IMPROVING SEARCH ENGINES BY DEMO

International Journal on Natural Language Computing (IJNLC) Vol.8, No.4, August 2019
DOI: 10.5121/ijnlc.2019.8401 1
IMPROVING SEARCH ENGINES BY DEMOTING
NON-RELEVANT DOCUMENTS
Fadi Yamout and Mireille Makary
The International University of Beirut, Department of Computer Science, Beirut
ABSTRACT
A good search engine aims to have more relevant documents on the top of the list. This paper describes a
new technique called “Improving search engines by demoting non-relevant documents” (DNR) that
improves the precision by detecting and demoting non-relevant documents. DNR generates a new set of
queries that are composed of the terms of the original query combined in different ways. The documents
retrieved from those new queries are evaluated using a heuristic algorithm to detect the non-relevant ones.
These non-relevant documents are moved down the list which will consequently improve the precision. The
new technique is tested on WT2g test collection. The testing of the new technique is done using variant
retrieval models, which are the vector model based on the TFIDF weighing measure, the probabilistic
models based on the BM25, and DFR-BM25 weighing measures. The recall and precision ratios are used
to compare the performance of the new technique against the performance of the original query.
KEYWORDS
Information retrieval, TFIDF, BM25, DFR-BM25.
1. INTRODUCTION
Search engines extract user-specified information from documents and files, ranging from books
to online blogs, journals, and academic articles [1]. The primary objective of search engines is to
quickly and precisely retrieving relevant documents related to the user‘s request [2]. Search
engines cannot be 100% accurate because the document relevance is subjective and depends on
the user's judgment, which depends on many factors such as his knowledge about the topic, the
reason for searching, and his satisfaction with the returned result [3].There are many challenges
involved in making a search engine successful [2,4]. These challenges include acquiring lots of
relevant documents from many sources, extracting useful representations of the documents to
facilitate search, ranking documents in response to a user request, and presenting the search
results effectively by posting the most relevant document on the top of the list [5,6,7].
This paper describes a new technique to improve search engines performance. The new technique
locates non-relevant documents among the documents retrieved and moves them down the list.
As a result, more relevant documents are lifted up the list, and consequently, the performance of
the search engine [8] improves. This is done by generating new queries from the original query,
retrieve a set of document for each new generated query, combine them into one set and use a
heuristic to determine the most non-relevant documents. The new technique is tested on WT2g1
test collection using the vector model [9,10,11,12] based on the TFIDF weighing measure[13,14],
the probabilistic models [15] based on the BM25, and DFR-BM25 weighing measures[16,17,18].
The recall and precision ratios are used to compare the performance of the new technique against
the performance of the original query.
1
http://ir.dcs.gla.ac.uk/wiki/Terrier/WT2G

2
2. MODELS
The Information Retrieval (IR) model defines a way to represent the documents and queries to
compare them [19, 20]. The most common IR models are vector [9, 10] and probabilistic
models[15].
2.1. Vector Model
In a vector model, both the documents and the queries are represented as vectors in multi-
dimensional space, where the terms become the dimensions’ vector [9, 10, 11, 12]. Therefore,
document d is represented as a vector of terms, as shown in Equation [11]:
d = (t1, t2 … tm)
(1)
In this equation, m represents the number of unique terms in the collection and ti denotes the
presence or absence of term i in document d. Vector model is based on linear algebra allowing
documents to be ranked based on their possible relevance to the query [11].
2.2. Probabilistic
In a probabilistic model, the documents and queries are viewed as vectors. However, the weight
of a term in a document is given by a probability [15]. Probabilistic models have been extended to
different models; Best Match 25 (BM25) [16, 17, 18] Okapi [21], Statistical Language Modelling
(SLM) [22], and Divergence From Randomness (DFR-BM25) [17].
3. WEIGHTING TERMS
Information retrieval system has various methods for weighting terms [23, 24, 25]. The primary
weights are TFIDF [14, 13], BM25 [18] and DFR_BM25 [16, 17, 18], weighting measure.
Assigning a weight for each term in each document in the collection has shown great potential for
improving retrieval effectiveness [24].
3.1. TFIDF
TFIDF weighting measure is a combination of local and global weights [13, 14]. The term
frequency (TFij) is based on the notion that terms that frequently occur in the text of documents
are essential in that text. Therefore, it represents the occurrences of a term i in a document j. The
global weight is the document frequency (DocFreqi), which represents in how many documents
the term i occurs. Inverse Document Frequency (IDFi) of term i has DocFreqi scaled to the total
number of documents in the collection (N) as shown in Equation 2:
𝐼𝐷𝐹𝑖 = log10 (
N
DocFreqi
)
(2)
In equation (2), N is the total number of documents, DocFreqi is the total number of documents
containing term i, and the log is based 10. The logarithm reduces the large value obtained due to
N since IDFi is to be multiplied by the small value TFij to derive the TFIDF as shown in Equation
3:
𝑇𝐹𝐼𝐷𝐹 = 𝑇𝐹𝑖𝑗 × 𝐼𝐷𝐹𝑖 (3)

