Vector Space Model : Ranking Revisited

Adapted from Lectures by Prabhakar Raghavan (Yahoo and Stanford) and Christopher Manning (Stanford)

Prasad

L09VSM-Ranking

1

Recap: tf-idf weighting
n 

The tf-idf weight of a term is the product of its tf weight and its idf weight.

w t ,d = (1 + log tft ,d ) × log10 ( N / df t )
n  n 

n 

Best known weighting scheme in information retrieval Increases with the number of occurrences within a document Increases with the rarity of the term in the collection
Prasad L09VSM-Ranking 2

Recap: Queries as vectors
n 

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

n 

n 

Prasad

L09VSM-Ranking

3

Recap: cosine(query,document)
Dot product Unit vectors

      q•d q d cos( q, d ) =   =  •  = q d qd

∑ ∑
V

V

i =1 i i

qd

2 i =1 i

q

V

i =1

d

2 i

cos(q,d) is the cosine similarity of q and d … or, equivalently, the cosine of the angle between q and d.

Prasad

L09VSM-Ranking

4

This lecture
Speeding up vector space ranking n  Putting together a complete search system
n 
n 

Will require learning about a number of miscellaneous topics and heuristics

Prasad

L09VSM-Ranking

5

Computing cosine scores

Efficient cosine ranking
n 

n 

Find the K docs in the collection “nearest” to the query ⇒ K largest query-doc cosines. Efficient ranking:
Computing a single cosine efficiently. n  Choosing the K largest cosine values efficiently.
n 
n 

Can we do this without computing all N cosines?
L09VSM-Ranking 7

Prasad

(cont’d)
n 

What we’re doing in effect: solving the Knearest neighbor problem for a query vector In general, we do not know how to do this efficiently for high-dimensional spaces But it is solvable for short queries, and standard indexes support this well
L09VSM-Ranking 8

n 

n 

Prasad

Special case – unweighted queries
n 

No weighting on query terms n  Assume each query term occurs only once Then for ranking, don’t need to normalize query vector n  Slight simplification

n 

Prasad

L09VSM-Ranking

9

7.1

Faster cosine: unweighted query

Read redundant?

7.1

Computing the K largest cosines: selection vs. sorting
n 

Typically we want to retrieve the top K docs (in the cosine ranking for the query) n  not to totally order all docs in the collection Can we pick off docs with K highest cosines? Let J = number of docs with nonzero cosines n  We seek the K best of these J
L09VSM-Ranking 11

n 

n 

Prasad

7.1

Use heap for selecting top K
n 

n 

n 

Binary tree in which each node’s value > the values of children Takes 2J operations to construct, then each of K “winners” read off in 2log J steps. For J=10M, K=100, this is about 10% of the cost of sorting.
1 .9 .3 .8 .1 .3 .1
12

Prasad

L09VSM-Ranking

7.1

Bottlenecks
n 

Primary computational bottleneck in scoring: cosine computation Can we avoid all this computation? Yes, but may sometimes get it wrong n  a doc not in the top K may creep into the list of K output docs n  Is this such a bad thing?
L09VSM-Ranking

n  n 

Prasad

7.1.1

13

Cosine similarity is only a proxy
n  n  n 

User has a task and a query formulation Cosine matches docs to query Thus cosine is anyway a proxy for user happiness If we get a list of K docs “close” to the top K by cosine measure, should be ok

n 

Prasad

L09VSM-Ranking

7.1.1

14

Generic approach
n 

Find a set A of contenders, with K < |A| << N
n 

n 

A does not necessarily contain the top K, but has many docs from among the top K Return the top K docs in A

n  n 

n 

Think of A as pruning non-contenders The same approach is also used for other (non-cosine) scoring functions Will look at several schemes following this approach
L09VSM-Ranking

Prasad

7.1.1

Index elimination
n 

Basic algorithm of Fig 7.1 considers only docs containing at least one query term Take this further:
n  n 

n 

Consider only high-idf query terms Consider only docs containing many query terms

