Improving Information Retrieval Performance using Word Embedding

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Improving Information Retrieval Performance using
Word Embedding
Dwaipayan Roy
Indian Statistical Institute, Kolkata
December 28, 2018
D. Roy (ISI, Kolkata) WE for IR December 28, 2018 1 / 78

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Information Retrieval
www.trec.nist.gov

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Information Retrieval: A Practical Application
www.google.com

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Information Retrieval: A Graphical Representation
Corpus
Index
IndexingInformation
Need
Query
Query
(Re)Formulation
Retrieval
System
Ranked
Docs.
d1
d2
d3
d4
d5
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Corpus
Index
IndexingInformation
Need
Query
Query
(Re)Formulation
Retrieval
Model
Ranked
Docs.
d1
d2
d3
d4
d5
.
.
.
Input

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Corpus
Index
IndexingInformation
Need
Query
Query
(Re)Formulation
Retrieval
Model
Ranked
Docs.
d1
d2
d3
d4
d5
.
.
.
Input
Information
Retrieval
System

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Corpus
Index
IndexingInformation
Need
Query
Query
(Re)Formulation
Retrieval
Model
Ranked
Docs.
d1
d2
d3
d4
d5
.
.
.
Input
Information
Retrieval
System
Output

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Retrieval Models
Language Model with Jelinek Mercer Smoothing (LM-JM)
Language Model with Dirichlet Smoothing (LM-Dir)
BM25
DFR
...

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Retrieval Models
BM25
DFR
...
Score(D,Q) =



∑
t∈Q f(D, t)
∏
t∈Q f(D, t)
...

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Retrieval Models
BM25
DFR
...
Score(D,Q) =



∑
t∈Q f(D, t)
∏
t∈Q f(D, t)
...
LM-JM(D,Q) =
∏
t∈Q
[
λ ∗ MLE(t in Document)
+ (1 − λ) ∗ MLE(t in Collection)
]

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Evaluation: Datasets
Collection Name Documents # documents # topics # rel-docs
TREC123 Tipster disks 1, 2 741,856
150
37836
(51-200)
TREC678
Tipster disks 4, 5
528,155
150
13692
exclude docs. from CR (301-450)
Robust
Tipster disks 4, 5
528,155
100
3720
exclude docs. from CR (601-700)
TREC910 WT10G 1,692,096
100
5980
(451-550)
GOV2 GOV2 25,205,179
150
26917
(701-850)
ClueWeb09B ClueWeb09 Disk1 - English 50,220,423
200
11037
(1-200)
Table: Overview of datasets used in experiments reported in this presentation.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Evaluation: Metrics
MAP: Mean average precision
P@5: Precision upto rank 5
Recall: Percentage of relevant documents retrieved

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Word Embedding

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Word Embedding
Represents every word as low dimensional vector in an abstract space.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Word Embedding
Similarity between vectors reflect semantic similarity between terms.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Word Embedding
Effect of conceptual composition by simple addition of vectors.
u
v
w1
w2

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Word Embedding
u
v
w1
w2
German + Airlines ∼ Lufthansa

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Word Embedding
Words have close vector representations if they share similar contexts.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Outline
1 Improving Baseline Retrieval Model
2 Improving Relevance Feedback based Query Expansion
3 Query Performance Prediction using Word Embedding

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Improving Baseline Retrieval Model
Outline

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Traditional Language Model
Given:
Collection C: A set of documents
Query Q
Documents ranked:
in decreasing order by the posterior probabilities P(D|Q)

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Traditional Language Model
P(D|Q) =
∏
t∈Q
λ
tf(t, D)
|D|
+ (1 − λ)
cf(t)
cs
Query terms generated by independent sampling from either
the document or the collection

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Language Model with Word Embedding
A generative process in which a noisy channel may transform a term
t′ into a term t

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Term Transformation Events : via Document Sampling
Figure: Schematics of generating a query term t from d
P(t, t′|d) = P(t|t′, d)P(t′|d)
P(t|t′, d) = sim(t′,t)
Σ(d)

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Term Transformation Events : via Collection Sampling
Figure: Schematics of generating a query term t from the collection
P(t, t′|C) = P(t|t′, C)P(t′|C)
P(t|t′, C) = sim(t,t′)∑
t′′∈Nt
sim(t,t′′)