3
3.2. BM25
BM25 [18] also known as “Best Match 25”, is the main weighting measures for probabilistic
models. BM improved from the traditional probabilistic weighting scheme to BM25 through
BM11 and BM15. It is the best of the known probabilistic weighting schemes by recent TREC
tests.
The weights assigned to the documents’ terms are given by a probability shown in Equation 4:
𝐵𝑀25 =
TFij × (k3 + 1) × QFiq
((k3 + QFiq) × K)
× log2 (
N − DocFreqi + 0.5
DocFreqi + 0.5
) (4)
Where QFiq represents the occurrences of a term i in a query q, k3 is set to 1000, as proposed in
[18], and K is shown in Equation 5:
K = k1 × ((1 − b) + b ×
DocLj
averageDocL
) + TFij (5)
Where DocLj is the length of document j, averageDocL is the average length of all documents in
the test collection, k1 is set to 1.2, and b is set to 0.75 as proposed in [18].
3.3. DFR-BM25
DFR_BM25 [16, 17], is derived by measuring the divergence of the actual term distribution from
that obtained under a random process. The weight of the term in a document is computed as a
function of two probabilities Prob1 and Prob2. Equation 6 shows the weight of a term as a product
of two components.
𝑤 = (1 − 𝑃𝑟𝑜𝑏2) × (−log2(𝑃𝑟𝑜𝑏1)) (6)
Prob2 measures the information gain of the term concerning the set of all documents in which the
term occurs. It is measured by the counter-probability (1- Prob2), where the less the term is
expected in a document concerning its frequency in the set of all documents in which the term
occurs, the more the amount of information is gained with this term. The counter-probability 1 –
Prob2 is computed, as shown in Equation 7:
(1 − Prob2) = (
TFij + 1
DocFreqi × (TFi + 1)
) (7)
Prob1 measures the information content of the term in a document. The component (–log2 (Prob1))
provides the equivalent amount of information and is computed, as shown in Equation 8:
(− log2 𝑃𝑟𝑜𝑏1) = TFi × log2 (
TFi
𝜆
) + (𝜆 +
1
12 × TFi
− TFi) × log2 ℯ + 0.5 × log2(2 × 𝜋 × TFi) (8)
Where TFi is the term frequency of the term i in the collection, and λ= TFi/N.
4. TEST COLLECTION
A test collection consists of a large collection of documents, a set of queries, and a relevance
judgment list which matches each query to its relevant documents [7, 26, 27]. In this paper,
WT2g2
test collection is used for the experiments. It has a size of 2 GB and consists of 247491
2
http://ir.dcs.gla.ac.uk/wiki/Terrier/WT2G

4
documents. WT2g has 50 topics with a variant number of terms in each query. In our technique,
we use the queries that are composed of three terms, as shown in Table 1, since we can generate
more queries out of three terms.
Table 1. Queries of WT2g Composed of Three Terms
Query
number
Query content Query number Query content
401 foreign minorities, Germany 423 Milosevic, Mirjana Markovic
404 Ireland, peace talks 426 law enforcement, dogs
407 poaching, wildlife preserves 427 UV damage, eyes
409 legal, Pan Am, 103 428 declining birth rates
411 salvaging, shipwreck, treasure 430 killer bee attacks
414 Cuba, sugar, exports 432 profiling, motorists, police
415 drugs, Golden Triangle 433 Greek, philosophy, stoicism
416 Three Gorges Project 435 curbing population growth
419 recycle, automobile tires 437 deregulation, gas, electric
420 carbon monoxide poisoning 439 inventions, scientific discoveries
421 industrial waste disposal 443 U.S., investment, Africa
422 art, stolen, forged 450 King Hussein, peace
Query 428, for example, has the following three terms: “declining”, “birth”, and “rates”. WT2g
also has a Relevance Judgment List (RJL) that indicates the relevant documents for each query [7,
27]. For example, RJL indicates that query 401 “foreign minorities, Germany“, has 2739 relevant
documents.
An example of a few relevant documents to query 401 is shown in Table 2.
Table 2. A Few Relevant Documents to Query 401
Relevant documents
FBIS3-100090, FBIS3-106290, FBIS3-115210,
FBIS3-100590, FBIS3-107660, FBIS3-127970,
FBIS3-131970, FBIS3-148320, FBIS3-153870,
FBIS3-133220, FBIS3-149260, FBIS3-155350
When a query is submitted, many documents are retrieved with the most relevant ones on the top
of the list. Table 3 shows a few documents retrieved by query 401. The two notations “R” and
“NR” are used to indicate a relevant document and non-relevant one respectively. Document
FBIS4-18372, for example was not found in the RJL whereas FBIS3-20090 was found in the
RJL.
Table 3. A Few Documents Retrieved by Query 401
Rank Documents RJL Rank Documents RJL
1 FBIS4-18372 NR 4 LA022790-0091 NR
2 FBIS3-20090 R 5 FT922-14939 R
3 FT941-1403 NR
5. ASSESSMENT
The evaluation measures used to assess the effectiveness of Information Retrieval (IR) are
precision and recall [5].

