@article{ 10.

5120/11587-6925,
author = { R. V. V. Murali Krishna, S. Y. Pavan Kumar, Ch. Satyananda Reddy },
title = { A Hybrid Method for Query based Automatic Summarization System },
journal = { International Journal of Computer Applications },
issue_date = { April 2013 },
volume = { 68 },
number = { 6 },
month = { April },
year = { 2013 },
issn = { 0975-8887 },
pages = { 39-43 },
numpages = {9},
url = { https://ijcaonline.org/archives/volume68/number6/11587-6925/ },
doi = { 10.5120/11587-6925 },
publisher = {Foundation of Computer Science (FCS), NY, USA},
address = {New York, USA}

International Journal of Computer Applications (0975 – 8887)

Volume 68– No.6, April 2013
39
ABSTRACT
Automatic text summarization is one of the research goals of
Natural Language Processing which relieves humans from
studying each and every line in a text document to understand the
underlying concepts in it. Automatic text summarization is aimed
to create a brief outline of a given text covering the important
points in the text. Automatic text summarization can be generic or
query specific. This paper is focused on Query specific text
summarization where a summary of the given text is constructed
based on the given query. Query specific text summarization is
based on the calculation of the relationship between sentences in
the text document and the query given. Several statistical
techniques and linguistic techniques have been developed to find
the relationship between the given query and the sentences in the
document. These methods when used alone could not give desired
accuracy in the results. In this paper a sentence scoring method is
defined based on existing sentence scoring methods. It attempts to
combine the individual results of these methods to give a better
assessment of the relationship between the sentences.
General Terms
Automatic text summarization ;Natural language processing;
statistical techniques; linguistic techniques; stemming; n-gram;
clustering
1. INTRODUCTION
Text summarization is the process of extracting key
phrases/sentences from the given text and presenting them in a
readable format to the user. The text presented to the user is called
as summary. There are 2 types of summaries 1) indicative 2)
informative. Indicative summaries do not carry the essence of the
original text but give an idea about the information. Informative
summaries carry the essence of the original text. There can be
two types of informative summaries. They are 1) Extractive
summary 2) abstractive summary. Extractive summaries carry the
essence of the original text though the words/sentences present in
the original text. Abstractive summaries carry only the essence in
the original text but not its words/sentences.
1.1 Main tasks in extractive text summarization are
1.1.1 Keyword extraction
1.1.2 Sentence scoring
1.1.3 Sentence ordering
1.2 Main tasks in abstractive text summarization are
1.2.1 Content selection
1.2.2 Sentence compression
1.2.3 Sentence fusion
1.2.4 Sentence generation
Various commercial tools are available for extractive text
summarization. There are no commercial tools available for
abstractive text summarization. Lot of research has been done on
extractive summarization areas like keyword extraction, sentence
scoring methods and sentence ordering.
Little research has been done in the areas like content selection,
sentence compression, sentence fusion and sentence generation of
abstractive text summarization. Depending on what the
summarization program focuses to make the summary of the text,
summaries can be generic or query relevant. Summary can be
generated for an individual document or for a collection of
documents. On the basis how summary is generated, Summary
generation can be broadly divided as abstractive and extractive. In
abstractive summary generation, the abstract of the document is
generated. The summary so formed need not have exact sentences
as present in the document. In extractive summary generation,
important sentences are extracted from the document based on
sentence similarity methods. The generated summary contains all
such extracted sentences arranged in a meaningful order. This paper
mainly focused on query specific extractive summarization of
source document. Extractive techniques copy the information
deemed most important by the system to the summary (for
example, key clauses, sentences or paragraphs)
2. RELATED WORK
Gholamrezazadeh, Saeedeh et. Al have done comprehensive survey
on text summarization systems. They have given a detailed
understanding of various approaches used in text summarization
systems. In statistical approaches the sentence selection is done
based on statistics of the text like word frequency, position of the
sentence, location of the word etc… [1].These methods are based
on the idea that text surface cues are the most obvious indication of
the text contents. There are several methods for determining the
key sentences such as, The Title Method [2], The Location Method
[3], The Aggregation Similarity Method [5], The Frequency
Method [4], TF- Based Query Method [2] etc. Some examples of
different systems which use these methods for generating a
summary are IN XIGHT’s [6] Summary Server is an application
that creates extraction-based summaries offline. It uses statistical
extraction techniques based on features such as sentence position,
sentence length and keywords. University of southern California
developed the SUMMARIST [6] system which produces
summaries of web documents. It first identifies the main topics of
the document using statistical techniques based on features such as
position, and word counts.