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Term Transformation Events
Noisy Channel
d
C
t
λ
t′
1 − λ − α − β
α
β
Figure: A Generalized Language Model (GLM). GLM degenerates to LM when
α = β = 0.
λ: Standard Jelinek Mercer weighting parameter
α: Probability of sampling term t via a transformation through a term
t′ sampled from the document d
β: Probability of sampling term t via a transformation through a term
t′ sampled from collection

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Generalized Language Model : Evaluation
Query Method
Metrics
MAP P@5 Recall
TREC 123
LM 0.1967 0.4213 0.4477
GLM 0.2104†
0.4562†
0.4893†
TREC 678
LM 0.2189 0.4160 0.5300
GLM 0.2214†
0.4187 0.5327
Robust
LM 0.2658 0.4364 0.7881
GLM 0.2777†
0.4586†
0.7990
WT10G
LM 0.1747 0.3152 0.6377
GLM 0.1906†
0.3782†
0.6689†
Table: Comparative performance of GLM on the basis of mean average precision
(MAP) precision at 5 (P@5) and recall at 1000. A † indicates the significance of
the metric value with respect to the baseline LM based retrieval model.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Evaluation: Summary
GLM > LM

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Evaluation: Summary
GLM > LM
A Word Embedding based Generalized Language Model for
Information Retrieval (SIGIR 2015)

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Retrieval Models: Keyword Matching
Score(D,Q) =



∑
t∈Q f(D, t)
∏
t∈Q f(D, t)
...

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Retrieval Models: Keyword Matching
Score(D,Q) =



∑
t∈Q f(D, t)
∏
t∈Q f(D, t)
...
Query: Vehicle Smoke Pollution
Score(D,Q) = f(D, Vehicle) ∗ f(D, Smoke) ∗ f(D, Pollution)

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Vocabulary Mismatch
Vehicle
Smoke
Pollution
Doc1
Passenger vehicles & heavy-duty trucks
are a major source of air pollution
which includes ozone, particulate matter,
and other smog-forming emissions
in the form of smoke.
Doc2
The exhaust gas of cars powered by
fossil fuels are a major source of
toxic materials in the air that causes
severe damage to the eco-system.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Vocabulary Mismatch: Solution
Query Expansion
Add more words of similar meaning to initial query for better retrieval.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Outline

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Improving Relevance Feedback based Query Expansion
Outline

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Pseudo-Relevance Feedback based Query Expansion
Given a query Q
Perform initial retrieval using some retrieval model
Consider top K documents as (pseudo-)relevant
Select terms from those documents as candidate expansion terms
Perform re-retrieval with the expanded query

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Relevance based Language Model (RLM)
Assumes that both query and (pseudo-)relevant documents sampled
from a latent relevance model R.
The task → to find (estimate) the density function for R.
P(w|R) =
∑
D∈M
P(w|D)
∏
q∈Q
P(q|D)

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Motivation of this work
Word Vector Compositionality based Relevance Feedback
using Kernel Density Estimation

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Motivation of this work
using Kernel Density Estimation
Word embedding:
Captures the semantic relationship between terms.
Relevance feedback:
Captures co-occurence information of terms with query in top docs.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Kernel Density Estimation (KDE)
76543210-1
5
4
3
2
1
0
76543210-1
5
4
3
2
1
0
76543210-1
5
4
3
2
1
0
Estimate a distribution that generates the given data (data points).
Place kernel function (e.g. Gaussian) centered around each observed
data point.
Combine the Gaussians to get a function peaked at the observed data
point.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Kernel Density Estimation (KDE)
f(x) =
1
nh
n∑
i=1
K(
x − xi
h
)
h - bandwidth (smoothing parameter)
n - number of data points
xi - ith data point
K() - the kernel

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
KDE with Gaussian Kernel
f(x) =
1
nh
n∑
i=1
K(
x − xi
h
)

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
f(x) =
1
nh
n∑
i=1
K(
x − xi
h
)
f(x) =
1
nh
n∑
i=1
1
√
2πσ
e− 1
2σ2 (
x−xi
h
)
2

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
f(x) =
1
nh
n∑
i=1
K(
x − xi
h
)
f(x) =
1
nh
n∑
i=1
1
√
2πσ
e− 1
2σ2 (
x−xi
h
)
2
h and σ: tunable parameters