5
The precision is the number of relevant documents retrieved over the retrieved documents, as
shown in Equation 9:
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 =
𝑅𝑒𝑙𝑒𝑣𝑎𝑛𝑡 𝑑𝑜𝑐𝑢𝑚𝑒𝑛𝑡𝑠 𝑅𝑒𝑡𝑟𝑖𝑒𝑣𝑒𝑑
𝑅𝑒𝑡𝑟𝑖𝑒𝑣𝑒𝑑 𝑑𝑜𝑐𝑢𝑚𝑒𝑛𝑡𝑠
(9)
The recall is the number of relevant documents retrieved over the total number of relevant
documents, as shown in Equation 10:
𝑅𝑒𝑐𝑎𝑙𝑙 =
𝑅𝑒𝑙𝑒𝑣𝑎𝑛𝑡 𝑑𝑜𝑐𝑢𝑚𝑒𝑛𝑡𝑠 𝑅𝑒𝑡𝑟𝑖𝑒𝑣𝑒𝑑
𝑅𝑒𝑙𝑒𝑣𝑎𝑛𝑡 𝑑𝑜𝑐𝑢𝑚𝑒𝑛𝑡𝑠
(10)
The precision in this paper is represented using the precision-recall curve with pre-established
recall levels (0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, and 1.0) where all queries are put
together and averaged at these levels.
6. THE NEW TECHNIQUE: DNR
In this section, an example is used to describe the new technique proposed in this paper,
“improving search by demoting non-relevant documents (DNR). The example has the sample
query 555 that contains three terms “improved”, “search”, and “engines”. Table 4 lists the
documents retrieved by query 555.
Table 4. List of Documents Retrieved by Query 555
Rank Document RJL Rank Document RJL Rank Document RJL Rank Document RJL
1 1872 N 4 0091 R 7 1796 N 10 1883 R
2 2090 R 5 1439 N 8 1882 R
3 1403 N 6 5536 N 9 6528 N
3 1403 N 6 5536 N 9 6528 N
Note : R: Relevant – N: Non-relevant
The documents 2090, 0091, 1882, and 1883 are considered relevant with the notation ’R’ in
column RJL whereas non-relevant documents have the notation ‘N’.
The first step is to generate queries from the terms of the initial query 555. The first set of queries
generated contains only one term, and are called them query-size-one. Table 5 shows three query-
size-ones generated: The first query generated is query 5551 and contains the term “improved”.
The second query, 5552, contains the term “search”. The third query, 5553, contains the term
“engines”.
Table 5. Query-Size-One Generated from Query 555
Query generated Content
Query 5551 improved
Query 5552 search
Query 5553 engines
The second set of queries generated contains two terms and are generated by combing the terms
in all possible ways. These set of queries are called query-size-two. Table 6 shows the three
query-size-two generated. Query 5554 contains the terms “improved”, and “search”, query 5555
contains the terms “improved”, and “engines”, query 5556 contains the terms “search”, and
“engines”.