Linguistic approaches are based on considering the semantic
relationship between the sentences. This semantic relationship can
be calculated by using several methods like Word Net [7], How
Net, Google similarity distance [8] etc.
The following are the various statistical and linguistic
techniques used in this paper for calculating similarity score
between the two sentences.
A Hybrid Method for Query based Automatic
Summarization System
R. V. V. Murali Krishna
G V P College of Engineering(A)
Visakhapatnam
AP, INDIA
S. Y. Pavan Kumar
G V P College of Engineering(A)
Visakhapatnam
AP, INDIA
Ch. Satyananda Reddy
Andhra University
Visakhapatnam
AP, INDIA
International Journal of Computer Applications (0975 – 8887)
Volume 68– No.6, April 2013
40
2.1 Sentence scoring methods
Sentence scoring methods are very important in a document
summarization. The efficiency of a summarization system mostly
depends on the sentence scoring methods. The main task of the
sentence scoring methods is to identify set of sentences which will
carry important data in the given document. The scoring methods
are also known as sentence similarity measures. The following are
the various statistical and linguistic techniques used in this paper
for calculating similarity score between the two sentences.
2.1.1 Statistical Techniques:
2.1.1.1 Word form similarity [9]
2.1.1.2 N-gram based similarity [12]
2.1.1.3 Word Order Similarity [10]
2.1.2 Linguistic Techniques:
2.1.2.1 Semantic similarity [10]
2.1.1.1 Word form similarity:
The word form similarity is mainly used to describe the form
similarity between two sentences, is measured by the number of
same words in two sentences. The sentences are preprocessed to
filter the keywords from the overall words in the sentence. If S1
and S2 are two sentences, the word form similarity is calculated by
using the below formula
Word form similarity (s1, s2) =
(2*Number of Same words (s1, s2)) / (Len (s1) + Len (s2)) *100
Example: The following example shows how the wordform
similarity is calculated between a sample query and a sentence
in a document
Query: Computer design contains memory chips. Sentence:
Memory chips are the main parts of a computer.
Query is partitioned in to tokens: Computer
|design|contains | memory | chips.
After applying stop list: Computer |design |memory |chips.
Applying Stemming: comput | design | memori | chip.
Sentence is partitioned in to tokens: Memory |chips| are |the|main
|parts |of |a |computer.
After applying stop list: Memory| chips |computer.
Applying stemming algorithm: Memori| chip|
comput.Calculation of Word Form Similarity for above two
sentences:
Word form similarity (s1, s2) =
(2*Number of Same words (s1, s2)) / (Len (s1) + Len (s2)) *100
Similarity (S1, S2) = ((2*3)/ (4+3)) * 100
Similarity (S1, S2) =6/7=0.8571* 100 = 85.71%
Similarity (S1, S2) = 85.71%
Therefore we can say that the above two sentences are
almost similar.
2.1.1.2. N-gram based similarity:
N-grams are fixed length consecutive series of “n” characters.
An n-gram of size 1 is referred to as a “unigram”, an n-gram of
size 2 is a “bigram”, an n-gram of size 3 is a “trigram”, size 4 or
more is simply called an “n-gram”. The n-gram based similarity
can be calculated after removing the stop words from the two
sentences by using the formulae.
N-gram similarity (s1, s2) =
( 2 * (no of common N-grams in s1 & s2) ) / (Total no of N –
grams in s1 & s2 ) * 100
Example: The following example shows how the N-grambased
similarity is calculated between a sample query and a sentence in a
document
Query: Computer design contains memory chips. Sentence:
Memory chips are the main parts of a computer.
Query is partitioned in to tokens: Computer |design|contains |
memory | chips.
After applying stop list: Computer |design |memory |chips.
