KR100703193B1 - Apparatus for summarizing generic text summarization using non-negative matrix factorization and method therefor - Google Patents

Abstract

The present invention relates to a document summarization apparatus using non-negative matrix factorization, which is a preliminary step for summarizing a document, after decomposing the given document into respective sentences, removing unnecessary strings, extracting roots, and A preprocessor configured to summarize the document based on the non-negative matrix factorization by calculating weights for the preprocessing; And a document summarizer for summarizing the entire document by non-negative matrix factorization of matrix A; It includes.

According to the present invention, by using the non-negative semantic variable matrix extracted by using the non-negative matrix factorization, it is possible to accurately summarize the document contents.

Nonnegative matrix, factorization, document, summary, nonnegative semantic feature matrix, nonnegative semantic variable matrix

Description

Apparatus for summarizing generic text summarization using non-negative matrix factorization and method therefor}

1 is a block diagram showing a document summary device using non-negative matrix factorization according to an embodiment of the present invention.

2 is a flowchart illustrating a document summary method using non-negative matrix factorization according to an embodiment of the present invention.

3 is a detailed flowchart illustrating a non-negative matrix factorization step according to an embodiment of the present invention.

4 is a detailed flowchart illustrating a sentence selection step according to an embodiment of the present invention.

The present invention relates to an apparatus and method for document summary using non-negative matrix factorization, and more particularly, to comprehensively summarize the entire contents of a document automatically using a non-negative semantic variable matrix obtained by non-negative matrix factorization. The present invention relates to a document summary apparatus and method using non-negative matrix factorization.

Recently, due to the development of information and communication, the use of computer networks has increased, and the use of digitized electronic documents has spread, and the amount of documents that an individual must manage and check has been increasing. As a result, many computer users must spend more and more time identifying and managing documents related to individuals or their organizations. Therefore, there is an urgent need for tools for finding, classifying, and classifying documents related to individuals for effective information management.

On the other hand, the information retrieval system that can find the desired information from the vast amount of information is used, but in order to check in detail the search results presented by the user has a lot of information problems. Therefore, instead of the tendency to solve the problem by presenting only the introductory part of the document, it is possible to reduce the retrieval time of the information desired by the user. LSA, Latent Semantic Analysis (Document Summary), and 2) Document Topic (Topic).

First, the document summary method using latent semantic analysis uses the components of a singular vector to select a sentence, and the eigenvalues corresponding to the eigenvectors have positive and negative components. Since the ranking of the extracted sentences is composed of the component values of the small eigenvectors, the extracted sentences have a problem in that they depend only on the vector values without meaning.

In addition, the document summary method using the subject of the document summarizes the document by finding various subjects in the document, and then selects a sentence that matches each subject, and calculates the Saliency Scores from the sentences and the terms. The documents are summarized in groups (Topical Groups), which has the problem of significant cost of extraction.

SUMMARY OF THE INVENTION An object of the present invention is to solve the above-described problem, and to accurately summarize document contents by using a non-negative semantic variable matrix extracted using non-negative matrix factorization.

Another object of the present invention is to separate a large document into individual sentences, thereby facilitating document summaries using the separated sentences.

Another object of the present invention is to use a non-negative constraint similar to the human recognition process.

Another object of the present invention is a document summary by unsupervised learning, which does not require a learning sentence by a prior expert, and also makes it possible to easily extract a document through a low calculation cost.

In the document summarization apparatus using the non-negative matrix factorization according to the present invention for realizing such a technical idea, as a previous step for summarizing the document (D), after decomposing the given document into each sentence A preprocessing unit 120 capable of summarizing a document based on non-negative matrix factorization by removing unnecessary strings, extracting roots, and calculating weights for the sentences; And a document summary unit 130 for summarizing the entire document by factoring a matrix A into a non-negative matrix; It includes.

Preferably, the input unit 110 for receiving the setting signal of the document to be summarized and the document to be summarized from the user; And an output unit 140 for outputting the summary summary summarized through the document summary unit for the user to check. It characterized in that it further comprises.