6
Table 6. Query-Size-One Generated from Query 555
Query generated Content
Query 5554 improved, search
Query 5555 improved, engines
Query 5556 search, engines
Each of the generated queries will retrieve its unique list of documents; a few or a lot of those
retrieved documents are also retrieved by the original query 555. Table 7 compares the retrieved
documents by generating queries against the retrieved documents by the original query.
Table 7. Comparison between the Original Query and Generated Queries
Rank
Original Query 555
RJL
Query-size one Query-size two
Documents 5551 5552 5553 5554 5555 5556
1 1872 N 0 1 1 1 0 0
2 2090 R 1 1 1 1 1 1
3 1403 N 0 0 0 1 0 0
4 0091 R 0 1 1 1 1 1
5 1439 N 0 0 0 1 0 1
6 5536 N 0 0 0 0 0 1
7 1796 N 0 1 1 1 1 1
8 1882 R 0 0 0 0 0 1
9 6528 N 0 0 0 1 1 1
10 1883 R 0 1 1 1 1 1
Note: R: Relevant – N: Non-relevant
For example query-size-one 5552 retrieved the document “1272”, as indicated by the digit ‘1’,
which is also retrieved by the original query 555 as indicated by the notation ‘R’. Next, the list of
retrieved documents by the query-size-one and query-size-two generated queries are examined. In
Table 7, some of the documents were not retrieved by the query-size-one queries such as “1439”,
and “6528”. A few other documents were retrieved by query-size-one and query-size-two queries
such as “1872”, “2090”, “0091”, 1796”, and “1883”. DNR considers a document to be non-
relevant when it is retrieved by none of the query-size-one generated queries and at most by one
of the query-size-two queries. This condition detects the non-relevant documents as illustrated in
Table 8.
Table 8. Selection of the Non-Relevant Documents
Rank
Original Query 555 Query-size one Query-size two
Status
Documents
RJL
5551
5552
5553
S1
5554
5555
5556
S2
1 1872 N 0 1 1 1 0 0 T
2 2090 R 1 1 1 1 1 1
3 1403 N 0 0 0 T 1 0 0 T S
4 0091 R 0 1 1 1 1 1
5 1439 N 0 0 0 T 1 0 1
6 5536 N 0 0 0 T 0 0 1 T S
7 1796 N 0 1 1 1 1 1
8 1882 R 0 0 0 T 0 0 1 T S
9 6528 N 0 0 0 T 1 1 1
10 1883 R 0 1 1 1 1 1

7
In Table 8, the column S1 has a “true” value when a document is retrieved by at most one query-
size-one query. Column S2 has a “true” value when at most two query-size-two queries retrieve a
document. Finally, a document is considered non-relevant when S1 and S2 have a “true” value,
and this is indicated by “S” in the “Status” column. Therefore, three out of the ten documents are
selected as non-relevant documents. The documents are “1403”, “5536”, and “1882”. These
documents are labeled with “S” in column “Status”. The results are classified as “False Alarm”,
“Relevant Rejected”, “Missed”, and “Not-relevant Rejected” as shown in Table 9.
Table 9. Summary of the Results
Selected Not selected
R (Relevant) False Alarm Relevant Rejected
NR (Non-relevant) Non-relevant Selected Missed
For example, the non-relevant documents “1403” and “5536” in Table 8 are classified as “Non-
relevant Selected” since our technique selected them; the third document, “1882”, is classified as
“False Alarm” since it is relevant in the RJL and was classified by our technique as non-relevant.
The non-relevant documents “1872”, “1439”, “1796”, and “6528” are classified as “Missed”
since they are classified as non-relevant in the RJL but were not selected as non-relevant by our
technique. The relevant documents, “2090”, “0091”, and “1883”, are classified as “Relevant
Rejected” since our technique did not select them.
Table 10 lists the classifications of the documents
Table 10. Classifications of the Documents
Rank Original query 555 RJL Status Classification
1 1872 N Missed
2 2090 R Relevant Rejected
3 1403 N S Non-relevant Selected
5 1439 N Missed
6 5536 N S Non-relevant Selected
7 1796 N Missed
8 1882 R S False Alarm
9 6528 N Missed
The best categories are the “Non-relevant Selected” and “Relevant Rejected”; the “Non-relevant
Selected” detects the non-relevant documents based on our condition whereas the “Relevant
Rejected” detects the relevant documents that should not be selected. The “False Alarm”
documents affect the precision of the retrieved documents severely since relevant documents in
RJL were considered to be non-relevant by our technique. Although the “Missed” category
missed the non-relevant documents, however, it has no significant effect on precision as will be
shown in the experiments. Finally, the non-relevant documents are moved down the list to
improve the precision of the retrieved documents, as shown in Table 11.