Splitting the query in to bigrams:Computer: co
|om |mp |pu |ut |te |er Design: de|es |si
|ig |gn
Memory: me |em| mo| or| ry
Chips: ch |hi| ip |ps
Sentence is partitioned in to tokens: Memory |chips| are |the|main
|parts |of |a |computer.After applying stop list: Memory| chips
|computer.
Splitting the sentence in to bigrams:
Memory: me |em| mo| or| ry
Chips: ch |hi| ip |ps
Computer: co |om |mp |pu |ut |te |er
Calculation of bi-gram based Similarity for above two
sentences as:
N-gram similarity (s1, s2) =
(2 * (no of common bi
- grams in s1 & s2) )/(Total no of bi –
grams in s1 & s2 ) * 100
N-gram base similarity (s1, s2) = ((2 *16 )/ (21+16) )* 100
= (32/37 ) * 100
= 86.5%
Therefore we can say that the above two sentences are almost
similar.
2.1.1.3. Word Order similarity:
Sentences containing the same words but in different orders
may result in very different meanings. It is easy for humans to
process word order information. However the incorporation of
order information in to computational methods for understanding
natural language is a difficult challenge. This may be the reason
why most existing methods do not tackle this type of information.
In this section we introduce a method that takes word order
information into account when computing sentence similarity .For
two sentences s1 and s2, r1 and r2 are the two arrays where r1
refers to array containing indexes of the words in the sentence s1
and r2 is an array containing indexes of the words in the sentences
s2 w.r.t s1 then the word order similarity is calculated using the
formulae
Word Order Similarity (s1, s2) =
 (1 - (

))X 100

Example: The following example shows how the Word order

similarity is calculated between a sample query and a sentence in a
document.
Query: The dog jumps over a fox.
Sentence: The fox jumps over a dog.
Query is partitioned in to tokens: The |dog |jumps |over |a|fox.
After applying stop list: dog |jumps |fox.
Assigning the indexes: dog [0] |jumps [1] |fox [2].
Sentence is partitioned in to tokens: The |fox |jumps |over |a|dog.
After applying stop list: fox |jumps |dog.
Assigning the indexes w.r.t query: fox [2]| jumps [1] |dog[0]
Calculation of Word Order Similarity for above two sentences:
International Journal of Computer Applications (0975 – 8887)
Volume 68– No.6, April 2013
41
Word Order Similarity (s1, s2) =
( 1 – ( | 0 – 2 | + | 1 – 1 | + | 2 – 0 | ) / ((0+2)+(1+1)+(2+0) ) ) * 100
Word Order Similarity (s1, s2) = 1/3 * 100=33.4%.
Hence S1 and s2 are almost dissimilar
2.1.2.1. Semantic similarity:
Semantic similarity aims to find the similarity between two
sentences based on the meaning of the individual words in the
sentences. For a given query we create a set of related sentences
by replacing an important word (i.e. noun, verb, adverb and
adjective) by its synonym(s).so, the steps involved in calculating
the semantic similarity are parts of speech tagging to the query,
applying the stop list to remove insignificant words, finding the
synonyms for every key word in a query, finally computing the
similarity between the query synonyms and the sentences. So here
for parts of speech tagging we use standard Stanford NLP POS
tagger and for identifying the synonyms we use standard WordNet
dictionary.
Example: The following example shows how the semantic
similarity is calculated between a sample query and a sentence in
a document.
Query: Computer is logical device.
Sentence: laptop is legitimate machine. `
Parts of speech tagging to the query:
Computer –noun is –verb logical–adjective device–noun. Query is
partitioned in to tokens:
Computer –noun| is –verb| logical–adjective| device–noun.
After applying stop list:
Computer –noun| logical–adjective| device–noun.
Finding the synonyms for every key word in the sentence using
WordNet Dictionary:
Computer: processor, CPU, mainframe,, laptop, pc Logical:
rational, reasonable, sound, legitimate, valid.
Device: machine, tool, mechanism, gadget.
Sentence is partitioned in to tokens: laptop |is |
legitimate|machine.
After applying stop list: laptop | legitimate |machine.
Finally finding the number of common key words between the
words (keywords along with their synonyms) obtained from the
query and the sentence.