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Comparing the Estimations
Relevance Model Estimation
Given set of terms from (pseudo-)relevant documents, estimate the density
function that generates R.
76543210-1
5
4
3
2
1
0
Kernel Density Estimation
Given set of n data samples, estimates the density function that generates
the observed data.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
KDE in Relevance Feedback Scenario
f(x) =
1
nh
n∑
i=1
K(
x − xi
h
)
The Points
Observed data points (xi): Q = {q1, . . . , qn}
Points at which probability (density) is to be estimated (x):
w → Candidate expansion term

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
KDE in Relevance Feedback Scenario
f(x) =
1
nh
n∑
i=1
K(
x − xi
h
)
The Points
f(w) =
1
nh
n∑
i=1
K(
w − qi
h
)

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Relevance Feedback with KDE
f(w1)
f(q2)
f(w2)
q1 q2 q3
w1 w2
One dimensional projection of embedded vectors

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Kernel Density Estimate (Unweighted)
f(w) =
1
nh
n∑
i=1
K(
w − qi
h
)
f(w1)
f(q2)
f(w2)
q1 q2 q3
w1 w2
n - number of samples
h - bandwidth
K(·) - kernel function

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Kernel Density Estimate (Weighted)
f(w, α) =
1
nh
n∑
i=1
αiK(
w − qi
h
)
f(w2)
f(q2)
f(w1)
q1 q2 q3
w1 w2
n - number of samples
h - bandwidth
K(·) - kernel function
αi - weight of local
Kernel function around
ith data point

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
One dimensional KDE
f(w2)
f(q2)
f(w1)
q1 q2 q3
w1 w2
f(w, α) =
1
nh
n∑
i=1
αi
1
√
2πσ
e−
(w−qi)2
2σ2h2
αi : weight of the ith query term

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
One dimensional KDE
f(w, α) =
n∑
i=1
P(w|M)P(qi|M)
1
σ
√
2π
exp(−
(w − qi)T(w − qi)
2σ2h2
)
Captures co-occurence information of terms with query in top
docs.
Captures the semantic relationship between terms.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Two dimensional KDE
In 1D-KDE, set of all top ranked documents considered as a single
document model (M).
We now generalize the 1D-KDE to two dimensions so as to weight
the contributions obtained from different documents in the ranked list
separately.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Two dimensional KDE
q1 q2 q3 Terms
Documents
D1
D2
2
1
D1
D3
The x-axis corresponds to the query term vectors similar to one
dimensional KDE (unordered).
The y-axis represents the normalized term frequency of query terms in
respective feedback documents.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Two dimensional KDE
The Points: One Dimensional KDE

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Two dimensional KDE
The Points: Two Dimensional KDE
Observed data points (xij) = (qi, Dj): Encapsulates word vector for
i-th query word qi and its normalized term frequency in feedback
document Dj (P(qi|Dj)).
Points at which probability (density) is to be estimated (x)
= (w, Dj): Encapsulates word vector for candidate expansion term w
and its normalized term frequency in feedback document Dj
(P(w|Dj)).

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Two dimensional KDE
f(x, α) =
n∑
i=1
M∑
j=1
P(w|Dj)P(qi|Dj)
2πσ2
exp(
(w − qi)2 + (P(w|Dj) − P(qi|Dj))2
−2σ2h2
)

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Two dimensional KDE
f(x, α) =
n∑
i=1
M∑
j=1
P(w|Dj)P(qi|Dj)
2πσ2
exp(
(w − qi)2 + (P(w|Dj) − P(qi|Dj))2
−2σ2h2
)
w in document Dj, denoted as x, will get a high value of the density
function if:
(i) w is semanticlly close to query terms qi;
(ii) w frequently co-occurs with query terms in each top ranked document
Dj.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Term Composition in KDE
Extending the set of pivot points:
q1 q2q1+q2
w
(a) Without composition
q1 q2q1+q2
w
(b) With composition

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Importance of Composition
ISI Inter Services
Intelligence
Infrared
Spatial In-
terferometer
Indian
Statistical
Institute
Institute for
Scientific
Interchange
Islamic
State of Iraq
Integral
Satcom
Initiative
Ice Skating
Institute
Islamic
School
of Irving

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Importance of Composition
ISI
Kolkata
Inter Services
Intelligence
Infrared
Spatial In-
terferometer
Indian
Statistical
Institute
Institute for
Scientific
Interchange
Islamic
State of Iraq
Integral
Satcom
Initiative
Ice Skating
Institute
Islamic
School
of Irving