8
Table 11. Relevant Documents Moved Higher
Rank Initial results RJL Status Rank New results RJL Status
1 1872 N 1 1872 N
2 2090 R 2 2090 R
3 1403 N S 3 0091 R
4 0091 R 4 1439 N
5 1439 N 5 1796 N
6 5536 N S 6 6528 N
7 1796 N 7 1883 R
8 1882 R S 8 1403 N S
9 6528 N 9 5536 N S
10 1883 R 10 1882 R S
Note: S: Selected - R: Relevant – N: Not-relevant
The new technique moves a few of the selected non-relevant documents to the bottom of the list
to improve the precision. For example, documents “1403“, “5536”, and “1882” are moved to the
bottom of the list as shown in Table 10. Consequently, documents “0091” and “1883”, which are
relevant (R), are lifted. As a result, more relevant documents are moved to the top.
7. EXPERIMENTS AND RESULTS
The following sections describe the experimental results of “Improving search engines by
demoting non-relevant documents” (DNR) against the baseline in the vector and probabilistic
models. Although the experiments were done on all WT2g’s documents, this paper shows the
results of the top twenty documents on one of the queries.
7.1 Using the vector model based on TFIDF
When DNR is tested in the vector model based on TFIDF weighting measure [14, 13], 3781
documents were found non-relevant in the RJL. These documents were pushed down the list and
consequently precision improved. The technique also classified 116 relevant documents as non-
relevant. Also, 506 documents, that are relevant in the RJL, were detected to be as relevant and
therefore, were not selected. Finally, 15419 documents that are non-relevant in the RJL were
missed. Table 12 summarizes the results, and Table 13 shows only the top twenty documents
retrieved for query 451.
Table 12. Statistics on the Vector Model, based On TFIDF
R (Relevant) False Alarm : 116 Relevant Rejected : 506
NR (Non-relevant) Non-Relevant Selected: 3781 Missed : 15419

9
Table 13. Selection of the Non-Relevant Documents for Query 451
Rank
Status
Classified
Documents
RJL
4511
4512
4513
S1
4514
4515
4516
S2
100 FT932-4802 R 0 0 1 0 1 1 RR
101 FT933-496 R 0 0 1 0 1 1 RR
102 FT924-13548 N 0 0 1 0 1 1 M
103 LA090190-0081 N 1 0 0 1 1 0 M
104 LA092990-0090 N 0 0 1 0 1 1 M
105 LA102190-0152 N 0 0 1 0 1 1 M
106 FT921-16122 R 0 0 0 T 1 0 0 T S FA
107 FT943-11390 R 0 0 1 0 1 1 RR
108 FT911-4070 R 0 0 1 0 1 1 RR
109 FT943-12373 N 0 0 1 0 1 1 M
110 FT944-6889 N 0 0 0 T 1 0 0 T S NS
111 LA090690-0256 N 0 0 1 0 1 1 M
112 LA051989-0015 R 0 0 1 0 1 1 RR
113 LA092290-0079 N 0 0 1 0 1 1 M
114 LA111290-0065 N 0 0 1 0 1 1 M
115 LA082490-0156 N 0 1 0 1 0 1 M
116 FT941-15071 N 0 0 1 0 1 1 M
117 FT942-4603 R 0 0 1 0 1 1 RR
118 LA053190-0100 R 0 0 1 0 1 1 RR
119 LA042289-0128 N 1 0 1 1 1 0 M
Note: FA: False Alarm – M: Missed – RR: Relevant Rejected NS: Non-relevant Selected –
T: True – S: Selected – R: Relevant – N: Not-relevant
In Table 13, two non-relevant documents “FT921-16122”, and “FT944-6889” were detected by
our technique to be non-relevant and therefore were shifted down the list. Consequently, the
relevant documents “FT943-11390”, “FT911-4070”, “LA051989-0015”, “FT942-4603” and
“LA053190-0100” will be lifted two ranks to be positioned higher in the list. Therefore, precision
will improve. It should be noted here that the relevant document “FT921-16122” is classified as a
non-relevant and is a “False Alarm”.
7.2 Using the probabilistic model based on BM25
When DNR is tested in the probabilistic model based on BM25 weighting measure [18] it
classified 3631 non-relevant documents as non-relevant. These documents were pushed down the
list and consequently precision improved. The technique also classified 97 relevant documents as
non-relevant. Also, 526 relevant documents were classified to be as relevant and therefore, were
not selected.
Finally, 15568 documents that are non-relevant in the RJL were missed as shown in Table 14, and
Table 15 shows only the top twenty documents retrieved for query 451.
Table 14. Statistics on the Probabilistic Model, based On BM25
NR (Non-relevant) Non-Relevant Selected 3631 Missed : 15568

