Using the F-measure as similarity measure for automatictext summarization Текст научной статьи по специальности «Языкознание и литературоведение»

Вычислительные технологии

Том 13, № 3, 2008

Using the F-measure as similarity measure for automatic text summarization

R. M. ÄLIGÜLIYEV Institute of Information Technology of the National Academy of Sciences of Azerbaijan, Baku e-mail: a.ramiz@science.az

Цель статьи — показать, что качество реферирования текстов непосредственно зависит от выбора меры подобия, и значит, использование F-меры как меры подобия является эффективным подходом. Эффективность выбора F-меры в качестве меры подобия подтверждена экспериментальным путем. Эксперименты показывают, что использование F-меры, с точки зрения точности реферирования, дает лучшие результаты, чем мера косинуса.

Introduction

Text mining is a research area that is currently extremely active. Automatic text summarization is one important task in this field. Automatic text summarization plays an important role in information retrieval (IR). The technology of automatic text summarization is maturing and may provide a solution to the information overload problem [1, 2]. With a large volume of texts, presenting the user with a summary of each document greatly facilitates the task of finding the desired documents. Similar to the tasks of IR, automatic text summarization can be regarded as how to find out salient pieces (here piece could be a phrase, sentence, or a paragraph of the document) from a document. Its goal is to include in that summary the most significant pieces in the text. In general, automatic text summarization takes an original text(s) as input, extracts the essence of the original text(s), and presents a well-formed summary to the user. Mani and Maybury [1] formally defined automatic text summarization as a process that produces a condensed version of its input's for user(s) consumption while preserving the main information content of source text(s).

1. Related Work

A variety of automatic text summarization methods have been proposed and evaluated. They can be broadly categorized into two approaches: abstraction and extraction. The goal of abstraction is to understand the text using knowledge-based methods and compose a coherent summary comparable to a human authored summary. This is very difficult to achieve with current natural language processing (NLP) techniques. In contrast to abstraction that requires heavy machinery from NLP, extraction can be easily viewed as the process of selecting salient excerpts from the source document [1]. Extraction systems analyze a

© Институт вычислительных технологий Сибирского отделения Российской академии наук, 2008.

source document using techniques derived from IR (e. g. frequency analysis and keyword identification) to determine significant sentences that constitute the summary.

A summary can also either be a user-oriented (or query-based) or generic [3]. A generic summary locates the main topics and key contents covered in the source text. A query-based summary locates the contents pertinent to user's seeking goals [2]. Query-based text summaries are useful for answering such questions as whether a given document is relevant to the user's query, and if relevant, which part(s) of the document is relevant. On the other hand, a generic summary provides an overall sense of the text's contents. A good generic summary should contain the main topics of the document while keeping redundancy to a minimum.

Sentence based summarization techniques are commonly used in automatic text summarization to produce extractive summaries [4-11]. The generic summarization methods that extract the most relevance sentences from the source document to form a summary in papers [7-10] are proposed. The proposed methods are based on clustering of sentences. Effective techniques for sentence extraction have been proposed in papers [4-6, 12]. The techniques first break a document into a list of sentences (paragraphs). Important sentences are then detected by some sentence weighting scheme, and the highly weighted sentences are selected to form a summary. A sentence weighting scheme can be variously formulated by employing many components and distributing them with different parameters. For example, Term Frequency, Sentence Order and Sentence Length are common components. The paper [13] focus on investigating and comparing effectiveness between Query Term Frequency (QTF) and Query Term Order (QTO). QTF in sentence weighting algorithm means the number of times the query terms appear in a sentence, and each term is equally weighted. QTO means the number of times the query terms appear in a sentence, with those terms appearing earlier in the query being assigned higher scores than those appearing later. Various criteria maybe used to associate importance with paragraphs, giving rise to different paths. To achieve automatic text summarization in paper [11] proposed two novel methods: modified corpus based approach (MCBA), and LSA-based TRM (Text Relationship Map) approach (LSA+TRM). The first is based on a score function combined with the analysis of salient features, and the genetic algorithm is employed to discover suitable combinations of feature weights. The second one exploits LSA and a TRM to derive semantically salient structures from a document. Both approaches concentrate on single-document summarization and generate extract-based summaries. The method TRM proposed by Salton et al. [12] is a graphical represent at, on of textual structure, in which paragraphs (in general, pieces of text) are represented by nodes on a graph and related paragraphs are linked by edges.

