JP2003108571A - Document abstraction apparatus, control method of document abstraction apparatus, control program of document abstraction apparatus, and recording medium - Google Patents

Abstract

(57) [Summary] [Problem] To create a more accurate summary corresponding to each field even when the frequency of occurrence of an important word in each field is low in a sentence. A sentence vector generation unit generates a sentence vector of a sentence constituting a sentence to be summarized and outputs the sentence vector to a sentence vector generation unit and an important sentence generation unit. The sentence vector generation unit 15 generates a sentence vector corresponding to the sentence to be summarized based on the sentence vector, and outputs the sentence vector to the sentence vector comparison unit 16. The sentence vector comparison unit 16 extracts an important sentence from the sentence to be summarized based on the sentence vector, the feature vector for each field, and the sentence vector. As a result, the summary sentence output unit 13 generates and outputs a summary sentence from the important sentences.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a document summarizing device, a control method for the document summarizing device, a control program for the document summarizing device, and a recording medium, and more particularly to a document summarizing device capable of easily creating an accurate summarization. , Control method of document summarizing device,
The present invention relates to a control program for a document summarizing device and a recording medium.

[0002]

2. Description of the Related Art Conventionally, various proposals have been made for automatically creating a summary by inputting a document to be summarized and using an information processing device such as a computer. For example, the summarizing device disclosed in Japanese Patent No. 2944346 generates a feature vector of a sentence based on a feature vector of a word in a document to be summarized, and uses the similarity between the sentence appearing in the sentence and the feature vector. The summary is adopted to create a summary. The document summarizing device disclosed in Japanese Patent Laid-Open No. 11-102372 calculates the similarity of context vectors,
It adopts a configuration that creates a summary based on the similarity of the obtained context vectors.

[0003]

In the above conventional technique, a feature vector or a context vector is created based on the frequency information of words appearing in a text, and a summary text is created. Therefore, even if an important word is included in the field to which the sentence belongs (for example, engineering, economics, machinery, electricity, finance, etc.), it is likely that the important word will occur at a low frequency and the summary sentence There is a possibility that it will not be selected as a word to be used for, and eventually a sentence including the word will not be used as a summary sentence. As a result, there is a problem that an incorrect summary sentence is created. Therefore, an object of the present invention is to provide a document summarizing device and a document capable of creating a more accurate summary corresponding to each field even when the frequency of appearance of an important word in each field in a sentence is low. An object of the present invention is to provide a control method for a summarizing device, a control program for a document summarizing device, and a recording medium.

[0004]

In order to solve the above-mentioned problems, a document summarization apparatus includes a sentence vector generation unit that generates a sentence vector of a sentence that constitutes a sentence to be summarized, and a summarization target of the summarization target based on the sentence vector. A sentence vector generation unit that generates a sentence vector corresponding to a sentence, the sentence vector, an area-specific feature vector that characterizes a field to which the sentence to be summarized belongs, and an important sentence from the sentence to be summarized based on the sentence vector. And an important sentence extracting unit for extracting. According to the above configuration, the sentence vector generation unit generates the sentence vector of the sentence that constitutes the sentence to be summarized, and outputs the sentence vector to the sentence vector generation unit and the important sentence generation unit. The sentence vector generation unit generates a sentence vector corresponding to the sentence to be summarized based on the sentence vector and outputs it to the important sentence generation unit. The important sentence generator extracts an important sentence from the sentence to be summarized based on the sentence vector, the field-specific feature vector, and the sentence vector.

In this case, the sentence vector generation unit may set an average vector of sentence vectors of sentences constituting the sentence to be summarized as the sentence vector. Also, the important sentence extraction unit calculates a first inner product which is an inner product of the sentence vector and the sentence, and calculates a second inner product which is an inner product of the sentence vector and the field-specific feature vector, The important sentence may be extracted by using the sum of the first inner product and the second inner product as the similarity and comparing the similarity with a predetermined reference similarity. Further, a summary sentence creating unit that creates a summary sentence based on the extracted important sentence may be provided. Furthermore, a field-specific feature vector generation unit that generates the field-specific feature vector based on a plurality of learning sentences for each field may be provided. Further, the field-specific feature vector generation unit generates a sentence vector corresponding to each of the learning sentences, and an average vector of the sentence vectors corresponding to the plurality of learning sentences is generated as the field-specific feature vector. May be.