10
Table 15. Selection of the Non-Relevant Documents for Query 451
Rank
Status
Classified
Documents
RJL
4511
4512
4513
S1
4514
4515
4516
S2
250 LA121590-0052 R 0 1 0 1 0 1 RR
251 LA110690-0017 N 0 0 0 T 1 0 0 T S NS
252 LA092290-0008 N 0 0 1 0 1 1 M
253 FT942-11348 N 0 1 0 1 0 0 T M
254 LA082290-0085 N 0 0 0 T 1 0 0 T S NS
255 LA120290-0218 N 0 0 0 T 1 0 0 T S NS
256 FT934-10273 N 0 1 0 1 0 1 M
257 LA040490-0001 R 0 1 0 1 0 1 RR
258 LA090290-0168 N 0 1 0 1 0 1 M
259 LA111990-0044 N 0 1 0 1 0 1 M
260 FT911-4579 N 0 1 0 1 0 1 M
261 LA110190-0016 N 0 0 1 0 1 1 M
262 LA100690-0121 N 0 1 0 1 0 1 M
263 LA120690-0183 N 0 0 1 0 1 1 M
264 FT943-16550 N 0 1 0 1 0 1 M
265 LA120190-0115 N 0 1 0 1 0 1 M
266 FT934-10071 N 0 0 1 0 1 1 M
267 LA092190-0053 R 0 1 0 1 0 1 RR
268 LA101690-0044 N 0 1 0 1 0 1 M
269 FT923-12577 N 0 1 0 1 0 1 M
Three documents, “LA110690-0017”, “LA082290-0085”, and “LA120290-0218” were detected
by our technique to be non-relevant and therefore, are shifted down the list. Consequently, the
relevant documents “LA040490-0001” and “LA092190-0053” are lifted three ranks to be
positioned higher in the list.
7.3 Using the probabilistic model based on DFR_BM251
When DNR is tested in the probabilistic model based on DFR_BM25 weighting measure [16, 17,
18] DNR classified 3632 non-relevant documents as non-relevant. These documents were pushed
down the list and consequently precision improved. The technique also classified 93 relevant
documents as non-relevant. Also, 533 documents, that are relevant in the RJL, were detected to be
as relevant and therefore, were not selected. Finally, 15564 documents that are non-relevant in the
RJL were missed as shown in Table 16, and Table 17 shows only the top 20 documents retrieved
for query 451.
Table 16. Statistics on the Probabilistic Model, based On DFR-BM25
NR (Non-relevant) Non-Relevant Selected 3632 Missed : 15564

11
Table 17. The Selection of the Non-Relevant Documents for Query 451
Rank
Status
Classified
Documents
RJL
4511
4512
4513
S1
4514
4515
4516
S2
100 FBIS4-6284 N 0 0 0 T 1 0 0 T S NS
101 LA090190-0081 N 0 1 0 1 0 1 M
102 LA092990-0090 N 0 1 0 1 0 1 M
103 LA090690-0256 N 0 0 0 T 1 0 0 T S NS
104 FT933-499 R 0 1 0 1 0 1 RR
105 FT943-11390 R 0 1 0 1 0 1 RR
106 LA082490-0156 N 0 1 0 1 0 1 M
107 FT933-496 R 0 1 0 1 0 1 RR
108 LA092290-0079 N 0 1 0 1 0 1 M
109 FT923-7518 N 0 0 0 T 1 0 0 T S NS
110 LA081090-0021 N 0 1 0 1 0 1 M
111 FT932-4802 R 0 0 1 0 1 1 RR
112 FT941-15071 N 0 0 1 0 1 1 M
113 LA051989-0015 R 0 1 0 1 0 1 RR
114 LA081690-0050 N 0 1 0 1 0 1 M
115 FT921-16122 R 0 0 0 T 1 0 0 T S F
116 LA042289-0128 N 0 0 1 0 1 1 M
117 FT911-4070 R 0 1 0 1 0 1 RR
118 FT923-6463 N 0 1 0 1 0 1 M
119 LA053190-0100 R 0 1 0 1 0 1 RR
Four documents, “FBIS4-6284”, “LA090690-0256”, “FT923-7518”, and “FT921-16122” were
classified as non-relevant and therefore are shifted down the list. Consequently, the relevant
documents “FT933-499”, “FT943-11390”, “FT933-496”, “FT932-4802“, “LA051989-0015”,
“FT911-4070”, and “LA053190-0100” are lifted four ranks to be positioned higher in the list.
Therefore, precision will improve. It should be noted here that the document “FT921-16122”
detected by our technique as non-relevant is actually relevant in the RJL and therefore, is
classified as “False Alarm”.
7.4 Analysis of the precision and recall
Table 18 compares the results of the baseline and the new technique for all models using
precision and recall at pre-established recall levels.
Table 18. Precision Values at Different Recall Levels
Rank
Vector model
using TFIDF
Probabilistic model
using BM25
Probabilistic model
using DFR-BM25
Precision %
Improve
Precision %
Improve
Precision %
ImproveBase DNR Base DNR Base DNR
0.0 0.025 0.029 15 0.047 0.045 -5 0.032 0.033 3
0.1 0.021 0.022 4 0.021 0.026 27 0.021 0.024 15
0.2 0.019 0.020 4 0.019 0.023 20 0.019 0.021 12
0.3 0.018 0.018 1 0.018 0.021 22 0.018 0.018 0
0.4 0.017 0.017 0 0.016 0.020 23 0.017 0.017 0
0.5 0.015 0.015 0 0.015 0.016 6 0.015 0.015 0