In this paper, we propose a simply and effective sentence extractive technique to achieve automatic text summarization. This method is based on evaluation of relevance score of sentence. The relevance score of each sentence is calculated in relation to all other sentences. We concentrate our presentation on choice a similarity measure.

2. Extractive Generic Summarization by Relevance Measure

Extractive summarization works by choosing a subset of the sentences in the original document. This process can be viewed as identifying the most salient sentences in a document, that give the necessary and sufficient amount of information related to the main theme of the document. To assess the importance of the sentences use a several similarity measures. One of the similarity measure widely used in text mining is the cosine measure. The cosine

similarity between two sentences Sj and Si is defined as:

cos(Si, Si) =

E

j=1

i, l = 1,..., n,

m m

E

j=i j=i

(1)

where Wj is the term weight of the tj in the sentence Sj = (w^, wj2,..., wjm), i =1, 2, j = 1, 2,..., m.

A typical weighting scheme is to use the frequency-based formula, such as [14]:

..., n,

w.

fjj log (—

(2)

where /¿j — the number of occurrences of term tj in the sentence Si and nj — the number of sentences containing the term tj.

In our paper to evaluate the importance of the sentences we use classical IR precision and recall measures. To calculate the similarity measure each sentence must at first be represented in a suitable form. In our method, a sentence Sj is represented as bag-of-terms, instead of the term-based frequency vector. Let a document D is decomposed into individual sentences D = (Si, S2,..., Sn), where n is the number of sentences in a document D. Let T = (t1,t2, ...,tm) represent all the terms occurred in a document D, where m is the number of terms. Let a sentence Sj is represented as bag-of-terms Sj = (t1,t2, ...,tmi), where mj is the number of terms in a sentence Sj.

The similarity between pair of sentences Sj and Si is evaluated to determine if they are semantically related. The similarity between sentences Sj and Si we define as

F (Sj,Sj) =

2P (Sj,Si)R(Sj,Sj) P (Sj, Si) + R(Sj,Si)

i = l = 1, 2,

n.

(3)

In formula (3) P(Sj,Si) and R(Sj,Si) are classical IR precision and recall measures, which we compute as follows:

P (Sj, Si)

R(Sj, Si)

|Sj n Si | = |Sj n Si | | Sj | mj :

|Sj n Si | = |Sj n Si| |Si | mi :

i=l i=l

1, 2,

1, 2,

n,

n,

(4)

(5)

where |A| is the cardinality of a set A.

In view of (4) and (5), the formula (3) becomes:

F (Sj, Si) = 2 , i,l

mj + mi

1, 2,..., n.

(6)

Our approach to text summarization allows generic summaries by scoring sentences. Each sentence is scored according to the formula (7). The relevance score of Sj with regard to all sentences in a documentD as (based on F-measure), we compute as:

Fscore(Sj) = £ F(Sj, Si), i = 1, 2,..., n.

(7)

i=1 i=j

n

Since the main purpose to show the effectiveness of the application of an F-measure as the similarity measure, that analogously we determine the relevance score of Si with regard to all sentences in a document D (based on cosine measure) [4-6]:

n

Cscore (Si) = y^cos(ffj,ffl), i = 1, 2, ...,n. (8)

l=1

l=i

Finally, as to selection of sentences to generate a summary all sentences are ranked according to their relevance scores calculated from formula (7) ((8)), and a designated number of top-weight sentences are picked out to form the summary.