In addition, the control method of the document summarizing device includes a sentence vector generation process for generating a sentence vector of a sentence constituting a sentence to be summarized, and a sentence vector corresponding to the sentence to be summarized based on the sentence vector. A sentence vector generation process, the sentence vector, an important sentence extraction process of extracting an important sentence from the summary target sentence based on the field-specific feature vector characterizing the field to which the summary target sentence belongs and the sentence vector; It is characterized by having. In this case, the sentence vector generation process is
A step of setting an average vector of sentence vectors of sentences constituting the sentence to be summarized as the sentence vector may be provided.

Further, in the important sentence extracting step, a step of calculating a first inner product which is an inner product of the sentence vector and the sentence, and a second inner product which is an inner product of the sentence vector and the feature-specific vector A step of calculating the sum of the first inner product and the second inner product, and a step of extracting the important sentence by comparing the similarity with a predetermined reference similarity. May be provided. Furthermore, a summary sentence creating step of creating a summary sentence based on the extracted important sentence may be provided.

A control program of the document summarizing device for causing the computer to function as a document summarizing device for generating the summarized sentence data based on the input summarizing target sentence data corresponds to the summarizing target sentence data. Generate a sentence vector of the sentences that make up the sentence to be summarized,
A sentence vector corresponding to the sentence to be summarized is generated based on the sentence vector, and the sentence vector, the feature vector for each field that characterizes the field to which the sentence to be summarized belongs, and the sentence to be summarized based on the sentence vector It is characterized by extracting important sentences from. In this case, the average vector of the sentence vectors of the sentences constituting the sentence to be summarized may be calculated and used as the sentence vector.

Further, a first inner product which is an inner product of the sentence vector and the sentence is calculated, and a second inner product which is an inner product of the sentence vector and the feature vector for each field is calculated, and the first inner product is calculated. The important sentence may be extracted by setting the sum of the second inner products as the similarity and comparing the similarity with a predetermined reference similarity. Further, a summary sentence may be created based on the extracted important sentence. Further, the control program of each document summarizing device may be recorded in a recording medium.

[0010]

BEST MODE FOR CARRYING OUT THE INVENTION Next, preferred embodiments of the present invention will be described with reference to the drawings. FIG. 1 shows a schematic functional configuration block diagram of the document summarizing apparatus of the embodiment. The document summarization device 10 is roughly classified into a morpheme analysis unit 11, a sentence analysis unit 12, a summary sentence output unit 13, and a sentence vector generation unit 14.
A sentence vector generation unit 15, a field-specific feature vector generation unit 17, a sentence vector comparison unit 16, and a field-specific feature vector dictionary 18. Here, the document summarization device 10 can be realized in a computer system, and includes a morpheme analysis unit 11, a sentence analysis unit 12, a summary sentence output unit 13, a sentence vector generation unit 14, and a sentence vector generation unit 1.
5. The functions of the sentence vector comparison unit 16 and the field-specific feature vector generation unit 17 are realized by a program executable by a microprocessor corresponding to each unit. Further, such a program may be directly executed from a recording medium such as a semiconductor memory or a CD-ROM. It is also possible to install the program in an external storage device in advance and execute it. Furthermore, the program may be installed each time it is executed prior to execution or only once at the beginning via a network such as the Internet.

The input of the text to be summarized is direct input using a keyboard, tablet, etc., input from a storage device such as a flexible disk, a hard disk, character recognition input using a scanner and OCR, the Internet, LAN, etc. It is performed by inputting communication via the network. Furthermore, the output of the obtained summary is
Output is performed by a display device such as a CRT, print output by a printer or the like, recording output to a storage device such as a flexible disk or a hard disk, or communication output via a network such as a LAN. The field-specific feature vector dictionary is actually provided as a database on a storage device such as a hard disk connected to the computer system, the Internet, or LA.
It is configured on a storage device such as a hard disk connected to a database server or the like via a network such as N. The morpheme analysis unit 11 performs morpheme analysis on the input summary target sentence, divides it into word units, and extracts nouns, adjective verbs, and the like required for generating sentence vectors.

