Más contenido relacionado
La actualidad más candente (19)
Similar a A domain specific automatic text summarization using fuzzy logic (20)
Más de IAEME Publication (20)
A domain specific automatic text summarization using fuzzy logic
- 1. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 4, July-August (2013), © IAEME
449
A DOMAIN-SPECIFIC AUTOMATIC TEXT SUMMARIZATION USING
FUZZY LOGIC
1
Mrs. A.R.Kulkarni
Assistant Professor, Computer Science & Engg Department,
Walchand Institute of Technology, Solapur
2
Dr. Mrs. S.S.Apte
HEAD, Computer Science & Engg Department,
Walchand Institute of Technology, Solapur
ABSTRACT
The amount of information on World Wide Web has increased enormously. In this context
there is a need for text summarization. It creates summaries of the documents that consist of
important sentences in the document. The summaries help the readers to make decision as to read the
whole document or not thus acting as a time saver. Various Techniques have been proposed for text
summarization by researchers that can be broadly classified into two types: Extraction and
Abstraction. This Paper focuses on Text Summarization by Extraction using Fuzzy Logic.. Many
Automatic text Summarization techniques have used either Statistics or Linguistics. Very Few works
has used a combination of both. Our Paper uses the idea of both Statistical and Linguistic methods.
This hybrid approach has been applied to news article dataset in the domain of technical news and
we have evaluated their performances by using precision and recall method. It is found that this
method generates good quality of summary.
Keywords: Summarization, Statistics, Linguistics, fuzzifier, Defuzzifier, Rule-Base, Extraction.
INTRODUCTION
“Text Summarization” is a process of creating a shorter version of original text that contains
the important information. The amount of information on the web is growing day by day. A
considerable amount of time is wasted in searching for relevant documents. Hence text
summarization technique came into existence which created a short summary for the text document
INTERNATIONAL JOURNAL OF COMPUTER ENGINEERING &
TECHNOLOGY (IJCET)
ISSN 0976 – 6367(Print)
ISSN 0976 – 6375(Online)
Volume 4, Issue 4, July-August (2013), pp. 449-461
© IAEME: www.iaeme.com/ijcet.asp
Journal Impact Factor (2013): 6.1302 (Calculated by GISI)
www.jifactor.com
IJCET
© I A E M E
- 2. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 4, July-August (2013), © IAEME
450
by choosing important sentences of the document. An Automatic text summarization works very
well on structured documents such as news articles, research publications and reports.
Text summarization has two approaches: Extraction and Abstraction. Extraction involves selecting
sentences of high relevance (rank) from the document based on word and sentence features and put
them together to generate summary. It uses mostly statistical methods. Abstraction procedure
examines the text, interprets it and generates summary using different sentences. It uses Linguistic
methods. This paper focuses on extractive summarization technique. It uses a combination of both
Statistical and Linguistic methods on fusion of various features to generate a better quality summary.
RELATED WORK
Since late 50s text summarization has been a crucial and important research area. The first
Automatic text summarization was created by Luhn in 1958[1] based on term frequency. Then G. J.
Rath, A. Resnick, and T. R. Savage[2] have proposed the evidences of problems in generating the
summaries using term frequency feature in 1961. Both studies are characterized by surface level
approaches. In late 60s, entity level approaches appeared: the first of its kind used syntactic analysis
proposed by Climenson [3].This was followed by Edmundson’s work[4] which used term
frequency, location features and cue words .Earliest instances of research on summarization was
done on scientific documents followed by various works published in other domains, mostly on
newswire data. In 1990s. with the advent of machine learning techniques in Natural Language
Processing, many publications came that used statistical techniques to produce document summaries.
They have used a combination of appropriate features and learning algorithms. Other approaches
have used hidden Markov models[5] and log-linear models to improve extractive summarization.
Recently, neural networks are used to generate summary for single documents using
extraction[6]. Very little work is done on automatic text summarization based on Artificial
Intelligence and evolutionary techniques. M.S.Binwale l[7] has designed automatic text
summarization using integrated hybrid model. He has used Diversity-based methods and Swarm
based methods followed by Fuzzy logic. Experimental results have shown that this model produces
good quality of summary.
Ladda Suanmali[8] in his work has used sentence weight ,a numerical measure assigned to
each sentence and then selecting sentences in descending order of their sentence weight for the
summary.
L.Antiqueira [9] has proposed a method for extractive summarization using concept of
complex networks and its metrics. It has shown that this method is capable of capturing important
text features as expected.
For MEDLINE citations, .an automatic summarization system has been introduced by
Marcelo Fiszman[10] . It is an domain-specific abstractive summarization which outperformed the
baseline summarizer considerably.