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Summary
Given:
Corpora : Set of documents.
Query Q : q1, . . . , qn.
Vector embeddings of each of the terms.
First level retrieval performed using any LM.
ET = All terms of top ranked M documents considered as potential
expansion terms.
Calculate KDE(w) for each w ∈ ET.
Terms with high estimated value taken as expansion terms.
Linearly interpolate derived density function with underlying query
model.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Evaluation
Query Method
LM-JM
MAP P@5 Recall
TREC 123
LM 0.1967 0.4213 0.4477
RM3 0.2507∗ 0.4653∗ 0.5154∗
KDERLM 0.2661∗† 0.4798∗ 0.5373∗†
TREC 678
LM 0.2189 0.4160 0.5300
RM3 0.2351∗ 0.4547∗ 0.5366
KDERLM 0.2536∗† 0.4601∗† 0.5520∗†
Robust
LM 0.2658 0.4364 0.7881
RM3 0.3309∗ 0.4929∗ 0.8596∗
KDERLM 0.3420∗† 0.5043∗ 0.8596∗
WT10G
LM 0.1747 0.3152 0.6377
RM3 0.2094∗ 0.3394∗ 0.6743∗
KDERLM 0.2221∗† 0.3419∗ 0.6910∗†

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Evaluation: Summary
KDERLM > RM3 > GLM > LM

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Evaluation: Summary
KDERLM > RM3 > GLM > LM
using Kernel Density Estimation (CIKM 2016)

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Query Expansion for Simple Query
Searching data set
National Geographic Readers: Snakes!

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Searching data set
Query
python life span

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Searching data set
Query
python life span
python setup in windows 10

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Searching data set
Query
python life span
Query expansion → overkill or, not work at all.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Searching data set
Query
python life span
Query expansion → overkill or, not work at all.
How to understand whether query actually need expansion? → QPP

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Query Performance Prediction using Word Embedding
Outline

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Query Performance Prediction (QPP)
Definition
To quantify the quality of search results when no relevance feedback is
given.
For an ambiguous query, it may be difficult for a search engine to
return satisfactory results for the query.
This can lead to poor IR effectiveness.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Query Performance Prediction (QPP)
Why we need QP Predictor?
▶ Feedback to Users.
▶ Feedback to Search Engine.
▶ Feedback to System Administrator.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
What Makes a Query Difficult?
Figure: A general model of a topic based on a query Q expressing a specific
information need, the relevant documents R for Q, the entire collection C, and
the distances between the sets involved.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Characteristics of Difficult Query
d(Q, C) - Distance between query and collection; Reflects likelihood of generation of Q
from C.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
from C.
d(R, C) - Distance between retrieved documents and collection; Reflects likelihood of
generation of documents of R from C.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
from C.
d(Q, Q) - Diversity of query. Reflects the ambiguity of Q; more the diameter, more
ambiguous the query.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
from C.
d(R, R) - Distance among retrieved documents; Manifestation of ambiguity of retrieved
documents.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
from C.
d(R, R) - Distance among retrieved documents; Manifestation of ambiguity of retrieved
documents.
d(Q, R) - Distance between query and retrieved documents R; Equivalent to distribution
of retrieval scores of R.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
from C.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Word Embedding based QPP
Ambiguous Terms
Terms with multiple senses
Hypothesis
For an ambiguous term w, in terms of the embedded space, w is more
likely to be a peripheral point of a word cluster rather than being an
interior point close to the cluster centre.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Ambiguous Term in Embedded Space
Figure: The neighbourhood of an ambiguous word with multiple senses

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Figure: Histogram of number of terms in the ϵ-neighbourhood of a term

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Figure: Histogram of number of terms in the ϵ-neighbourhood of a term
Cluster the set of points in the ϵ-neighbourhood of query term
For an ambiguous word, the distribution of points in the clusters likely
to be close to uniform

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
AMBWV(Q; K, ϵ) =
1
K
K∑
k=1
(|Nϵ(Q)| − µ)2
;
Q - The query.
Nϵ(Q) - ϵ-neighborhood of Q.
K - Number of clusters.
ϵ - Neighbourhood size.
µ - Average cardinality of clusters. (|Nϵ(Q)|
K )

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
AMBWV(Q; K, ϵ) =
1
K
K∑
k=1
(|Nϵ(Q)| − µ)2
;
High variance → One sense predominates others → Non-ambiguous
Low variance → Having multiple senses → Ambiguous