The sentence analysis unit 12 is a sentence vector generation unit 1.
4, sentence vector generation unit 15 and sentence vector comparison unit 1
Among the plurality of sentences that compose the summary target sentence that is controlled by inputting 6, the sentence that is considered to have a high degree of importance in the field to which the summary subject sentence belongs (hereinafter referred to as important sentence) is output to the summary sentence output unit 13. It will be output. Sentence vector generator 1
Reference numeral 4 vectorizes each sentence that constitutes the summary target sentence, and outputs the generated sentence vector to the sentence vector generation unit 15. The text vector generation unit 15 generates a text vector of the summary target document based on the input text vector. Then, the sentence vector comparison unit 16 functions as an important sentence extraction unit, and is input based on the input sentence vector, the generated sentence vector, and the feature vector of the field to which the summary target document belongs read from the field-specific feature vector dictionary. The important sentence in the field to which the relevant summary sentence belongs is extracted from the plurality of sentences constituting the relevant summary sentence and the summary sentence output unit 1
Output to 3.

The field-specific feature vector generating section 17 is provided with a field-specific feature vector dictionary 1 prior to the operation of the document summarizing apparatus 10.
8 is to generate a field-specific feature vector to be stored in the field 8. In this case, the operation manager of the document summarization device 10 determines that a predetermined number of learning sentences, which are sentences that are known to be included in the field to which the summarization target sentence that uses the document summarization device 10 belongs, for each field. prepare. In this case, the number of sentences to be prepared is empirically determined, and is a sufficient number so that the characteristics of each field are statistically significant. Next, a sentence vector is generated from the learning sentence. As a specific vectorization method, for example, there is a method of using a vector having an appearing word as a dimension.
The value of the element of the vector is a TFIDF value (a value calculated from the appearance frequency of the word and the feature amount of the word in the sentence)
To use.

For example, the vector D1 (d1) of the sentence d1 in the database storing a plurality of sentences can be expressed by the following equation. D1 (d1) = (TF (d1, t1) * IDF (t1), TF (d1, t2) * IDF (t2), TF (d1, t3) * IDF (t3), ......, TF (d1, tn ) * IDF (tn)) where t1, t2, t3, ..., Tn are words that appear in the database, and t1 to tn correspond to all words in the database. Then, for each field, for example, the average of vector values of learning sentences is calculated, and the average vector corresponding to the obtained average vector value is stored in the field-specific feature vector dictionary 18 as the field-specific feature vector of each field. Specifically, as shown in FIG. 2, a plurality of sentence vectors corresponding to a plurality of learning sentences belonging to the field A are obtained. Similarly, a plurality of sentence vectors corresponding to a plurality of learning sentences belonging to the field B are obtained. Note that, in FIG. 2, a plurality of text vectors in each field are two-dimensionally represented for simplification of the drawing, and are illustrated as a rectangle including a code (A or B) indicating a field to which each text vector belongs. There is. However, in reality, each sentence vector is the T of the first word.
It is represented as an N-dimensional vector with the FIDF value as the first axis, the TFIDF value of the second word as the second axis, ..., And the TFIDF value of the Nth word as the Nth axis. Here, N corresponds to the number of words that appear in all target documents. Then, the obtained plurality of text vectors are averaged for each field, and the obtained average vector is set as a field-specific feature vector of each field. In FIG. 2, the field-specific feature vector VCA of the field A and the field-specific feature vector VCB of the field B are displayed.

When actually storing the feature vector of each field in the field-specific feature vector dictionary 18 as a database, as shown in FIG. 3, field ID data for specifying the field and the field ID data are specified for each field. Word T for field
It is configured as a database including n pieces of FIDF data in descending order of TFIDF value (in descending order of appearance frequency). More specifically, for the field of field ID data = 1, in order from the word “PC”, which is the word with the highest TFIDF value in the field, and the corresponding TFIDF data representing the value of “0.0001” of the TFIDF,
Up to the word "software", n word TFIDF data are stored. Similarly, for the field of field ID data = 2, the word “send” is sequentially provided from the TMAIL data having the highest TFIDF value in the field, “mail” and the corresponding TFIDF value “0.00015”. Up to n words TFIDF data are stored.

