Scoring, Term Weighting

and Vector space model

Dr. Subrat Kumar Nayak
Associate Professor
Dept. of CSE, ITER, SOADU
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.
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” → 200,000 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.
 AND gives too few; OR gives too many
Ranked retrieval models
 Rather than a set of documents satisfying a query
expression, in ranked retrieval, the system returns an
ordering over the (top) documents in the collection for a
query
 Free text queries: Rather than a query language of
operators and expressions, the user’s query is just one or
more words in a human language
 In principle, there are two separate choices here, but in
practice, ranked retrieval has normally been associated
with free text queries and vice versa
Feast or famine: not a problem in ranked
retrieval
 When a system produces a ranked result set, large result
sets are not an issue
 Indeed, 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

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, 1] – to each document
 This score measures how well document and query
“match”.
Jaccard coefficient
 A common measure of overlap of two sets A and B
 jaccard(A,B) = |A ∩ B| / |A ∪ B|
 jaccard(A,A) = 1
 jaccard(A,B) = 0 if A ∩ B = 0
 A and B don’t have to be the same size.
 Always assigns a number between 0 and 1.
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
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. Jaccard doesn’t consider this
information
 We need a more sophisticated way of normalizing for
length
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, the
higher the score (should be)
 We will look at a number of alternatives for this.
Recall (Lecture 2): Binary term-document
incidence matrix

Antony and Cleopatra Julius Caesar The Tempest Hamlet Othello Macbeth
Antony 1 1 0 0 0 1
Brutus 1 1 0 1 0 0
Caesar 1 1 0 1 1 1
Calpurnia 0 1 0 0 0 0
Cleopatra 1 0 0 0 0 0
mercy 1 0 1 1 1 1
worser 1 0 1 1 1 0

Each document is represented by a binary vector ∈ {0,1}|V|

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 157 73 0 0 0 0
Brutus 4 157 0 1 0 0
Caesar 232 227 0 2 1 1
Calpurnia 0 10 0 0 0 0
Cleopatra 57 0 0 0 0 0
mercy 2 0 3 5 5 1
worser 2 0 1 1 1 0
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.
 In a sense, this is a step back: The positional index was able to
distinguish these two documents.
Term frequency tf
 The term frequency tft,d of term t in document d is defined as the
number of times that t occurs in d.
 Note: Frequency means count in IR
 We want to use tf when computing query-document match scores.
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.
 But not 10 times more relevant.
 Relevance does not increase proportionally with term frequency.
Log-frequency weighting
 The log frequency weight of term t in d is
1 + log10 tf t,d , if tf t,d  0
wt,d =
 0, otherwise
 0 → 0, 1 → 1, 2 → 1.3, 10 → 2, 1000 → 4, etc.
 Score for a document-query pair: sum over terms t
in both q and d:
 score = tqd (1 + log tf t ,d )

 The score is 0 if none of the query terms is present in

the document.
Rare terms are more informative
 Rare terms are more informative than frequent terms
 Recall stop words
 Consider a term in the query that is rare in the collection (e.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.
Collection vs. Document frequency
 Collection frequency of t is the number of occurrences of t in the
collection
 Document frequency of t is the number of documents in which t
occurs
Word Collection Document
 Example: frequency frequency

insurance 10440 3997

try 10422 8760

 Which word is for better search (gets higher weight)

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

 We use log (N/dft) instead of N/dft to “dampen” the effect of idf.

idf example, suppose N = 1 million
term dft idft
calpurnia 1 6
animal 100 4
sunday 1,000 3
fly 10,000 2
under 100,000 1
the 1,000,000 0

idf t = log10 ( N/df t )