12
0.6 0.013 0.013 0 0.013 0.013 2 0.013 0.013 0
0.7 0.011 0.011 0 0.007 0.007 0 0.011 0.011 0
0.8 0.004 0.004 0 0.004 0.004 0 0.004 0.004 0
0.9 0.001 0.001 0 0.001 0.001 0 0.001 0.001 0
1.0 0.000 0.000 0 0.000 0.000 0 0.000 0.000 0
Figure 1. Comparing the baseline to DNR based on TFIDF
Figure 2. Comparing the baseline to DNR based on BM25
Figure 3. Comparing the baseline to DNR based on DFR-BM25
0.00
0.01
0.01
0.02
0.02
0.03
0.03
0.04
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Precision
Recall
Probablistic model based DFR-BM25
baseline
DNR
0.00
0.01
0.01
0.02
0.02
0.03
0.03
0.04
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Precision
Recall
Vector model based TFIDF
baseline
DNR
0.00
0.01
0.02
0.03
0.04
0.05
0.06
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Precision
Recall
Probablistic model based BM25
baseline
DNR

13
8. CONCLUSIONS
In this paper, we have introduced a new technique, Demoting Non-Relevant documents (DNR)
that improves the precision of search engines by detecting and demoting non-relevant documents.
The new technique is tested on WT2g test collection using variant retrieval models. The results
show that the new technique outperformed the baseline on low recall level when tested using the
vector model based on the TFIDF weighing measure, the probabilistic models based on the
BM25, and DFR-BM25 weighing measures.
When DNR is tested in the vector model based on TFIDF weighting measure, 3781 documents
were found non-relevant and were pushed down the list. When DNR is tested in the probabilistic
model based on BM25 weighting measure, 3631 documents were found non-relevant and were
pushed down the list. When DNR is tested in the probabilistic model based on DFR_BM25
weighting measure 3632 documents were found non-relevant and were pushed down the list.
The main limitation of the DNR technique is the time required to generate the new queries and to
apply the heuristic to the documents retrieved from each query. Further research should be done
on larger test collections to determine if the precision and recall values found in this paper can
also be applied to different collections. The experiments will be done on GOV23
which consist of
25,205,179 documents.
ACKNOWLEDGMENTS
This work was fully supported by the National Council for Scientific Research (CNRS) in Beirut,
Lebanon
REFERENCES
[1] R. P. S. H. Manning CD, An Introduction to Information Retrieval DRAFT, vol. 1, Cambridge:
Cambridge University Press, 2008.
[2] R.-N. B. Baeza-Yates R., Modern Information Retrieval, vol. Vol. 463. , New York: New York:
ACM Press , 1999.
[3] D. M. R. R. G. B. B. Carpineto C., ““An Information Theoretic Approach to Automatic Query
Expansion”,” ACM Transactions on Information Systems (TOIS), vol. 19, no. 1, pp. Pages 1-27,
2001.
[4] R. P. S. H. Manning C., ““An Introduction to Information Retrieval”,” Natural Language
Engineering, vol. 16, no. 1, pp. 100-103, 2010.
[5] B. W. Croft, D. Metzler and T. Strohman, “"Search engines",” Information retrieval in practice, vol.
2, no. 2, pp. 13-28, 2010.
[6] B. Croft, D. Metzler and T. Strohman, Search Engines: Information Retrieval in Practice, Pearson
Education, Inc., 2015.
[7] G. G. Chowdhurry , Introduction to Modern Information Retrieval, Neal-Schuman Publishers, 2010.
[8] B. W. Croft, D. Metzler and T. Strohman, “"Search engines",” Information retrieval in practice, vol.
2, no. 2, pp. 13-28, 2010.
[9] Q. Ai, L. Yang, J. Guo and W. B. Croft, ““Analysis of the Paragraph Vector Model for Information
Retrieval",” in Proceedings of the ACM SIGIR International Conference on the Theory of
Information Retrieval (ICTIR 2016), Italy, 2016.
[10] C. N. S. V. T. N. Jain A, “Information Retrieval using Cosine and Jaccard Similarity Measures in
Vector Space Model,” vol. 164, no. 6, 2017.
3
http://ir.dcs.gla.ac.uk/wiki/Terrier