Also preferably the preprocessing unit 120, sentence decomposition means 121 for decomposing the document into each of the sentences; Extraction sentence setting means (122) for setting the number of sentences to be extracted; Stopword removal means (123) for removing stopwords for each of the sentences decomposed through the sentence decomposition means; Root extracting means (124) for extracting roots with respect to the sentence; Term-frequency vector generating means (125) for generating a term-frequency vector according to the frequency of use of the term excluding the stop word with respect to the sentence; Weight calculation means (126) for calculating a weight for the sentence using the term-frequency vector; Term-sentence matrix generating means (127) for generating the matrix A consisting of the term and the sentence in the document; Characterized in that it comprises a.

Also preferably, the document summary unit 130 includes: factoring means (131) for factoring the matrix A using non-negative matrices W and H; Non-negative matrix calculation means (132) for calculating the non-negative matrices W and H; In the matrix H that contains the p-th row haengbek emitter H _p. The sentence vector A of the matrix A in the same column as the q column with the largest element value in _. sentence selection means 133 for selecting and extracting sentences corresponding to _q and adding them to the summary S; And document summary generating means for generating the summary of the document using the set of sentences selected by the sentence selecting means. Characterized in that it comprises a.

Also preferably, the weight is characterized in that it comprises a regional weight and a global weight.

Also preferably, the non-negative matrix factoring means 131 factorizes the matrix A using a non-negative semantic feature matrix W and a non-negative semantic variable matrix H,

Here, the matrix W is a matrix designated by the number of terms and semantic features, and the matrix H is a matrix factored by an approximation value from the matrix A.

Also preferably, the non-negative matrices W and H can is characterized in that the value by ∥A-WH∥ ² is minimized.

And preferably, the sentence selection means 133, the row vector H _{p .} The row vectors H _p in the order of the highest element sum of the elements _. Characterized in that the selection.

On the other hand, in the document summary method using the non-negative matrix factorization of the present invention, (a) decomposing and extracting the document (D) to be summarized into each sentence, and setting the number of sentences (k) to be summarized; (b) removing meaningless words from the sentence and extracting a root included in the sentence; (c) generating a term-frequency vector according to the frequency of use of the term in the sentence; (d) calculating weights for the sentences including local weights and global weights using the term-frequency vectors; (e) generating a term-sentence matrix (A) consisting of m terms and n sentences; (f) factoring the term-sentence matrix using a non-negative matrix; (g) a row vector in the matrix H _p H included in the p-th _row. Selecting a sentence corresponding to the sentence vector q _.q A of the matrix A in the column such as heat with the largest value element in; And (h) outputting a summary (S) of the document using the sentence selected in the step (g); It includes.

Also preferably, prior to the step (a), (a-1) receiving a setting signal of the number of sentences to be summarized and the unnecessary character strings and root types from the user to summarize and the documents to be summarized. ; It characterized in that it further comprises.

Also preferably, the step (f) comprises: (f-1) factoring using the non-negative semantic feature matrix W and the non-negative semantic variable matrix H; (f-2) wherein the number of non-negative matrices W and H is determined that the minimum by ∥A-WH∥ ^2; And (f-3) calculating the values of the non-negative matrices W and H when they are minimized as a result of the determination in the step (f-2), and updating them at the same time; Characterized in that it comprises a.

Also preferably, the step (g) may include (g-1) the row vector H _{p .} Calculating a sum of elements of each row vector; (g-2) k row vectors H _{p .} Selecting and extracting; (g-3) After determining whether the selected and extracted sentences are the number of sentences to be summarized (k), in each of the k row vectors selected in the case of k, the matrix A in the same column as the q column having the largest element value in the row. Sentence vector A _.q Selecting a sentence corresponding to; And (g-4) calculating the similarity between the column vectors of the non-negative semantic feature matrix W and the subject when the selected number is less than k, as a result of the determination in the step (g-3), and the column vector W having a high similarity (W). _{Searching for p} ) and repeatedly selecting and extracting sentences; Characterized in that it comprises a.

Preferably, the step (g-4) is repeated as many times as the number (k) of the decomposition-extracted sentences.

The features and advantages of the present invention will become more apparent from the following detailed description based on the accompanying drawings. Prior to this, the terms or words used in the present specification and claims are defined in the technical spirit of the present invention on the basis of the principle that the inventor can appropriately define the concept of the term in order to explain his invention in the best way. It should be interpreted to mean meanings and concepts. In addition, when it is determined that the detailed description of the known function and its configuration related to the present invention may unnecessarily obscure the subject matter of the present invention, it should be noted that the detailed description is omitted.

Hereinafter, the overall technical spirit of the present invention with reference to the drawings.