Thus the generation summary process consist the following steps:

1. Decompose the document into individual sentences.

2. Represent each sentence as bag-of-terms.

3. Using the formulae described in (6) for each pair of sentences Si and Sl compute the similarity measure.

4. Using the formula (7) ((8)) for each sentence Si, compute the relevance score.

5. Rank all sentences according to their relevance score.

6. Starting with the sentence which has a highest relevance score the sentences add to the summary. If the compression rate (CR), which is defined as ratio of summary length to original length, reaches the predefined value, terminate the operation; otherwise, continue the process adding of the sentences to the summary.

The methods (7) and (8) are conditionally we call Method1 and Method2, respectively.

3. Experiments and discussion

In this section, we describe the experiment results to evaluate our text summarization algorithm. In our experiments, using human-generated and NewslnEssence-generated summaries, we employed four text-summarization methods — Method1, Method2, MS Word Summarizer and Copernic Summarizer [15]. The document collection used in this experiment consisted of fourteen documents, partitioned into two groups. The first group contained four documents (doc1.. .doc4), taken from http://oswinds.csd.auth.gr, www.actapress.com, and http://www.mitre.org. The second group contained ten news articles (news1.. .news10), were randomly selected from the NewslnEssence [16]. For first group we compared the summaries produced by methods Method1, Method2, MS Word Summarizer, and Copernic Summarizer against the human-generated summaries. For second group we compared the summaries produced by methods Method1, Method2, MS Word Summarizer, and Copernic Summarizer against the summaries produced by NewsInEssence Summarizer. These are an important point, since there is no standard measure of summary quality. To quote [2]: "Text summarization is still an emerging field, and serious questions remain concerning the appropriate methods and types of evaluation".

3.1. Preprocessing

Each document in the first group has been transformed to text format and the abstracts, keywords and references have been removed. One of the major problems in text mining is that a document can contain a very large number of words. If each of these words is represented as a vector coordinate, the number of dimensions would be too high for the text

mining algorithm. Hence, it is crucial to apply preprocessing methods that greatly reduce the number of dimensions (words) to be given to the text mining algorithm. Our system can apply several preprocessing methods to the original documents, namely stemming and removal stopwords.

Stemming (i.e. removing word affixes such as 'ing', 'ion', 's') consists of converting each word to its stem, i.e. a natural form with respect to tag-of-speech and verbal/plural inflections. In essence, to get the stem of a word it is necessary to eliminate its suffixes representing tag-of-speech and/or verbal/plural inflections. We have used Porter's algorithm [17], originally developed for the English language.

Stopwords (i.e. insignificant words like 'can', 'in', 'this', 'from', 'then', 'or', 'the', 'by') are words that occur very frequently in a document. Since they are so common in many documents, they carry very little information about the contents of a document in which they appear.

3.2. Comparison between human-generated and automatically-generated summaries

Four independent professional evaluators were employed to conduct manual summarization. For each document docl.. .doc4, each professional evaluator was requested to select 15 and 30 % sentences which (s)he deemed the most relevant for summarizing the document. Table 1 shows the statistics of the documents and the summarization results.

We employ the standard measures to evaluate the performance of summarization, i.e. precision, recall and F-measure. We assume that a human would be able to identify the most important sentences in a document most effectively. If the set of sentences selected by an automatic extraction method has a high overlap with the human-generated extract, the automatic method should be regarded as effective. Assume that Sman is the manual summary and Sauto is the automatically-generated summary, the measurements are defined as [14]:

D |Sman ^ Sauto1 /„•,

P =—is—|—, (9)

| sauto |

0 1 Sman ^ Sauto1 /1ft\

R =-is—i—' (10)

1 sman1

F = (11)

P + R y '