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Pre-retrieval QPP
d(Q, C) - Reflects likelihood of generation of Q from C.
d(Q, Q) - Reflects the ambiguity of Q;

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Pre-retrieval QPP
d(Q, C) - Reflects likelihood of generation of Q from C.
d(Q, Q) - Reflects the ambiguity of Q;
Pre-retrieval QPP Methods
Ignores any post-retrieval information
Hybrid of pre-retrieval and post-retrieval approaches outperforms both

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Combining with NQC
NQC
Post-retrieval QPP.
Variance of similarity score proportional to query performance.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Word Embedding and NQC based Hybrid Predictor
AMBWV-NQC = αAMBWV + (1 − α)NQC

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Evaluation: Metrics
Correlation between QPP scores and true score (AP value)
Evaluation Metrics
Pearson’s ρ: Correlation coefficient
Kendall’s τ: Rank correlation coefficient
High correlation → Better prediction

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Results on Some of the Topic sets
TREC3 TREC7 TREC8 TREC10
ρ τ ρ τ ρ τ ρ τ
AvgIDF 0.2353 0.2490 0.4262 0.3796 0.5910 0.3518 0.3736 0.2359
MaxIDF 0.2285 0.2772 0.3524 0.2662 0.4938 0.2996 0.2219 0.1233
WordNet 0.0297 -0.0375 0.2696 0.2509 0.0746 0.1014 0.2221 0.0896
AMBWV 0.3222 0.2457 0.2043 0.2132 0.2132 0.1559 0.1544 0.0033
NQC 0.3146 0.1559 0.4580 0.3502 0.6717 0.4335 0.4676 0.2686
AvgIDF-NQC 0.2782 0.2620 0.4551 0.4253 0.6315 0.3878 0.4119 0.2588
MaxIDF-NQC 0.2862 0.3159 0.4204 0.3094 0.5815 0.3714 0.2157 0.2336
WordNet-NQC 0.1794 0.1069 0.4095 0.3453 0.3457 0.3078 0.3292 0.1673
AMBWV-NQC 0.4741 0.3648 0.5362 0.3844 0.7070 0.4498 0.4936 0.2914
Table: Comaprisons of the word embedding based QPP method against various
baselines on the test topic sets.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Contribution
An embedding based approach to quantify underlying ambiguity of
user query.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Contribution
An embedding based approach to quantify underlying ambiguity of
user query.
Estimating Gaussian Mixture Models in the Local
Neighbourhood of Embedded Word Vectors for Query
Performance Prediction
(Information Processing and Management - in press)

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Conclusion

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Conclusion
WE & I R Great together!

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Conclusion
WE & I R Great together!
https://github.com/dwaipayanroy
dwaipayan_r@isical.ac.in

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
THANK YOU!

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
GLM: A less-heavy QE method

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
GLM Parameter Details
λ empirically set to 0.2
GLM parameters α and β varied within range of [0.1, 0.4] to ensure
α + β + λ < 1
Word vectors embedded in a 200-dimensional space with
negative-sampling using 5-word window on continuous bag-of- words
model.
Nt set to 3
Back

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
1D-KDE Details
P(w|M) : tf(w,M)
|M|
P(qi|M) : tf(qi,M)
|M|
h: Bandwidth set to one (1)
σ: Trained on the development set (TREC 6 and TREC 9)
Return

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
2D-KDE Details
(w − qi)2
= (w − qi)T(w − qi)
P(w|Dj) :
tf(w,Dj)
|Dj|
P(qi|Dj) :
tf(qi,Dj)
|Dj|
h: Bandwidth set to one (1)
σ: Trained on the development set (TREC 6 and TREC 9)
covariance matrix σ as a diagonal matrix with equal covariance for
both the dimensions.
Return

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Extra slide
Return

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Pearson’s ρ
ρ(X,Y) =
∑n
i=1 (xi − ¯x)(yi − ¯y)
√∑n
i=1 (xi − ¯x)2
√∑n
i=1 (yi − ¯y)2
Range: (−1, 1)
Back

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Kendall’s τ
τ(X,Y) =
(No. of concordant pair) - (No. of discordant pair)
n(n − 1)/2
Range: (−1, 1)
Back

Improving Information Retrieval Performance using Word Embedding

More Related Content

Similar to Improving Information Retrieval Performance using Word Embedding (20)

Recently uploaded (20)

Improving Information Retrieval Performance using Word Embedding