A document summarizing apparatus using non-negative matrix factorization according to an embodiment of the present invention will be described with reference to FIG. 1.

1 is a block diagram showing a document summary device using non-negative matrix factorization according to an embodiment of the present invention.

First, referring to FIG. 1, the document summarizing apparatus 100 capable of summarizing a general document D using a non-negative semantic variable matrix by extracting sentences based on non-negative matrix factorization may include an input unit. 110, a preprocessing unit 120, a document summary unit 130, and an output unit 140.

First, the input unit 110 performs a function of receiving a setting signal of the document D to be summarized and the document D to be summarized from the user side.

In this embodiment, the setting signal of the document D to be summarized is set to include the number of sentences to be summarized, an unnecessary character string, a kind of root, and the like, but the present invention is not limited thereto.

As a preliminary step for summarizing a general document (D), the preprocessor 120 decomposes a given document (D) into individual sentences, removes unnecessary strings, extracts roots, By calculating the weight for the function, it is possible to summarize the document based on the non-negative matrix factorization bar, sentence decomposition means 121, extraction sentence setting means 122, stop word removal means ( 123, root extracting means 124, term-frequency vector generating means 125, weight calculating means 126, and term-sentence matrix generating means 127.

First, the sentence decomposing means 121 performs a function of decomposing a given document D into each sentence.

Further, the extraction sentence setting unit 122 performs a function of setting the number of sentences to be extracted in the document D. In this embodiment, the number of extraction sentences is set to k, but the present invention is not limited thereto. .

In addition, the stop word removing means 123 performs a function of removing the stop word through the stop word list A for each sentence decomposed through the sentence decomposition means 121.

At this time, in the present embodiment, although Rijsbergen is set as the stop word list A, the present invention is not limited thereto.

For reference, stop-words or noise world refers to words that are ignored during search or words that search engines exclude from the index when building the database. For example, '~', ~ in Korean Verbs, including, in, surveys, suffixes, conjunctions, endings, etc., and in English, including a, an, is, are, if, for, but, this, may, There are modal verbs, personal pronouns, instructional pronouns, and prepositions.

The root extracting means 124 also performs a function of extracting roots for each sentence.

For reference, the root is a morpheme representing the actual meaning of the word among the morphemes constituting the word, and is a real real morpheme synthesized in the case of a compound word, and the rest except the affix in the case of a derivative word. For example, the roots include the cover, the cover, the person, the person, and the generous.

In addition, the term-frequency vector generating means 125 performs a function for generating a term-frequency vector for each sentence in accordance with the frequency of use of the term except the stopword. In this case, the generated term-frequency vector is shown in Equation 1 below.

T _i = [t _1i , t _2i , ..., t _ni ] ^T

In the above formula, the term-frequency vector T _i represents the frequency of use of the term for the i th sentence, and t _ij is An element of T _i , which indicates the frequency of the j th term appearing in the i th sentence.

In addition, the weight calculation means 126 calculates a weight including a local weight and a global weight for the sentence using the generated term-frequency vector.

More specifically, the regional weight L (t _ji ) represents the j th term in the i th sentence, and the global weight G (t _ji ) represents the j th term throughout the document D, 2 and (3).

L (i) = tf (i)

G (i) = log (N / n (i))

In this case, tf (i) represents the frequency of appearance of the i-th term, N represents the total number of sentences in the document D, and n (i) represents the number of sentences including the i-th term.

On the other hand, the weight for each sentence including the local weight and the global weight is expressed by the following Equation 4 using Equation 2 and Equation 3 described above.

a _ji = L (t _ji ) · G (t _ji )

In this case, a _i is an element of the i th sentence A _i represented by Equation 5 below.

A _i = [a _1i , a _2i , ..., a _ni ] ^T

The term-sentence matrix generating means 127 performs a function of generating a matrix A consisting of m × n consisting of m terms and n sentences in total in the document D.

In this case, the matrix A (m × n) is composed of a component A _ij, A _ij element Is the weighted frequency of the i th sentence and the j th term.

In addition, the document summary unit 130 summarizes the entire document D by non-negative matrix factorization (NMF) of the matrix A generated through the term-statement matrix generating means 127. The non-negative matrix factoring means 131, the non-negative matrix calculating means 132, the sentence selecting means 133, and the document summary generating means 134 are included.