Prasad

L09VSM-Ranking

7.1.2

16

High-idf query terms only
n  n 

For a query such as catcher in the rye Only accumulate scores from catcher and rye Intuition: in and the contribute little to the scores and don’t alter rank-ordering much Benefit:
n 

n 

n 

Postings of low-idf terms have many docs → these (many) docs get eliminated from A
L09VSM-Ranking

Prasad

7.1.2

17

Docs containing many query terms
n 

Any doc with at least one query term is a candidate for the top K output list For multi-term queries, only compute scores for docs containing several of the query terms
n  n 

n 

Say, at least 3 out of 4 Imposes a “soft conjunction” on queries seen on web search engines (early Google)

n 

Easy to implement in postings traversal
L09VSM-Ranking

Prasad

7.1.2

18

3 of 4 query terms
Antony Brutus Caesar Calpurnia 3 2 1 4 4 2 8 8 3 16 32 64 128 16 32 64 128 5 8 13 21 34

13 16 32

Scores only computed for 8, 16 and 32.
Prasad L09VSM-Ranking

7.1.2

19

Champion lists
n 

Precompute for each dictionary term t, the r docs of highest weight in t’s postings
n  n 

Call this the champion list for t (aka fancy list or top docs for t)

n  n 

Note that r has to be chosen at index time At query time, only compute scores for docs in the champion list of some query term
n 

Pick the K top-scoring docs from amongst these
L09VSM-Ranking

Prasad

7.1.3

20

Exercises
n 

How do Champion Lists relate to Index Elimination? Can they be used together? How can Champion Lists be implemented in an inverted index?
n 

n 

Note the champion list has nothing to do with small docIDs

Prasad

L09VSM-Ranking

7.1.3

21

Static quality scores
n 

n  n 

n 

We want top-ranking documents to be both relevant and authoritative Relevance is being modeled by cosine scores Authority is typically a query-independent property of a document Examples of authority signals
n  n  n  n  n 

Wikipedia among websites Articles in certain newspapers Quantitative A paper with many citations Many diggs, Y!buzzes or del.icio.us marks (Pagerank) 7.1.4

Modeling authority
n 

Assign to each document a query-independent quality score in [0,1] to each document d
n 

Denote this by g(d)

n 

Thus, a quantity like the number of citations is scaled into [0,1]
n 

Exercise: suggest a formula for this.

Prasad

L09VSM-Ranking

7.1.4

23

Net score
n 

Consider a simple total score combining cosine relevance and authority net-score(q,d) = g(d) + cosine(q,d)
n 

n 

n 

Can use some other linear combination than an equal weighting Indeed, any function of the two “signals” of user happiness – more later

n 

Now we seek the top K docs by net score
L09VSM-Ranking

Prasad

7.1.4

24

Top K by net score – fast methods
n  n  n 

First idea: Order all postings by g(d) Key: this is a common ordering for all postings Thus, can concurrently traverse query terms’ postings for
n  n 

Postings intersection Cosine score computation

n 

Exercise: write pseudocode for cosine score computation if postings are ordered by g(d)
L09VSM-Ranking

Prasad

7.1.4

25

Why order postings by g(d)?
n 

Under g(d)-ordering, top-scoring docs likely to appear early in postings traversal In time-bound applications (say, we have to return whatever search results we can in 50 ms), this allows us to stop postings traversal early
n 

n 

Short of computing scores for all docs in postings

Prasad

L09VSM-Ranking

7.1.4

26

Champion lists in g(d)-ordering
n 

Can combine champion lists with g(d)-ordering Maintain for each term a champion list of the r docs with highest g(d) + tf-idftd Seek top-K results from only the docs in these champion lists

n 

n 

Prasad

L09VSM-Ranking

7.1.4

27

High and low lists
n 

For each term, we maintain two postings lists called high and low
n 

Think of high as the champion list

n 

When traversing postings on a query, only traverse high lists first
n 

n 

If we get more than K docs, select the top K and stop Else proceed to get docs from the low lists

n 

n 

Can be used even for simple cosine scores, without global quality g(d) A means for segmenting index into two tiers