A lot of work has been done in single document and multi document summarization using
statistical methods. A lot of researchers are trying to apply this technology to a variety of new and
challenging areas, including multilingual summarization and multimedia news broadcast.
SURVEY ON NEED AND SCOPE OF TEXT SUMMARIZATION
Text Summarization is increasingly being used in the commercial sector such as
• Telephone communication industry, e.g BT’s ProSum.
• In data mining of text databases, E.g. Oracle’s Context.
• In filters for web-based information retrieval, E.g. Inxight’s summarizer used in Alta Vista
Discovery
- 3. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 4, July-August (2013), © IAEME
451
• In word Processing tools e.g. Microsoft’s AutoSummarize
• A variety of new applications are using multilingual summarization, multimedia news
broadcast, audio scanning services for the blind etc.
• To summarize news to SMS or WAP-format for mobile phones.
Many approaches differ on the manner of their problem formulations.
A BETTER APPROACH TO SUMMARIZATION
This approach uses both statistical and Linguistic methods [11]to improve the quality of
generated summary. It uses Fuzzy logic for effective Text Summarization[12]. Fuzzy logic uses
decision module that determines the degree of importance of each sentence based on its rated
features. Decision module is designed using a fuzzy inference system.
This approach is illustrated in figure 4.1
Text summarization approach consists of following stages:
• Preprocessing
• Feature Extraction
• Fuzzy logic scoring
• Sentence selection and Assembly
PREPROCESSING
It has 4 steps:
Segmentation: It is a process of dividing a given document into sentences.
Removal of Stop words: Stop words are frequently occurring words such as ‘a’ an’, the’ that
provides less meaning and contains noise. The Stop words are predefined and stored in an array.
Tokenization and POS Tagging: A standard Parser cum Tagger is used to generate tokens and tag
them with proper parts of speech such as such as nouns(NN), verbs(VBZ), adjectives(JJ) and
adverbs(ADVB), determiners(DT) coordinating conjunction(CC) etc. It also groups syntactically
correlated words into phrases such as noun phrase, verb phrase, adjective phrase etc.
Word Stemming: converts every word into its root form by removing its prefix and suffix so that it
can be used for comparison with other words.
Preprocessing Feature
selection
Fuzzification
Rule Basedefuzzification
Selection of
sentences &
assembly
Summary
Document
- 4. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 4, July-August (2013), © IAEME
452
FEATURE EXTRACTION
The text document is represented by set, D= {S1, S2,- - - , Sk} where, Si signifies a sentence
contained in the document D .The document is subjected to feature extraction. The important word
and sentence features to be used are decided .This work uses features such as Title word, Sentence
length, Sentence position, numerical data, Term weight, sentence similarity, existence of Thematic
words and proper Nouns .
1. Title word:
A high score is given to the sentence if it contains words occurring in the title as the main
content of the document is expressed via the title words. This feature is computed as follows:
If Nt is the number of words in the sentence that occur in the title and Ntotal is the total number of
words in the title, then
ࡲ ൌ
ܜۼ
ܔ܉ܜܗܜۼ
2. Sentence Length:
We eliminate the sentences which are too short such as datelines or author names. For every
sentence the normalized length of sentence is calculated as
۴ ൌ
ܚ܍܊ܕܝۼ ܗ ܛ܌ܚܗܟ ܑܖ ܍ܐܜ ܍܋ܖ܍ܜܖ܍ܛ
ܚ܍܊ܕܝۼ ܗ ܛ܌ܚܗܟ ܑܖ ܍ܐܜ ܜܛ܍ܖܗܔ ܍܋ܖ܍ܜܖ܍ܛ
3. Sentence Position:
The sentences occurring first in the paragraph have highest score. Suppose a paragraph has n
sentences then the score of every sentence for this feature is calculated as follows:
F3(S1) = n/n; F3(S2)=4/5; F3(S3)=3/5; F3(S4)=2/5; and so on.
4. Numerical data:
The sentences having numerical data can reflect important statistics of the document and may
be selected for summary. Its score is calculated as:
۴ሺܑ܁ሻ ൌ
ܚ܍܊ܕܝۼ ܗ ܔ܉܋ܑܚ܍ܕܝܖ ܉ܜ܉܌ ܑܖ ܍ܐܜ ܍܋ܖ܍ܜܖ܍ܛ ܑ܁
܍܋ܖ܍ܜܖ܍܁ ܐܜܖ܍ۺ
5. Thematic words:
These are domain specific words with maximum possible relativity. The score for this feature
is calculated as the ratio of the number of thematic words that occurs in a sentence over the
maximum number of thematic words in a sentence.