The nonnegative matrix factoring means 131 receives the matrix A using a non-negative semantic feature matrix (NSFM) W and a non-negative semantic variable matrix (NSVM). It decomposes and is expressed as Equation 6 below.

At this time, A is an m × n matrix, W is an m × r matrix, and H represents a matrix factored by an approximation from the matrix A. Where r is specified as the number of semantic features.

And the non-negative matrix calculation means 132 performs a function of calculating the non-negative matrices W and H using the following equations (7) and (8).

At this time, the non-negative matrices W and H are calculated and updated at the same time until the value of | A-WH | ² becomes minimum.

For reference, the column vector A _{· j} of the matrix A forms a linear combination with the column vector W _{· l} and the element H _kj of the semantic variable matrix H, which is expressed by Equation 9 below.

In this case, the column vector A _{· j} is a column vector of the j th sentence of the matrix A, and the column vector W _{· l} represents the l th semantic feature column vector of the matrix W.

In addition, the sentence selecting means 133 is a row vector H _{p .} Select / extract the sentence corresponding to the sentence vector A _.q of the matrix A in the same column as the q column with the largest element value in, and add it to the summary.

More specifically, row vector H _{p .} The sum of the elements of is calculated for each row vector using Equation 10 below.

In this case, k row vectors H _{p .} Select.

This sentence selection / extraction is repeated as many times as the number of sentences k to be summarized, and if the selected number is less than k, the next-highest similarity column vector W _{· p} is found and the sentence is repeatedly selected / extracted.

In this case, the similarity between the column vectors of the non-negative semantic feature matrix W and the subject is calculated to select a column vector W _{· p} having a high p-th similarity, but the present invention is not limited thereto.

The document summary generating means 134 generates a summary S of the document D using the set of sentences selected by the sentence selecting means 133.

Here, since the semantic feature used for sentence extraction has a non-negative value, it is more accurate in document summary than latent meaning analysis.

For reference, every semantic variable can represent each sentence, which intuitively makes each sentence more meaningful with a small subset associated with a broadly arranged topic rather than just one or all topics. Thus, the characteristics of each semantic variable are clustered into terms that are semantically related by non-negative matrix factorization. Thus, semantically related clusters are combined into semantic features, using nonnegative matrix factorization to distinguish homonyms in context.

In addition, the output unit 140 performs a function of outputting the summary S through the document summary unit 130 so that the user can check it. In this embodiment, the output unit 140 is set to include a monitor or a printer. However, the present invention is not limited to this.

In this way, a nonnegative matrix factorization statement can be extracted to summarize the bulk of the document.

Hereinafter, a document summary method using non-negative matrix factorization through application software having the above-described configuration will be described with reference to FIGS. 2 to 4.

2 is a flowchart illustrating a document summary method using non-negative matrix factorization according to an embodiment of the present invention, and FIG. 3 is a detailed flowchart illustrating a non-negative matrix factorization step according to an embodiment of the present invention. 4 is a detailed flowchart illustrating a sentence selection step according to an embodiment of the present invention.

First, as shown in FIG. 2, the preprocessor 120 decomposes / extracts the document D, which is the summary object, into each sentence (S10).

At this time, the preprocessing unit receives a setting signal for the document to be summarized and the document to be summarized from the user. In this case, the setting signal is set to include the number of sentences to be summarized, unnecessary strings, types of roots, and the like. It is not limited.

The preprocessor 120 sets the number of sentences to be summarized using the decomposed / extracted sentences (S20).

Here, although the number of sentences to be summarized is set to k, the present invention is not limited thereto.

Next, the preprocessing unit 120 removes a meaningless meaning word from each of the decomposed sentences, and extracts a root included in each sentence (S30).

The preprocessor 120 generates a term-frequency vector according to the frequency of use of the term using the decomposed sentence as shown in Equation 1 (S40).

The preprocessor 120 calculates a weight for each decomposed sentence including the regional weight and the global weight using the term-frequency vector using Equation 4 (S50).

The preprocessor 120 generates the term-sentence matrix A consisting of m terms and n sentences but reflecting the weight calculated through Equation 4 (S60).

Next, the document summary unit 130 factorizes the generated term-phrase matrix A using a non-negative matrix (S70).

More specifically, as shown in FIG. 3, after factoring using the non-negative semantic feature matrix W and the non-negative semantic variable matrix H in step S70 (S71), the document summary unit 130 determines the non-negative matrix. determines the values of the W and H by ∥A-WH∥ ² that the minimum (S72).