Semantic similarity (s1, s2)=2* No Of Common Keywords in s1
& s2/ (total no of keywords in s1 and s2))
Semantic similarity (s1,s2)= 2*3/(3+3)=1
2.2 Sentence Clustering
After retrieving the important sentences by applying the
proposed sentence scoring method, there is a need to check for
redundancy among the sentences. Two or more sentences may
convey same or similar meaning. One way to remove redundancy
is though sentence clustering. In this paper clustering of sentences
is done based on the clustering method proposed by Harish
Karnicket. al [11]. The extracted sentences S are arranged in
ascending order on the basis of score, then first sentence is
selected and its similarity is measured with all the other sentence,
sentences having similarity above the threshold are removed from
the set S Similarly the procedure is repeated for all the sentences
and then the outcome will be a summary without redundancy.
2.3 Sentence Ordering
To generate a readable coherent summary it is very important
to order the sentences correctly. For a single document the order of
sentence in the original document can be used as order to generate
summaries. Alternately the sentences can be presented in
descending order of their score [11].
3. PROPOSED SENTENCE SCORING
METHOD
The proposed method finds relationship between a user query
and a sentence in the text document. This method is defined in
terms of the existing statistical and linguistic techniques. Sentences
are extracted from the document based on the score given by the
proposed sentence scoring method. These extracted sentences may
contain redundant information and it is handled by using an
iterative clustering algorithm. The final outcome contains the most
important sentences in the document related to the query without
redundancy which is known as summary of the document. The
resultant sentences can be ordered according to their sentence score
or order of sentences in the original document. The entire
summarization process is shown in Figure 1.
Figure 1． Summarization Process
Proposed sentence scoring method = ((a+b+c)/3+d)/2
where a, b, c and d refer to the similarities obtained by the sentence
scoring methods Word form similarity, N-gram based similarity ,
Word Order Similarity and Semantic similarity respectively.
4. EVALUATION METRICS
There are many measures that can calculate the topical
similarities between two summaries. For evaluation the results we
use two methods. The first one is by precision (P), recall (R) and
F1-measure which are widely used in Information Retrieval. For
each document, the manually extracted sentences are considered as
the reference summary (denoted by Summ_ref). In this approach
the candidate summary (denoted by Summ_cand)is compared with
the reference summary. The formulas for the measures P, R and F1
are as follows[9].Depending on what the summarization program
focuses to make the summary of the text, summaries can be generic
or query relevant. This paper mainly focused on query specific
extractive summarization of source document. In extractive
summarization each sentence in a document is scored based on
some scoring method with respect to the query. The final outcome
contains the most important sentences in the document related to
the query without redundancy which is known as summary of the
document. For single document order of sentence in the original
document can be used as order to generate summaries.

P = ) /
International Journal of Computer Applications (0975 – 8887)
Volume 68– No.6, April 2013
42
R = ( )
/
F1= (2*P*R) / (P+R)
The second measure we use the ROUGE (Recall-Oriented
Understudy for Gisting Evaluation), which was adopted by NIST
(DUC) for evaluating automatic summarization system. It has
been shown that ROUGE is very effective for measuring
document summarization .It measures summary quality by
counting overlapping units such as the N-gram, word sequences
and word pairs between the candidate summary and the reference
summary. The ROUGE-N [9] measure compares N-grams of two
summaries, and counts the number of matches. The measure is
defined by formula
ROUGE-N=
( ) / (
)
Where N stands for the length of the N-gram, Count_match
(N-grams) is the maximum number of N-grams co-occurring in
candidate summary and a set of reference summaries. Count (Ngrams) is the number of
N-grams in the reference summaries.
5. IMPLEMENTATION
The entire system is implemented in java programming
language (J2SE). The following are the inputs and outputs of the
developed system.
Input Document : Automatic summarization of a text
document based on user query.doc
Query : what is automatic summarization of
a text document Based on a user query?
Output (Summary):
Technology of automatic text
summarization plays an important role in Information Retrieval
Systems and Text Mining, and provides a solution to the
information overload problem. Automatic text summarization is
the creation of a shortened version of a text by a computer
program. Automatic text summarization can be generic or query
specific. This paper is focused on Query specific text
summarization. Query specific text summarization is based on the
calculation of the relationship between sentences in the text
document and the query given. The technology of automatic text
summarization is maturing and may provide a solution to the
information overload problem. Automatic document
summarization aims to condense the original text into essential
content and to assist in filtering and selection of necessary
information.