The evaluation results are shown in Tables 2 and 3. Tables 2 and 3 show a summary of precision (P), recall (R) and F-measure (F) for each system, when CR is 15 and 30%, respectively. The MS Word Summarizer reaches an average of 0.433 (0.511) P, 0.540 (0.538) R and 0.479 (0.524) F, when CR is 15% (30%). The Copernic Summarizer reaches an average of 0.514 (0.535) P, 0.540 (0.560) R and 0.527 (0.547) F, when CR is 15% (30%). Our approach using cosine measure (Method2) achieves an average of 0.530 (0.510) P, 0.560 (0.532) R and 0.544 (0.520) F, while using the F-measure as similarity measure achieves an average of 0.615 (0.614) P, 0.649 (0.666) R and 0.630 (0.652) F, when CR is 15% (30%). Interestingly, the Method1 gives the best results in both cases when CR is 15 and 30 %, than other methods. It can be observed that when F-measure is considered, on average, Method1 outperforms Method2 about 15.8% F and 25.4% F when CR is 15 and 30%, respectively.

Table 1. Statistics of the documents doc1.. .doc4 and summaries

Docu- Number of sentences in the

ments Docu- summaries created by summarizers

ments Human Method2 MS Word Copernic Method1

CR = CR = CR = CR = CR =

15% 30% 15% 30% 15% 30% 15% 30% 15% 30%

doc1 158 23 47 23 47 29 50 23 47 23 47

doc2 151 22 45 23 45 23 43 24 48 23 45

doc3 111 17 34 17 34 21 36 17 34 17 34

doc4 195 27 55 32 64 40 63 30 61 32 64

Table 2. Evaluation measures for automatic extraction methods, CR = 15%

Docu-

Overlap with the human-generated extracts

ments Method1 Method2 MS Word summarizer Copernic summarizer

P R F P R F P R F P R F

doc1 0.652 0.652 0.652 0.478 0.478 0.478 0.379 0.478 0.423 0.435 0.435 0.435

doc2 0.565 0.591 0.578 0.522 0.545 0.533 0.478 0.500 0.489 0.500 0.545 0.522

doc3 0.647 0.647 0.647 0.588 0.588 0.588 0.476 0.588 0.526 0.588 0.588 0.588

doc4 0.594 0.704 0.644 0.531 0.630 0.576 0.400 0.593 0.478 0.533 0.593 0.561

avg. 0.615 0.649 0.630 0.530 0.560 0.544 0.433 0.540 0.479 0.514 0.540 0.527

Table 3. Evaluation measures for automatic extraction methods, CR = 30%

Docu-

Overlap with the human-generated extracts

ments Method1 Method2 MS Word summarizer Copernic summarizer

P R F P R F P R F P R F

doc1 0.617 0.617 0.617 0.447 0.447 0.447 0.400 0.426 0.412 0.383 0.383 0.383

doc2 0.644 0.644 0.644 0.533 0.533 0.533 0.558 0.533 0.545 0.583 0.622 0.602

doc3 0.676 0.676 0.676 0.529 0.529 0.529 0.611 0.647 0.629 0.618 0.618 0.618

doc4 0.625 0.727 0.672 0.531 0.618 0.571 0.476 0.545 0.508 0.557 0.618 0.586

avg. 0.641 0.666 0.652 0.510 0.532 0.520 0.511 0.538 0.524 0.535 0.560 0.547

Table 4. Performance evaluation (F-measure) compared between Method1 and other methods

Docu- CR =15% CR = 30 %

ments Method2 MS Word Copernic Method2 MS Word Copernic

doc1 36.4 (+) 54.1 (+) 49.9 (+) 38.0 (+) 49.8 (+) 61.1 (+)

doc2 8.4 (+) 18.2 (+) 10.7 (+) 20.8 (+) 18.2 (+) 7.0 (+)

doc3 10.0 (+) 23.0 (+) 10.0 (+) 27.8 (+) 7.5 (+) 9.4 (+)

doc4 11.8 (+) 34.7 (+) 14.8 (+) 17.7 (+) 32.3 (+) 14.7 (+)

avg. 15.8 (+) 31.5 (+) 19.5 (+) 25.4 (+) 24.4 (+) 19.2 (+)

Hereafter, we use relative improvement

(Methodl — other methods)

100 for comparison.

other methods

Table 4 reports the performance results compared between Method1 and other methods.