7.1.4

Impact-ordered postings
n 

n 

We only want to compute scores for docs for which wft,d is high enough We sort each postings list by wft,d Now: not all postings in a common order! How do we compute scores in order to pick off top K?
n 

n 

n 

Two ideas follow
L09VSM-Ranking

Prasad

7.1.5

29

1. Early termination
n 

When traversing t’s postings, stop early after either
n  n 

a fixed number of r docs wft,d drops below some threshold

n 

Take the union of the resulting sets of docs
n 

One from the postings of each query term

n 

Compute only the scores for docs in this union
L09VSM-Ranking

Prasad

7.1.5

30

2. idf-ordered terms
n  n 

When considering the postings of query terms Look at them in order of decreasing idf
n 

High idf terms likely to contribute most to score

n 

As we update score contribution from each query term
n 

Stop if doc scores relatively unchanged

n 

Can apply to cosine or some other net scores

Prasad

L09VSM-Ranking

7.1.5

31

Cluster pruning: preprocessing
Pick √N docs at random: call these leaders n  For every other doc, pre-compute nearest leader
n 

Docs attached to a leader: its followers; n  Likely: each leader has ~ √N followers.
n 
Prasad L09VSM-Ranking

7.1.6

32

Cluster pruning: query processing
n 

Process a query as follows:
Given query Q, find its nearest leader L. n  Seek K nearest docs from among L’s followers.
n 

Prasad

L09VSM-Ranking

7.1.6

33

Visualization

Query

Prasad

Leader

L09VSM-Ranking

Follower

7.1.6

34

Why use random sampling
n  n 

Fast Leaders reflect data distribution

Prasad

L09VSM-Ranking

7.1.6

35

General variants
n 

n 

n 

Have each follower attached to b1=3 (say) nearest leaders. From query, find b2=4 (say) nearest leaders and their followers. Can recur on leader/follower construction.

Prasad

L09VSM-Ranking

7.1.6

36

Exercises
n 

To find the nearest leader in step 1, how many cosine computations do we do?
n 

Why did we have √N in the first place?

n 

What is the effect of the constants b1, b2 on the previous slide? Devise an example where this is likely to fail – i.e., we miss one of the K nearest docs.
n 

n 

Likely under random sampling.
L09VSM-Ranking

Prasad

7.1.6

37

Parametric and zone indexes
n  n 

Thus far, a doc has been a sequence of terms In fact documents have multiple parts, some with special semantics:
n  n  n  n  n  n 

Author Title Date of publication Language Format etc.
L09VSM-Ranking 38

n 

These constitute the metadata about a document
6.1

Prasad

Fields
n 

We sometimes wish to search by these metadata
n 

E.g., find docs authored by William Shakespeare in the year 1601, containing alas poor Yorick

n  n  n 

Year = 1601 is an example of a field Also, author last name = shakespeare, etc Field or parametric index: postings for each field value
n 

Sometimes build range trees (e.g., for dates) (doc must be authored by shakespeare)
L09VSM-Ranking 39

n 

Field query typically treated as conjunction
n 

Prasad

6.1

Zone
n 

A zone is a region of the doc that can contain an arbitrary amount of text e.g.,
n  n  n 

Title Abstract References …

n 

n 

Build inverted indexes on zones as well to permit querying E.g., “find docs with merchant in the title zone and matching the query gentle rain”
L09VSM-Ranking 40

Prasad

6.1

Example zone indexes

Encode zones in dictionary vs. postings.
6.1

Tiered indexes
n 

Break postings up into a hierarchy of lists
n  n  n 

Most important … Least important

n  n 

n 

Can be done by g(d) or another measure Inverted index thus broken up into tiers of decreasing importance At query time use top tier unless it fails to yield K docs
n 