6. RESULT ANALYSIS
TABLE 1.Output in terms of no of sentences when each
of the methods are applied individually on the source
document consisting of 186 sentences
Sentence Scoring methods Resulting No. of
sentences
from the source
document
Word Form Similarity 3
N-gram Based Similarity 30
Word Order Similarity 7
Semantic Similarity 9
Proposed Similarity 12 Measure
Figure 2: Graphical comparison of different sentence
scoring methods
TABLE 2．Comparision of sentence scoring methods
based on standard summary evaluation measures
Sentence Precision Recall F1- ROUGE-2
Scoring Measure
Method
Word Form 0.27 0.33 0.29 0.64
Similarity
N-gram 0.16 0.55 0.24 0.95
Based
Similarity
Word 0.14 0.11 0.12 0.85
Order
Similarity
Semantic 0.33 0.33 0.32 0.87
Similarity
Proposed 0.50 0.66 0.56 0.93
Similarity
Note : The abstract of the document is taken as the reference
summary to evaluate automatic summary.
7. CONCLUSION
It is observed from the results that the proposed sentence
scoring method has given more accuracy in sentence scoring than
the existing methods. Hence summary obtained is more relevant
and close to manually generated summary.
8. FUTURE SCOPE
The proposed sentence scoring method is based on the average
of the values scored using statistical techniques and linguistic
techniques. Instead of just calculating average of the values, a
weighted average can be taken where the weights for each of the
values can be chosen based on the appropriateness of the methods
applied for the content in the given document. Since a large size
document will usually result in large no of relevant sentences the
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
F1-measure
Rouge-2
International Journal of Computer Applications (0975 – 8887)
Volume 68– No.6, April 2013
43
clustering process can be implemented via a parallel processing
technique like Threads in java to speed up the overall process of
summarization.
9. REFERENCES
[1] Gholamrezazadeh, Saeedeh; Salehi, Mohsen Amini;
Gholamzadeh, Bahareh, "A Comprehensive Survey on Text
Summarization Systems," Computer Science and its
Applications, 2009. CSA '09. 2nd International Conference
on , vol., no., pp.1,6, 10-12 Dec. 2009
doi: 10.1109/CSA.2009.5404226
[2] YoungkoongKo, JungyunSeo, “An Effective SentenceExtraction Technique Using
Contextual Information and
Statistical Approaches for Text Summarization”, Pattern
Recognition Letters. doi:10.1016/j.patrec.2008.02.008
[3] Wasson, M., “Using leading text for news summaries:
Evaluation results and implications for commercial
summarization applications”, in Proc. 17th International
Conference on Computational Linguistics and 36th Annual
Meeting of the ACL, 1998, pp.1364-1368
[4] Salton, G, Automatic Text Processing: The Transformation,
Analysis, and Retrieval of Information by Computer,
Addison-Wesley Publishing Company, 1989.
[5] Waleed al-sanie, “Towards an infrastructure for Arabic text
summarization using rhetorical structure theory”, Master
Thesis, Department of computer science. King Saud
University, Riyadh, Kingdom of Saudi Arabia, 2005.
[6] http://web.science.mq.edu.au/~swan/summarization/projects_f
ull.htm
[7] Bellegarda, J., “Exploiting latent semantic information in
statistical language modeling,” in Proc. IEEE, August
2000.Vol. 88, No. 8,pp: 1279-1296.
[8] R. L. Cilibrasi and P. M. B. Vitanyi, “The google similarity
distance,” IEEE Trans. On Knowl. and Data Eng., vol. 19, no.
3, pp. 370–383, 2007.
[9] Zhang Pei-ying and LI un-he,”Automatic text summarization
based on sentences clustering and extraction” , IEEE2009
[10] Yuhua Li, Zuhair Bandar, David McLean and James O’Shea
,”A Method for Measuring Sentence Similarity and its
application to conversational agents”
[11] Harish Karnick and VaruneshMishra,”Query Specific MultiDocument
Summarization”, Indian Institute of Technology,
Kanpur, April 25, 2010
[12] Adamson, G and J. Boreham, ” The use of an Association
Measure Based on Character Structure to identify
semantically related pairs ofwords and document titles”.