3.3. Comparison between NewsInEssence-generated and automatically-generated summaries

We assume that the NewsInEssence summarizer would be able to identify the most relevant sentences in a document most effectively. For each news article newsl.. .news10 by NewsInEssence summarizer were created two summaries, at CR= 20 and CR= 30 %. Table 5 shows the statistics of the news articles and the summarization results. If the set of sentences selected by an automatic extraction method has a high overlap with the NewsInEssence-generated extract, the automatic method should be regarded as effective. Assume that SNIE is the NewsInEssence-generated summary and Sauto is the automatically generated summary, the measurements are defined as:

p = |SNIE H Sauto1 (12)

P = IS I ' (12)

I sauto |

|SNIE H Sauto1 /10s

R = —is—|—' (13)

1 sman1

F =iPR. (14)

P + R y J

Tables 6 and 7 show the evaluation results. When CR is 20 and 30 % the MS Word Summarizer reaches an average of 0.396 and 0.410 P, 0.436 and 0.444 R, 0.413 and 0.422 F, respectively. The Copernic Summarizer reaches an average of 0.479 (0.567) P, 0.425 (0.512) R and 0.447 (0.532) F, when CR is 20% (30%). The Method2 achieves an average of 0.376 (0.410) P, 0.360 (0.402) R and 0.367 (0.404) F, and the Method1 achieves an average of 0.512 and 0.575 P, 0.488 and 0.583 R, 0.498 and 0.575 F, when CR is 20 and 30%, respectively. It can be observed that when F-measure is considered, on average, Method1 outperforms Method2 about 35.7% F and 42.3 % F when CR is 20 and 30 %, respectively. Table 8 gives the performance results compared between Method1 and other methods. In the Tables 4 and 8 "+" means the result outperforms and "—" means the opposite.