If so drop to lower tiers
L09VSM-Ranking

Prasad

7.2.1

42

Example tiered index

Prasad

7.2.1

Query term proximity
n 

n 

n 

n 

n 

Free text queries: just a set of terms typed into the query box – common on the web Users prefer docs in which query terms occur within close proximity of each other Let w be the smallest window in a doc containing all query terms, e.g., For the query strained mercy the smallest window in the doc The quality of mercy is not strained is 4 (words) Would like scoring function to take this into account – how? L09VSM-Ranking

7.2.2

Query parsers
n 

Free text query from user may in fact spawn one or more queries to the indexes, e.g. query rising interest rates
n  n 

n 

n 

Run the query as a phrase query If <K docs contain the phrase rising interest rates, run the two phrase queries rising interest and interest rates If we still have <K docs, run the vector space query rising interest rates Rank matching docs by vector space scoring
L09VSM-Ranking

n 

Prasad

This sequence is issued by a query parser
7.2.3
45

Aggregate scores
n 

We’ve seen that score functions can combine cosine, static quality, proximity, etc. How do we know the best combination? Some applications – expert-tuned Increasingly common: machine-learned

n  n 

n 

Prasad

L09VSM-Ranking

7.2.3

46

Putting it all together

Prasad

L09VSM-Ranking

7.2.4

47

Evaluation of IR Systems
"

Adapted from Lectures by Prabhakar Raghavan (Yahoo and Stanford) and Christopher Manning (Stanford)

Prasad

L10Evaluation

48

This lecture
n 

Results summaries:
n 

Making our good results usable to a user

n 

How do we know if our results are any good?
n 

Evaluating a search engine
n  n 

Benchmarks Precision and recall

Prasad

L10Evaluation

49

Result Summaries
n 

n 

Having ranked the documents matching a query, we wish to present a results list. Most commonly, a list of the document titles plus a short summary, aka “10 blue links”.

50

Summaries
n 

The title is typically automatically extracted from document metadata. What about the summaries?
n  n 

This description is crucial. User can identify good/relevant hits based on description.

n 

Two basic kinds:
n 

n 

A static summary of a document is always the same, regardless of the query that hit the doc. A dynamic summary is a query-dependent attempt to explain why the document was retrieved for the query at hand.
L10Evaluation 51

Prasad

Static summaries
n 

In typical systems, the static summary is a subset of the document.
n 

Simplest heuristic: the first 50 (or so – this can be varied) words of the document
n 

Summary cached at indexing time

n 

More sophisticated: extract from each document a set of “key” sentences
n  n 

Simple NLP heuristics to score each sentence Summary is made up of top-scoring sentences.

n 

Most sophisticated: NLP used to synthesize a summary
n 

Seldom used in IR (cf. text summarization work)
52

Dynamic summaries
n 

Present one or more “windows” within the document that contain several of the query terms
n 

“KWIC” snippets: Keyword in Context presentation

n 

Generated in conjunction with scoring
n  n 

If query found as a phrase, all or some occurrences of the phrase in the doc If not, document windows that contain multiple query terms

n 

The summary itself gives the entire content of the window – all terms, not only the query terms.

Generating dynamic summaries
n 

If we have only a positional index, we cannot (easily)
reconstruct context window surrounding hits. If we cache the documents at index time, then we can find windows in it, cueing from hits found in the positional index.
n 

n 

E.g., positional index says “the query is a phrase in position 4378” so we go to this position in the cached document and stream out the content Note: Cached copy can be outdated

n 

Most often, cache only a fixed-size prefix of the doc.
n 

Prasad

L10Evaluation

54

Dynamic summaries
n 

Producing good dynamic summaries is a tricky optimization problem
n 

n  n  n 

The real estate for the summary is normally small and fixed Want snippets to be long enough to be useful Want linguistically well-formed snippets Want snippets maximally informative about doc

n 

But users really like snippets, even if they complicate IR system design
L10Evaluation 55