As a result of the determination in the step S72, in the case of minimization, the document summary unit 130 calculates the values of the non-negative matrices W and H (S73) and updates them at the same time (S74).

On the other hand, when it is determined that the step S72 is not minimized, the document summarizing unit 130 allows the non-negative matrices W and H to be minimized using Equations 7 and 8 (S75).

Next, the document summary unit 130 includes the row vector H _{p .} Selects the sentence corresponding to the sentence vector q _.q A of the matrix A in the column such as heat with the largest value element in the (S80).

More specifically, referring to FIG. 4, in step S80, the document summary unit 130 has a row vector H _{p .} The sum of elements is calculated for each row vector by using Equation 10 (S81).

At this time, the document summary unit 130 k k row vectors H _{p .} Select / extract (S82).

After determining whether the selected / extracted sentence is the number k of sentences to be summarized (S83), in case of k, the document summary unit 130 is in the same column as the q column having the largest element value in the row in each of the k selected row vectors. A sentence corresponding to the sentence vector A. _Q of the matrix A is selected (S84).

Using the selected sentence, the document summary unit 130 outputs the summary S of the document D (S90).

If the selected number is less than k, the document summary unit 130 finds a column vector W _{· p} having a high similarity between the subjects and non-negative semantic feature matrix W and repeatedly selects / extracts a sentence ( S85).

[Experimental Example]

To verify the performance of the document summary method proposed in the present invention, 129 articles were randomly extracted and used as experimental data. For performance evaluation, we used precision, recall, and F-measure which are mainly used in document summaries.

At this time, the experimental data used for the actual evaluation were manually summarized by three evaluators, and 81 documents including the sentences commonly selected among two or more of the evaluators each selected 3.8 sentences were used.

As a measure of performance evaluation, the following equation was used.

<Reproduction rate>

<Accuracy rate>

<F-measure>

Here, S _man is a sentence selected by the method proposed by each person, and S _sum is a sentence selected by the method proposed by the present invention. The following table shows the results of comparing the method of summarizing the document using the latent meaning analysis method and the method proposed by the present invention.

division LSA NMF Accuracy 0.445 0.700 Recall 0.340 0.650 F-measure 0.405 0.680

According to the F value of Table 1, the nonnegative matrix factorization (NMF) summarization method according to the present invention is about 0.275 better than the method using the latent significance analysis (LSA).

As shown in the experiment, the proposed method outperforms the potential semantic analysis method because the proposed method processes the document in a similar process to human recognition process using non-negative values and partial information. In addition, as shown in Equations 6, 7, and 8, the sentence can be extracted at a low cost.

As described above and described with reference to a preferred embodiment for illustrating the technical idea of the present invention, the present invention is not limited to the configuration and operation as shown and described as described above, it is a deviation from the scope of the technical idea It will be understood by those skilled in the art that many modifications and variations can be made to the invention without departing from the scope of the invention. Accordingly, all such suitable changes and modifications and equivalents should be considered to be within the scope of the present invention.

According to the present invention as described above by using the non-negative semantic variable matrix extracted by using the non-negative matrix factorization, it is possible to accurately summarize the document content.

In addition, as a document summary by non-supervised learning using non-negative constraints similar to human recognition process, there is no need for learning sentences by experts in advance, and its peculiar effect that can easily extract documents with low computational cost. have.

Claims (13)