Next, the operation of the sentence summarizing device will be described focusing on the operation of the sentence analyzing unit with reference to the processing flow chart of the sentence analyzing unit of FIG. The morphological analysis unit 11 morphologically analyzes the input summary target sentence, divides it into word units, extracts nouns, adjective verbs, and the like required for generating a sentence vector, and outputs them to the sentence analysis unit 12. Sentence analysis unit 12
Retrieves the sentences that make up the input sentence to be summarized (step S11). Then, it is determined whether or not all the sentences forming the summary target sentence have been extracted (step S1).
2). In the determination in step S12, if all the sentences that form the summary target sentence have not yet been extracted (step S12; No), the extracted sentence is sent to the sentence vector generation unit 14 to generate a sentence vector.

Here, the sentence vector generation operation of the sentence vector generation unit 14 will be described with reference to the processing flow chart of the sentence vector generation unit of FIG. Sentence vector generation unit 14
Acquires a noun, adjective verb, etc., required for generating a sentence vector input by the morphological analysis unit 11 (step S
21). Then, the sentence vector generation unit 14 calculates the TFIDF of the word obtained by the morphological analysis and vectorizes the sentence (step S22). The specific calculation of TFIDF is performed based on the following equation. TFIDF = TF (d, t) × IDF (t) where TF (d, t): Frequency of appearance of word t in text d IDF (t): log [DB (db) / f (t, db)] DB (db): Total number of texts stored in a certain database db f (t, db): Number of texts in which a word t appears in a certain database db.

As an example of a specific calculation, a case where the TFIDF value of the text of the 23rd song, which is the text d for the word t = “month”, is calculated at the Hyakunin Isshu will be described. . Text d = "If you look at the moon you can't do anything but fall in my own fall" In the text d of this 23rd song, the word t = "month" frequency of occurrence TF (d, t) is 1 (Times). In addition, the number of all songs of Hyakunin Isshu (total text number DB (db)) is 100, and the number of songs including the word t = “month” in all Hyakunin Isshu (appearance frequency f (t, db)) is 11. is there. Therefore, the word t =
TF for the text "Moon" d = "If you see the moon, things are small, but in my own fall I don't know"
The IDF value has the following results.

In the determination of step S12, when all the sentences that form the summary target sentence have already been extracted (step S12; Yes), the sentence analysis unit 12 causes the sentence analysis unit 12 to execute step S12.
All sentence vectors obtained in the process of 13 are sent to the sentence vector generation unit 15 to generate sentence vectors (step S14). Here, the text vector generation operation will be described with reference to the processing flowchart of the text vector generation unit 15 in FIG. The sentence vector generation unit 15
The average value of the sentence vectors generated by the sentence vector generation unit 14 is calculated (step S31). Next, the sentence vector generation unit 15 uses the calculated average value of the sentence vectors (average vector of sentence vectors) as the sentence vector, and the sentence analysis unit 1
2 (step S32). Next, the sentence analysis unit 12
Sends the sentence vector obtained in the process of step S13 and the sentence vector obtained in the process of step S14 to the sentence vector comparison unit 16 to extract an important sentence (step S15).

Here, the operation of extracting the important sentence used for creating the abstract will be described with reference to the processing flow chart of the sentence vector comparison unit 16 of FIG. The sentence vector comparison unit 16
A process for extracting one sentence vector from the sentence vectors sent from the sentence analysis unit 12 is performed (step S41). Next, the sentence vector comparison unit 16 performs step S4.
It is determined whether or not there is any sentence vector to be retrieved in 1 (step S42). In the determination of step S42, if there is no sentence vector to be extracted (step S42; Yes), the sentence vector comparison unit 16
Ends all the important sentence extraction operations, so the processing ends.

If there is a sentence vector to be extracted in the determination in step S42 (step S; N
o), the sentence vector comparison unit 16 calculates the similarity on the basis of the extracted sentence vector, the feature vector corresponding to the field to which the summary target sentence previously designated by the user belongs, and the sentence vector of the summary target sentence (step S43). ). As a concrete calculation procedure of the degree of similarity, first, the inner product (first inner product) of the sentence vector and the sentence vector is obtained. Next, the inner product (second inner product) of the sentence vector and the feature vector corresponding to the field is obtained. Then, the sum of the first inner product and the second inner product is set as the similarity.