Table 5. Statistics of the documents newsl.. .news10 and summaries

News Number of sentences in the

articles News summaries created by summarizers

articles NewsInEssence Method2 MS Word Copernic Method1

CR = CR = CR = CR = CR =

20% 30% 20% 30% 20% 30% 20% 30% 20% 30%

newsl 26 5 8 5 8 6 12 5 8 5 8

news2 36 7 10 7 10 9 12 7 10 7 10

news3 13 3 4 3 4 3 4 2 3 3 4

news4 13 3 5 3 4 3 5 2 3 3 4

news5 39 8 12 8 12 9 12 7 11 8 12

news6 36 7 10 7 11 9 13 7 10 7 11

news7 40 9 14 8 12 10 14 8 12 8 12

news8 25 5 7 5 8 7 10 5 8 5 8

news9 26 5 7 5 8 7 9 5 7 5 8

news10 26 7 11 5 8 5 8 5 7 5 8

Table 6. Evaluation measures for automatic extraction methods, CR = 20%

Overlap with the NewsInEssence-generated extracts

articles Method1 Method2 MS Word summarizer Copernic summarizer

P R F P R F P R F P R F

news1 0.800 0.800 0.800 0.400 0.400 0.400 0.500 0.600 0.545 0.400 0.400 0.400

news2 0.429 0.429 0.429 0.572 0.572 0.572 0.333 0.429 0.392 0.714 0.714 0.714

news3 0.667 0.667 0.667 0.333 0.333 0.333 0.667 0.667 0.667 0.500 0.333 0.400

news4 0.333 0.333 0.333 0.333 0.333 0.333 0.333 0.333 0.333 0.500 0.333 0.400

news5 0.375 0.375 0.375 0.375 0.375 0.375 0.333 0.375 0.353 0.286 0.250 0.267

news6 0.286 0.286 0.286 0.572 0.572 0.572 0.222 0.286 0.250 0.286 0.286 0.286

news7 0.625 0.556 0.589 0.375 0.333 0.353 0.400 0.444 0.421 0.500 0.444 0.470

news8 0.400 0.400 0.400 0.200 0.200 0.200 0.143 0.200 0.167 0.600 0.600 0.600

news9 0.600 0.600 0.600 0.200 0.200 0.200 0.429 0.600 0.500 0.600 0.600 0.600

news10 0.600 0.429 0.500 0.400 0.286 0.334 0.600 0.429 0.500 0.400 0.286 0.334

avg. 0.512 0.488 0.498 0.376 0.360 0.367 0.396 0.436 0.413 0.479 0.425 0.447

Table 7. Evaluation measures for automatic extraction methods, CR = 30%

Overlap with the NewsInEssence-generated extracts

articles Method1 Method2 MS Word summarizer Copernic summarizer

P R F P R F P R F P R F

news1 0.625 0.625 0.625 0.625 0.625 0.625 0.333 0.500 0.400 0.375 0.375 0.375

news2 0.400 0.400 0.400 0.400 0.400 0.400 0.417 0.500 0.455 0.500 0.500 0.500

news3 0.500 0.500 0.500 0.500 0.500 0.500 0.500 0.500 0.500 0.667 0.500 0.572

news4 0.600 0.750 0.667 0.500 0.400 0.444 0.400 0.400 0.400 0.667 0.400 0.500

news5 0.583 0.583 0.583 0.417 0.417 0.417 0.500 0.500 0.500 0.636 0.583 0.608

news6 0.455 0.500 0.476 0.364 0.400 0.381 0.231 0.300 0.261 0.400 0.400 0.400

news7 0.583 0.500 0.538 0.167 0.143 0.154 0.429 0.429 0.429 0.500 0.429 0.462

news8 0.625 0.714 0.667 0.250 0.286 0.267 0.111 0.143 0.125 0.500 0.572 0.534

news9 0.625 0.714 0.667 0.500 0.571 0.533 0.556 0.714 0.625 0.857 1.000 0.923

news10 0.750 0.545 0.631 0.375 0.273 0.316 0.625 0.455 0.527 0.571 0.364 0.446

avg. 0.575 0.583 0.575 0.410 0.402 0.404 0.410 0.444 0.422 0.567 0.512 0.532

Table 8. Performance evaluation (F-measure) compared between Method1 and other methods

News CR = 15%

articles Method2 MS Word Copernic

news1 100.0 (+) 46.8 (+) 100.0 (+)

news2 -25.0 (-) 9.4 (+) -39.9 (-)

news3 100.3 (+) 0.0 66.8 (+)

news4 0.0 0.0 -16.8 (-)

news5 0.0 6.2 (+) 40.4 (+)

news6 -50.0 (-) 14.4 (+) 0.0

news7 66.9 (+) 39.9 (+) 25.3 (+)

news8 100.0 (+) 139.5 (+) -33.3 (-)

news9 200.0 (+) 20.0 (+) 0.0

news10 49.7 (+) 0.0 49.7 (+)

avg. 35.7 (+) 20.6 (+) 11.4 (+)