In a document summary device using non-negative matrix factorization, As a previous step to summarize the document (D), it is based on non-negative matrix factorization by decomposing the given document into individual sentences, then removing unnecessary strings, extracting roots, and calculating weights for the sentences. A preprocessor 120 to summarize the document as; And A document summary unit 130 for summarizing the entire document by factoring a matrix A into a non-negative matrix; Document summarizing apparatus using a non-negative matrix factorization comprising a. The method of claim 1, An input unit (110) for receiving a setting signal of the document to be summarized and the document to be summarized from a user; And An output unit 140 for outputting the summary summarized through the document summary unit for the user to check; Document summarization apparatus using a non-negative matrix factorization further comprises. The method of claim 1, The preprocessing unit 120, Sentence decomposition means (121) for decomposing the document into the respective sentences; Extraction sentence setting means (122) for setting the number of sentences to be extracted; Stopword removal means (123) for removing stopwords for each of the sentences decomposed through the sentence decomposition means; Root extracting means (124) for extracting roots with respect to the sentence; Term-frequency vector generating means (125) for generating a term-frequency vector according to the frequency of use of the term excluding the stop word with respect to the sentence; Weight calculation means (126) for calculating a weight for the sentence using the term-frequency vector; And Term-sentence matrix generating means (127) for generating the matrix A consisting of the term and the sentence in the document; Document summarizing apparatus using a non-negative matrix factorization comprising a. The method of claim 1, The document summary unit 130, Factoring means (131) for factoring said matrix A using non-negative matrices W and H; Non-negative matrix calculation means (132) for calculating the non-negative matrices W and H; In the matrix H that contains the p-th row row vector H _p. Sentence selection means 133 for selecting and extracting a sentence corresponding to the sentence vector A _.q of the matrix A in the same column as the q column having the largest element value in and adding it to the summary S; And Document summary generating means (134) for generating the summary of the document by using the set of sentences selected by the sentence selecting means; Document summarizing apparatus using a non-negative matrix factorization comprising a. The method according to claim 1 or 3, The weight is, Document summarization apparatus using non-negative matrix factorization comprising local weights and global weights. The method of claim 4, wherein The non-negative matrix factoring means 131, Factorize the matrix A using a non-negative semantic feature matrix W and a non-negative semantic variable matrix H, Wherein the matrix W is a matrix designated by the number of terms and semantic features, and the matrix H is a matrix factored by an approximation from the matrix A. The method of claim 4, wherein The non-negative matrices W and H are A-WH Document summarization apparatus using non-negative matrix factorization, characterized in that the value by ² is minimized. The method of claim 4, wherein The sentence selection means 133, The row vector H _{p .} The row vectors H _p in the order of the highest element sum of the elements _. Document summarization apparatus using non-negative matrix factorization, characterized in that for selecting the. In the document summary method using non-negative matrix factorization, (a) decomposing and extracting the document (D) to be summarized into each sentence, and setting the number of sentences (k) to be summarized; (b) removing meaningless words from the sentence and extracting a root included in the sentence; (c) generating a term-frequency vector according to the frequency of use of the term in the sentence; (d) calculating weights for the sentences including local weights and global weights using the term-frequency vectors; (e) generating a term-sentence matrix (A) consisting of m terms and n sentences; (f) factoring the term-sentence matrix using a non-negative matrix; (g) a row vector in the matrix H _p H included in the p-th _row. Selecting a sentence corresponding to the sentence vector q _.q A of the matrix A in the column such as heat with the largest value element in; And (h) outputting a summary (S) of the document using the sentence selected in step (g); Document summary method using a non-negative matrix factorization comprising a. The method of claim 9, Before step (a), (a-1) receiving, from a user, a setting signal of the number of sentences to be summarized, the number of sentences to be summarized, and unnecessary character strings and root types of the documents to be summarized; Document summarization method using a non-negative matrix factorization further comprises. The method of claim 9, Step (f), (f-1) factoring using the non-negative semantic feature matrix W and the non-negative semantic variable matrix H; (f-2) wherein the number of non-negative matrices W and H is determined that the minimum by ∥A-WH∥ ^2; And (f-3) calculating the values of the non-negative matrices W and H when they are minimized as a result of the determination in the step (f-2) and updating them simultaneously; Document summary method using a non-negative matrix factorization comprising a. The method of claim 9, The step (g) is (g-1) the row vector H _{p .} Calculating a sum of elements of each row vector; (g-2) k row vectors H _{p .} Selecting and extracting; (g-3) After determining whether the selected and extracted sentences are the number of sentences to be summarized (k), in each of the k row vectors selected in the case of k, the matrix A in the same column as the q column having the largest element value in the row. Selecting a sentence corresponding to the sentence vector A _{· q of the} at least one sentence; And (g-4) As a result of the determination in the step (g-3), when the selected number is less than k, the similarity between the column vectors of the non-negative semantic feature matrix W and the subject is calculated, and a column vector having a high similarity (W _· searching for _p ) and repeatedly selecting and extracting sentences; Document summary method using a non-negative matrix factorization comprising a. The method of claim 12, The (g-4) step, The document summarization method using non-negative matrix factorization, characterized in that it is repeatedly performed as the number (k) of the extracted sentences.

Images