There is one idf value for each term t in a collection.
Effect of idf on ranking
 Does idf have an effect on ranking for one-term queries,
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, idf weighting makes
occurrences of capricious count for much more in the
final document ranking than occurrences of person.
tf-idf weighting
 The tf-idf weight of a term is the product of its tf weight and its
idf weight.
w t ,d = log(1 + tf t ,d )  log10 ( N / df t )
 Best known weighting scheme in information retrieval
 Note: the “-” in tf-idf is a hyphen, not a minus sign!
 Alternative names: tf.idf, tf x idf
 Increases with the number of occurrences within a document
 Increases with the rarity of the term in the collection
 Score for a document given a query

Score(q,d) =  tf.idft,d
t qd

There are many variants

How “tf” is computed (with/without logs)
Whether the terms in the query are also
 weighted
…
Binary → count → weight matrix

Antony and Cleopatra Julius Caesar The Tempest Hamlet Othello Macbeth
Antony 5.25 3.18 0 0 0 0.35
Brutus 1.21 6.1 0 1 0 0
Caesar 8.59 2.54 0 1.51 0.25 0
Calpurnia 0 1.54 0 0 0 0
Cleopatra 2.85 0 0 0 0 0
mercy 1.51 0 1.9 0.12 5.25 0.88
worser 1.37 0 0.11 4.15 0.25 1.95

Each document is now represented by a real-valued

vector of tf-idf weights ∈ R|V|
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 - most entries are zero.
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
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 . . .
 . . . because Euclidean distance is large for vectors of
different lengths.
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.
Use angle instead of distance
 Thought experiment: take a document d and append it
to itself. Call this document d′.
 “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.

 Key idea: Rank documents according to angle with

query.
From angles to cosines
 The following two notions are equivalent.
 Rank documents in decreasing order of the angle between
query and document
 Rank documents in increasing order of cosine(query,document)
 Cosine is a monotonically decreasing function for the
interval [0o, 180o]
From angles to cosines

 But how should we be computing cosines?

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= i i
x 2

 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.
 Long and short documents now have comparable weights
cosine(query,document)
Dot product Unit vectors
  


V
  q•d q d q di
cos( q, d ) =   =  •  = i =1 i
q d
 i=1 i
V V
qd q2
d 2
i =1 i

qi is the weight of term i in the query

di is the weight of term i in the document

cos(q,d) is the cosine similarity of q and d … or,

equivalently, the cosine of the angle between q and d.
 Cosine for length-normalized vectors
 For length-normalized vectors, cosine similarity is simply the dot
product (or scalar product):

r r r r
cos(q, d ) = q • d =  qi di
V

i=1

for q, d length-normalized.


Cosine similarity illustrated
 Sec. 6.3

Cosine similarity amongst 3 documents

 How similar are
the novels term SaS PaP WH
 SaS: Sense and affection 115 58 20
Sensibility
jealous 10 7 11
 PaP: Pride and
Prejudice, and gossip 2 0 6

 WH:Wuthering wuthering 0 0 38

Heights?
Term frequencies (counts)

Note: To simplify this example, we don’t do idf weighting.

3 documents example contd.
Log frequency weighting After length normalization

term SaS PaP WH term SaS PaP WH

affection 3.06 2.76 2.30 affection 0.789 0.832 0.524
jealous 2.00 1.85 2.04 jealous 0.515 0.555 0.465
gossip 1.30 0 1.78 gossip 0.335 0 0.405
wuthering 0 0 2.58 wuthering 0 0 0.588

dot(SaS,PaP) ≈ 12.1 cos(SaS,PaP) ≈ 0.94

dot(SaS,WH) ≈ 13.4 cos(SaS,WH) ≈ 0.79
dot(PaP,WH) ≈ 10.1 cos(PaP,WH) ≈ 0.69
Computing cosine scores
Computing cosine scores
 Previous algorithm scores term-at-a-time (TAAT)
 Algorithm can be adapted to scoring document-at-a-
time (DAAT)
 Storing wt,d in each posting could be expensive
 …because we’d have to store a floating point number
 For tf-idf scoring, it suffices to store tft,d in the posting and idft in
the head of the postings list
 Extracting the top K items can be done with a priority
queue (e.g., a heap)