CR = 30 %

Method2

MS Word

Copernic

0.0

56.2

+

66.7

+

0.0

-12.1

-20.0

0.0 0.(

50.2 (+) 66.8 (+

39.8 (+) 16.6 (+

24.9 (+) 82.4 (+

249.4 (+) 25.4 (+

149.8 (+) 433.6 (+

25.1 (+) 6.7 (+

99.7 (+) 19.7 (+

42.3 (+) 36.3 (+

-12.6

33.4

+

-4.1

19.0

+

16.5

+

24.9

+

-27.7

41.5

+

8.1

+

Conclusion

In this paper, we propose a practical approach for extracting the most relevant sentences from the original document to form a summary. The proposed text summarization method creates generic summaries by scoring and extracting sentences from the source documents. For sentence scoring most summarization systems use cosine measure as well a similarity measure. The idea of our approach is to exploit the classical IR precision and recall measures as similarity measure. In this paper, we show that using the classical IR precision and recall measures as similarity measure is a viable and effective technique. We provide experimental evidence that our approach achieves reasonable performance. The experimental result shows that the similarity measure may bias the score, and make the summarizer misjudge the importance of sentences. The effect of F-measure as similarity measure in text summarization is illustrated with an example shown in Tables 4 and 8. The experiments justify our assumption that the relevance score of a sentence directly depends on choice of similarity measure. We conclude that F-measure can be employed as similarity measure to promote text summarization.

References

[1] Mani I., Maybury M.T. Advances in automated text summarization. Cambridge: MIT Press, 1999. 442 p.

[2] Zhang Y., Zincir-Heywood N., Milios E. World Wide Web site summarization // International J. of Web Intelligence and Agents Systems. 2004. Vol. 2, N 1. P. 39-53.

[3] Afantenos S., Karkaletsis V., Stamatopoulos P. Summarization from medical documents: a survey // Artificial Intelligence in Medicine. 2005. Vol. 33, N 2. P. 157-177.

[4] Alguliev R.M., Aliguliyev R.M. Effective summarization method of text documents // Proc. of the 2005 IEEE/WIC/ACM Intern. Conf. on Web Intelligence. France: Compiegne,

2005. P. 264-271.

[5] Alguliev R.M., Aliguliyev R.M. Improvement of Text Information Classification by Definition Importance of Sentences // Information Technologies. 2006. N 3. P. 62-68 (in Russian).

[6] Alguliev R.M., Aliguliyev R.M. A new summarization method of text documents and evaluation of classification result in three aspects // Telecommunications. 2006. N 3. P. 7-16 (in Russian).

[7] Alguliev R.M., Aliguliyev R.M., Bagirov A.M. Global optimization in the summarization of text documents // Automatic Control and Computer Sci. N.Y.: Allerton Press, Inc. 2005. Vol. 39, N 6. P. 42-47.

[8] Alguliev R.M., Aliguliyev R.M. Summarization of text-based documents with a determination of latent topical sections and information-rich sentences // Automatic Control and Computer Sci. N.Y.: Allerton Press, Inc. 2007. Vol. 41, N 3. P. 132-140.

[9] Aliguliyev R.M. A novel partitioning-based clustering method and generic document summarization // Proc. of the 2006 IEEE/WIC/ACM Intern. Conf. on Web Intelligence and Intelligent Agent Technology (WI-IAT 2006 Workshops) (WI-IATW'06). China: Hong Kong,

2006. P. 626-629.

[10] Aliguliyev R.M. Automatic document summarization by sentence extraction // Comput. Technologies. 2007. Vol. 12, N 5. P. 5-15.

[11] Yeh J.-Y., Ke H.-R., Yang W.-P., Meng I.-H. Text summarization using a trainable summarizer and latent semantic analysis // Information Processing and Management. 2005. Vol. 41, N 1. P. 75-95.

[12] Salton G., Singhal A., Mitra M., Buckley C. Automatic text structuring and summarization // Information Processing and Management. 1997. Vol. 33, N 2. P. 193-207.

[13] Liang S.F., Devlin S., Tait J. Investigating sentence weighting components for automatic summarization // Information Processing and Management. 2006. Vol. 43, N 1. P. 146-153.

[14] Baeza-Yates R., Ribeiro-Neto R. Modern Information Retrieval. N.Y.: Addison Wesley. ACM Press, 1999. 513 p.

[15] :(CopernicSummarizer: http://www.copernic.com//en/products/summarizer)

[16] :(NewsInEssence Summary: http://lada.si.umich.edu)

[17] Porter M. An algorithm for suffix stripping // Program. 1980. Vol. 14, N 3. P. 130-137.

Received for publication 11 October 2007