieval Sec. 4.2

Sorting 10 blocks of 10M records

 First, 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.
 10 times this estimate – gives us 10 sorted runs of
10M records each.
 Done straightforwardly, need 2 copies of data on disk
 But can optimize this
Introduction to Information Retrieval Sec. 4.2
Introduction to Information Retrieval Sec. 4.2

How to merge the sorted runs?

 Can do binary merges, with a merge tree of log210 = 4 layers.
 During each layer, read into memory runs in blocks of 10M,
merge, write back.

1
1 2
2 Merged run.
3 4
3

4
Runs being
merged.
Disk
Introduction to Information Retrieval Sec. 4.2

How to merge the sorted runs?

 But it is more efficient to do a multi-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
Introduction to Information Retrieval Sec. 4.3

Remaining problem with sort-based

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, we could work with term,docID postings
instead of termID,docID postings . . .
 . . . but then intermediate files become very large.
(We would end up with a scalable, but very slow
index construction method.)
Introduction to Information Retrieval Sec. 4.3

SPIMI:
Single-pass in-memory indexing
 Key idea 1: Generate separate dictionaries for each
block – no need to maintain term-termID mapping
across blocks.
 Key idea 2: Don’t sort. Accumulate postings in
postings lists as they occur.
 With these two ideas we can generate a complete
inverted index for each block.
 These separate indexes can then be merged into one
big index.
Introduction to Information Retrieval Sec. 4.3

SPIMI-Invert

 Merging of blocks is analogous to BSBI.

Introduction to Information Retrieval Sec. 4.3

SPIMI: Compression
 Compression makes SPIMI even more efficient.
 Compression of terms
 Compression of postings
 See next lecture
Introduction to 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?
Introduction to Information Retrieval Sec. 4.4

Web search engine data centers

 Web search data centers (Google, Bing, Baidu)
mainly contain commodity machines.
 Data centers are distributed around the world.
 Estimate: Google ~1 million servers, 3 million
processors/cores (Gartner 2007)
Introduction to Information Retrieval Sec. 4.4

Massive data centers

 If in a non-fault-tolerant system with 1000 nodes,
each node has 99.9% uptime, what is the uptime of
the system?
 Answer: 63%
 Exercise: Calculate the number of servers failing per
minute for an installation of 1 million servers.
Introduction to Information Retrieval Sec. 4.4

Distributed indexing
 Maintain a master machine directing the indexing job
– considered “safe”.
 Break up indexing into sets of (parallel) tasks.
 Master machine assigns each task to an idle machine
from a pool.
Introduction to Information Retrieval Sec. 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)
Introduction to Information Retrieval Sec. 4.4

Parsers
 Master assigns a split to an idle parser machine
 Parser reads a document at a time and emits (term,
doc) pairs
 Parser writes pairs into j partitions
 Each partition is for a range of terms’ first letters
 (e.g., a-f, g-p, q-z) – here j = 3.
 Now to complete the index inversion
Introduction to Information Retrieval Sec. 4.4

Inverters
 An inverter collects all (term,doc) pairs (= postings)
for one term-partition.
 Sorts and writes to postings lists
Introduction to Information Retrieval Sec. 4.4

Data flow
assign Master assign
Postings

Parser a-f g-p q-z Inverter a-f

Parser a-f g-p q-z

Inverter g-p

splits Inverter q-z

Parser a-f g-p q-z

Map Reduce
Segment files
phase phase
Introduction to Information Retrieval Sec. 4.4

MapReduce
 The index construction algorithm we just described is
an instance of MapReduce.
 MapReduce (Dean and Ghemawat 2004) is a robust
and conceptually simple framework for distributed
computing …
 … without having to write code for the distribution
part.
 They describe the Google indexing system (ca. 2002)
as consisting of a number of phases, each
implemented in MapReduce.
Introduction to Information Retrieval Sec. 4.4

MapReduce
 Index construction was just one phase.
 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’ll discuss in the web part of the course, most
search engines use a document-partitioned index …
better load balancing, etc.
Introduction to Information Retrieval Sec. 4.4

Schema for index construction in

MapReduce
 Schema of map and reduce functions
 map: input → list(k, v) reduce: (k,list(v)) → output
 Instantiation of the schema for index construction
 map: collection → list(termID, docID)
 reduce: (<termID1, list(docID)>, <termID2, list(docID)>, …) →
(postings list1, postings list2, …)
Introduction to Information Retrieval

Example for index construction

 Map:
 d1 : C came, C c’ed.
 d2 : C died. →
 <C,d1>, <came,d1>, <C,d1>, <c’ed, d1>, <C, d2>,
<died,d2>
 Reduce:
 (<C,(d1,d2,d1)>, <died,(d2)>, <came,(d1)>, <c’ed,(d1)>)
→ (<C,(d1:2,d2:1)>, <died,(d2:1)>, <came,(d1:1)>,
<c’ed,(d1:1)>)

38
Introduction to Information Retrieval Sec. 4.5

Dynamic indexing
 Up to now, we have assumed that collections are
static.
 They rarely are:
 Documents come in over time and need to be inserted.
 Documents are deleted and modified.
 This means that the dictionary and postings lists have
to be modified:
 Postings updates for terms already in dictionary
 New terms added to dictionary
Introduction to Information Retrieval Sec. 4.5

Simplest approach
 Maintain “big” main index
 New docs go into “small” auxiliary index
 Search across both, merge results
 Deletions
 Invalidation bit-vector for deleted docs
 Filter docs output on a search result by this invalidation
bit-vector
 Periodically, re-index into one main index
Introduction to Information Retrieval Sec. 4.5

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.
 Merge is the same as a simple append.
 But then we would need a lot of files – inefficient for OS.
 Assumption for the rest of the lecture: The index is one big
file.
 In reality: Use a scheme somewhere in between (e.g., split
very large postings lists, collect postings lists of length 1 in one
file etc.)
Introduction to Information Retrieval Sec. 4.5

Logarithmic merge
 Maintain a series of indexes, each twice as large as
the previous one
 At any time, some of these powers of 2 are instantiated
 Keep smallest (Z0) in memory
 Larger ones (I0, I1, …) on disk
 If Z0 gets too big (> n), 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
Introduction to Information Retrieval Sec. 4.5
Introduction to 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, 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
Introduction to Information Retrieval Sec. 4.5

Further issues with multiple indexes

 Collection-wide statistics are hard to maintain
 E.g., 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
Introduction to Information Retrieval Sec. 4.5

Dynamic indexing at search engines

 All the large search engines now do dynamic
indexing
 Their indices have frequent incremental changes
 News items, blogs, 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, and
the old index is deleted
Introduction to Information Retrieval Sec. 4.5
Introduction to Information Retrieval Sec. 4.5

Other sorts of indexes

 Positional indexes
 Same sort of sorting problem … just larger Why?
 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.
 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
Introduction to Information Retrieval Ch. 4

Resources for today’s lecture

 Chapter 4 of IIR
 MG Chapter 5
 Original publication on MapReduce: Dean and
Ghemawat (2004)
 Original publication on SPIMI: Heinz and Zobel (2003)