۴ሺܑ܁ሻ ൌ
ܚ܍܊ܕܝۼ ܗ ܋ܑܜ܉ܕ܍ܐ܂ ܉ܜ܉܌ ܑܖ ܍ܐܜ ܍܋ܖ܍ܜܖ܍ܛ ܑ܁
ܠ܉ۻ ܗܖ ܗ ܋ܑܜ܉ܖ܍ܐܜ ܛ܌ܚܗܟ
6. Sentence to Sentence Similarity:
For each sentence S, the similarity between S and every other sentence is computed by the
method of token matching. The [N][N] matrix is formed where N is the total number of sentence in a
document. The diagonal elements of a matrix are set to zero as the sentence should not be compared
with itself. The similarity of each sentence pair is calculated as follows
- 5. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 4, July-August (2013), © IAEME
453
ࡲ ൌ
∑ሾSimሺSi, ሺSjሻሿ
MaxሾSimሺSi, ሺSjሻሿ
Where i=1 to N and j=1 to N.
7. Term weight:
The score of this feature is given by the ratio of summation of term frequencies of all terms in
a sentence over the maximum of summation values of all sentences in a document.
It is calculated by the following equation.
F7=∑TFI
--------------- Where i=1 to n, n is the number of terms in a sentence.
MAX(∑TFI )
8. Proper Nouns:
The sentence that contains maximum number of proper nouns is considered to be important.
Its score is given by
F8= Number of proper nouns in the sentence s
--------------------------------------------------
Sentence length(s)
Thus each sentence is associated with 8 feature vector. Using all the 8 feature scores, the
score for each sentence are derived using fuzzy logic method. The fuzzy logic method uses the fuzzy
rules and triangular membership function .The fuzzy rules are in the form of IF-THEN .The
triangular membership function fuzzifies each score into one of 3 values that is LOW,MEDIUM &
HIGH. Then we apply fuzzy rules to determine whether sentence is unimportant, average or
important. This is also known as defuzzification.
For example IF (F1is H) and (F2 is M) and (F3 is H) and (F4 is M) and (F5 is M) and (F6 is
M) and (F7 is H) and (F8 is H) THEN (sentence is important).
All the sentences of a document are ranked in a descending order based on their scores. Top n
sentences of highest score are extracted as document summary based on compression rate. Finally
the sentences in summary are arranged in the order they occur in the original document.
EVALUATION METHODOLOGY
The evaluation of the summaries is done based on two factors mentioned in Fig. 5. We used
2 documents from news articles belonging to technical domain as an input to the system. Here the
human generated summaries are used as reference summaries for evaluation of our results. The
human generated summary acts as a reference summary since humans can capture and relate deep
meanings of the text as compared to machines. We received human generated summaries for our
input documents from different Experts. Here we call the summaries of Fuzzy summarizer, online
summarizer 1,online summarizer 2 as the candidate summaries.
The performance of the proposed approach will be evaluated using precision, recall and F-
measure[12]. Precision evaluates the proportion of correctness for the sentences in the summary
whereas recall is utilized to evaluate the proportion of relevant sentences included in the summary.
- 6. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 4, July-August (2013), © IAEME
454
For precision, the higher the values, the better the system is in omitting irrelevant sentences.
Similarly, the higher the recall values the more successful the system would be in fetching the
relevant sentences. The weighted harmonic mean of precision and recall is called as F-measure. The
detail formula for Precision, recall and F-measure is as shown below.
Precision = │ {Retrieved sentences} ∩ {Relevant sentences}│
-------------------------------------------------------------
│ {Retrieved Sentences} │
Recall= │ {Retrieved sentences} ∩ {Relevant sentences} │
__________________________________________
│ {relevant sentences} │
F-measure= 2 x
୰ୣୡ୧ୱ୧୭୬ ୶ ୖୣୡୟ୪୪
୰ୣୡ୧ୱ୧୭୬ାୖୣୡୟ୪୪
EXPERIMENTAL RESULTS
The two sports news articles , their manual summaries, summaries generated by our
algorithm and summaries generated by two online summarizers are shown below. The chart showing
the comparision between results of online summarizers and our proposed summarizer.
Original Document 1
- 7. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 4, July-August (2013), © IAEME
455
Original Document 2
Manual summary for Document 1
- 8. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 4, July-August (2013), © IAEME
456
Manual summary for Document2
For document 1, the summary generated by our algorithm is:
- 9. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 4, July-August (2013), © IAEME
457
For document 2, the summary generated by our algorithm is:
Online summarizer1
• Document 1
- 10. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 4, July-August (2013), © IAEME
458
• Document 2
Online Summarizer 2
• Document 1
- 11. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 4, July-August (2013), © IAEME
459
• Document 2
Comparison Graphs
• Document 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Summarizer 1 Summarizer 2 Our
Summarizer
Precision
Recall
f-measure
- 12. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 4, July-August (2013), © IAEME
460
Document 2
CONCLUSION
Automatic summarization is a complex task that consists of several sub-tasks. Each of the
sub-task directly affects the ability to generate high quality summaries. In extraction based
summarization the important part of the process is the identification of important relevant sentences
of text. Use of fuzzy logic as a summarization sub-task improved the quality of summary by a great
amount. The results are clearly visible in the comparison graphs. Our algorithm shows better results
as compared to the output produced by two online summarizers.