Prasad

Alternative results presentations?
n  n 

An active area of HCI research An alternative: http://www.searchme.com / copies the idea of Apple’s Cover Flow for search results

Prasad

L10Evaluation

56

Evaluating search engines

Prasad

L10Evaluation

57

Measures for a search engine
n 

How fast does it index
n  n 

Number of documents/hour (Average document size) Latency as a function of index size Ability to express complex information needs Speed on complex queries

n 

How fast does it search
n 

n 

Expressiveness of query language
n  n 

n  n 

Uncluttered UI Is it free?
L10Evaluation 58

Prasad

Measures for a search engine
n 

n 

All of the preceding criteria are measurable: we can quantify speed/size; we can make expressiveness precise The key measure: user happiness
n 

What is this?
n  n 

Speed of response/size of index are factors But blindingly fast, useless answers won’t make a user happy

n 

Need a way of quantifying user happiness

Prasad

L10Evaluation

59

Data Retrieval vs Information Retrieval
n 

DR Performance Evaluation (after establishing correctness)
n  n 

Response time Index space

n 

IR Performance Evaluation
n 

How relevant is the answer set? (required to establish functional correctness, e.g., through benchmarks)
L10Evaluation 60

Prasad

Measuring user happiness
n 

Issue: who is the user we are trying to make happy?
n 

Depends on the setting/context

n 

Web engine: user finds what they want and return to the engine
n 

Can measure rate of return users

n 

eCommerce site: user finds what they want and make a purchase
n 

n 

Prasad

Is it the end-user, or the eCommerce site, whose happiness we measure? Measure time to purchase, or fraction of searchers who become buyers?
L10Evaluation

61

Measuring user happiness
n 

Enterprise (company/govt/academic): Care about “user productivity”
n 

n 

How much time do my users save when looking for information? Many other criteria having to do with breadth of access, secure access, etc.

Prasad

L10Evaluation

62

Happiness: elusive to measure
n 

Most common proxy: relevance of search results
n 

But how do you measure relevance?

n 

We will detail a methodology here, then examine its issues
1.  A benchmark document collection 2.  A benchmark suite of queries 3.  A usually binary assessment of either Relevant or Nonrelevant for each query and each document n  Some work on more-than-binary, but not the standard

n 

Relevant measurement requires 3 elements:

Prasad

L10Evaluation

63

Evaluating an IR system
n 

n 

Note: the information need is translated into a query Relevance is assessed relative to the information need, not the query
n 

n 

E.g., Information need: I'm looking for information on whether drinking red wine is more effective at reducing heart attack risks than white wine. Query: wine red white heart attack effective

n 

You evaluate whether the doc addresses the information need, not whether it has these words
L10Evaluation 64

Prasad

Difficulties with gauging Relevancy
n 

Relevancy, from a human standpoint, is: n  Subjective: Depends upon a specific user’s judgment. n  Situational: Relates to user’s current needs. n  Cognitive: Depends on human perception and behavior. n  Dynamic: Changes over time.
L10Evaluation 65

Prasad

Standard relevance benchmarks
n 

n 

TREC - National Institute of Standards and Technology (NIST) has run a large IR test bed for many years Reuters and other benchmark doc collections used

n 

“Retrieval tasks” specified
n 

sometimes as queries

n 

Human experts mark, for each query and for each doc, Relevant or Nonrelevant
n 

or at least for subset of docs that some system returned for that query
66

Prasad

Unranked retrieval evaluation: Precision and Recall
n 

n 

Precision: fraction of retrieved docs that are relevant = P(relevant|retrieved) Recall: fraction of relevant docs that are retrieved = P(retrieved|relevant)
Relevant Retrieved Not Retrieved
n  n 

Nonrelevant fp tn

tp fn

Prasad

Precision P = tp/(tp + fp) Recall R = tp/(tp + fn)
L10Evaluation

67