14
[11] C. W. B. Zamani H., ““Estimating Embedding Vectors for Queries",” in Proceedings of the ACM
SIGIR International Conference on the Theory of Information Retrieval (ICTIR), Newark, DE, USA,
2016.
[12] D. Z. J. E. Berry M. W., ““Matrices, Vector Spaces, and Information Retrieval”,” SIAM Review.,
vol. 41, no. 2, p. Vol. 41, 1992.
[13] J. . H. Paik, “A novel TF-IDF weighting scheme for effective ranking,” in SIGIR '13 Proceedings of
the 36th international ACM SIGIR conference on Research and development in information
retrieval, Dublin, 2013.
[14] J. A. B. A. Kumari M, “Synonyms Based Term Weighting Scheme: An Extension to TF.IDF,” vol.
89, no. 1, 2016.
[15] S. Robertson and H. Zaragoza, The Probabilistic Relevance Framework, Hanover, MA, USA: Now
Publishers Inc, 2009.
[16] G. Amati and C. J. Van Rijsbergen, “Probabilistic Models of Information Retrieval Based on
Measuring the Divergence from Randomness,” ACM Transactions on Information Systems, p. 57–
389, 2002.
[17] S. Clinchant and E. Gaussier, “Bridging Language Modeling and Divergence from Randomness
Models: A Log-Logistic Model for IR,” in Conference on the Theory of Information Retrieval -
ICTIR 2009: Advances in Information Retrieval Theory, Berlin, 2009.
[18] S. Robertson, “The Probabilistic Relevance Framework - BM25 and Beyond,” Foundations and
Trends in Information Retrieval, pp. 333-389, 17 December 2009.
[19] Hiemstra D., “Information Retrieval Models,” in Information Retrieval: Searching in the 21st
Century, Vols. pp 2-19, UK, A John Wiley and Sons, Ltd., Publication, 2009.
[20] T. Roelleke, Information Retrieval Models- Foundations and Relationships, London: Morgan &
Claypool Publishers, 2013.
[21] S. Robertson, S. Walker, M. Hancock-Beaulieu, M. Gatford and A. Payne, “OKAPI at TREC-4,” in
In the Proceedings og the 7th Text retrieval Conference (TREC7), Gaithersburg, 1995.
[22] R. Rosenfeld, “Two decades of statistical language modeling: where do we go from here?,” in
Proceedings of the IEEE, 2000.
[23] L. S. B. G. Beel J., “A Novel Term-Weighting Scheme for User Modeling based on Users Personal
Document Collections,” in Proceedings of the iConference 2017, China, 2016.
[24] S. M. H. D. K. W. Verberne S, “Evaluation and analysis of term scoring methods for term
extraction,” vol. 19, no. 5, 2016.
[25] Sparck-Jones K., ““Experiments in Relevance Weighting of Search Terms”,” Information Processing
& Management., vol. 15, no. 3, pp. 133-144. ISSN: 0306-4573, 1979.
[26] M. Sanderson, “Test Collection Based Evaluation of Information Retrieval Systems,” The essence of
knowledge, Sheffield, 2010.
[27] F. Yamout and M. Makary, “Building Relevant Judgment List with Minimal Human Intervention,”
International Journal of Advanced Computer Technology (IJACT), 2015.
[28] H. Harankhedkar, “Techspirited,” 01 01 2019. [Online]. Available: https://techspirited.com/internet-
its-uses-in-our-daily-life.
[29] H. C. M. M. Baeza-Yates R., "Query Recommendation Using Query Logs in Search Engines", vol.
16, Berlin, Heidelberg: Springer, 2005, pp. 588-596.

15
AUTHORS
Fadi Yamout is a PhD holder in Computer Sciences from the University Of
Sunderland, United Kingdom, in the area of Information Retrieval and Search Engines.
He worked as Chairman of Computer Science and for Middle East Airlines as Head of
planning and control department. He worked in the capacity of System Analyst and
Programmer for Societe de Service d'Informatique in Lebanon and Somapro in Canada.
Mireille Makary is an MS holder in Computer Sciences from the University Of
Balamand, Lebanon. Worked as Campus coordinator of Computer Science and have
extensive teaching experience in Computer Science courses at all different levels.

IMPROVING SEARCH ENGINES BY DEMO

More Related Content

What's hot

Similar to IMPROVING SEARCH ENGINES BY DEMO

Recently uploaded

IMPROVING SEARCH ENGINES BY DEMO