FUTURE SCOPE
The quality of summary can still be improved by using topic segmentation and semantic
analysis of the text in addition to the features considered above. We applied our method for single
document summarization which could be extended for multiple document summarizations.
REFERENCES
1. LUHN. H.P.1958. “Automatic Creation of Literature abstracts”,IBM Journal of Research &
Development 2 April p-159.
2. G.J.Rath, A Rensick and T.R.Savage “The formation of abstracts by selection of sentences”,
at IBM Foundation, Yorktown Heights, New York.
3. Climenson, W.D., Hardwick, N.H., Jacobson, S.N. (1961).”Automatic Syntax Analysis in
Machine Indexing and Abstracting”.
4. Edmundson, H.P. (1969).New Methods in Automatic Extracting.
0
0.2
0.4
0.6
0.8
1
Summarizer 1 Summarizer 2 Our
Summarizer
Precision
Recall
f-measure
- 13. International Journal of Computer Engineering and Technology (IJCET), ISSN 0976-
6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 4, July-August (2013), © IAEME
461
5. M. Suneeta & S.Sameen Fatima “Corpus based Automatic Text Summarization System with
HMM Tagger” at IJSCE ISSN: 2231- 2307,Volume-1, Issue-3, July 2011
6. Kaikhah.K “Automatic Text Summarization using neural networks” at Intelligent systems
2004 Proceedings,2004 2nd
International IEEE Conference Volume 1.
7. Binwahlan.M.S., Salim.N.,& Suanmali.L (2009d),”Fuzzy Swarm based text summarization”,
Journal of Computer Science, 5(5), 338-346
8. LaddaSuanmali, NaomieSalimand Mohammed Salem Binwahlan, “Fuzzy Logic Based
Method for Improving Text Summarization”, (IJCSIS) International Journal of Computer
Science and Information Security, Vol. 2, No. 1, 2009
9. L. Antiqueira, O. N. Oliveira Jr., L. F. Costa, and M. G. V. Nunes.”A complex network
approach to text summarization “at Information Sciences 179(5):584-599 (2009) ”.
10. Marcelo Fiszman,Thomas.C.Rendflesh, Halil Kilicoglu “Abstraction summarization for
managing the biomedical research literature”,CLS '04 Proceedings of the HLT-NAACL
Workshop on Computational Lexical Semantics Pages76-83 Association for Computational
Linguistics Stroudsburg, PA, USA ©2004.
11. Rushdi Shams, M.M.A. Hashem, Afrina Hossain, Suraiya Rumana Akter, and Monika
Gope,”A corpus based web document summarization using statistical & Linguistic
approach”,
12. Ladda Suanmali , Naomie Salim and Mohammed Salem Binwahlan, “Improving Text
Summarization using Fuzzy Logic”, (IJCSIS) International Journal of Computer Science and
Information ecurity, Vol. 2, No. 1, 2009.
13. Meghana.N.Ingole, M.S.Bewoor and S.H.Patil, “Context Sensitive Text Summarization using
Hierarchical Clustering Algorithm”, International Journal of Computer Engineering &
Technology (IJCET), Volume 3, Issue 1, 2012, pp. 322 - 329, ISSN Print: 0976 – 6367,
ISSN Online: 0976 – 6375.
14. Roma V J, M S Bewoor and Dr.S.H.Patil, “Automation Tool for Evaluation of the Quality of
NLP Based Text Summary Generated Through Summarization and Clustering Techniques by
Quantitative and Qualitative Metrics”, International Journal of Computer Engineering &
Technology (IJCET), Volume 4, Issue 3, 2013, pp. 77 - 85, ISSN Print: 0976 – 6367,
ISSN Online: 0976 – 6375.
15. V.Sujatha, K.Sriraman, K. Ganapathi Babu and B.V.R.R.Nagrajuna, “Testing and Test Case
Generation by using Fuzzy Logic and N.L.P Techniques”, International Journal of Computer
Engineering & Technology (IJCET), Volume 4, Issue 3, 2013, pp. 531 - 538, ISSN Print:
0976 – 6367, ISSN Online: 0976 – 6375.