Next, the sentence vector comparison unit 16 determines whether or not the degree of similarity is equal to or greater than a preset constant value (a value corresponding to the reference degree of similarity for determining whether or not the sentence is an important sentence) ( Step S44). In the determination of step S44, if the similarity is less than a certain value (step S
44; No), since the sentence corresponding to the sentence vector used in the calculation of the similarity in step S43 is not an important sentence in the abstract creation, the process proceeds to step S41 until the process for all sentence vectors is completed. The same processing is performed thereafter. In the determination of step S44, if the similarity is equal to or higher than a certain value (step S44; Ye
s), the sentence corresponding to the sentence vector used for the calculation of the degree of similarity in step S43 is extracted as an important sentence to be used for creating the abstract (step S45). Then, the process returns to step S41, and the same process is repeated until the process for all sentence vectors is completed. As a result, the sentence analysis unit 12 sends the important components extracted by the sentence vector comparison unit 16 to the summary sentence output unit 13. As a result, the summary output unit connects the transmitted important sentences, generates and outputs the summary, and ends the processing (step S16). A specific example of the abstract will be described in detail below.

Next, the summary sentence obtained by the method of the above-described embodiment is compared with the summary sentence obtained by using the important sentence extracted using the conventional similarity of sentence vectors. In the following description, it is assumed that a part of the conventional technique disclosed in the specification of Japanese Patent No. 2944346 is taken out and used as a text to be summarized. In this case,
A sentence number is added to the original sentence to facilitate explanation. The sentences to be summarized are as follows. "1: An idea is a new combination of known information, and nothing can be created from nothing. Therefore, when thinking when creating a document,
Frequently, citations are made by referring to existing documents. 3: Generally, the number of existing documents to be referred to is large, and the amount of sentences in each document is also large. 4: Therefore, if the entire existing document to be used as the reference is read as it is, it consumes time and labor, and the power to create the document, which is the original purpose, is reduced. 5: The number of documents to be referred to can be reduced by narrowing down the document contents using a search device. 6: Also, the large amount of sentences in each document can be reduced by using the abstract / abstract extracting device. 7: Here, consider a case where abstract / abstract extraction is extracted from a document in order to reduce the effort of reference by reducing the sentence amount of each document. 8: In this case, it is necessary to use a method in which the important content included in the original document is not damaged even if the text amount of the document is reduced. 9: As a method of document summarization that has been conventionally proposed,
There are the following two methods. 10: The first method is to analyze a sentence in a surface layer. 11: In this method, an important part of a sentence is determined from an appearance frequency analysis of words, and a summary sentence is generated by combining words included in the original document and extracting sentences, and sentence end expressions of sentences. It also includes those that extract emphasized / claimed sentences in a sentence by a verb. 12: The second method is to semantically analyze a sentence. 13: In this method, the format and context of a sentence are assumed in advance, the sentence is analyzed according to the assumption, and a summary is extracted. Includes defining significance and extracting summaries. The important sentences extracted by using the similarity of the conventional sentence vector with respect to the above sentence to be summarized are the following sentence numbers 7, 8, 13
Will be three sentences. “7: Here, in order to reduce the reference effort by reducing the sentence volume of each document,
Consider the case of extracting abstract extraction. "8: In this case, it is necessary to use a method that does not impair important contents contained in the original document even if the text amount of the document is reduced." In this paper, we assume the format and context of the sentence and analyze the sentence according to that assumption to extract the abstract, and we also use the coarseness of the dependency of the sentence to define the importance of the content. What is extracted is included. ”On the other hand, the important sentences extracted by the method using the feature vector according to the field of this embodiment are the following sentence numbers 8, 11, and 1.
It will be 3 sentences of 3. "8: In this case, it is necessary to use a method that does not impair important contents included in the original document even if the text amount of the document is reduced." Decide the important part of the sentence from the appearance frequency analysis and generate the summary sentence by combining the words contained in the original document and extracting the sentence, or emphasize in the sentence by the sentence end expression and the idiom Includes extraction of asserted sentences. "" 13: This method presumes the format and context of a sentence, analyzes the sentence according to the assumption, and extracts a summary, and a sentence. Includes those that define the importance of content by using the coarseness of dependency of the content and extract abstracts. ”

Here, the difference in the important sentence obtained by each method will be described. In the above-mentioned conventional method, since the frequency of appearance of words such as "appearance frequency", "emphasis", and "generation" is low in the entire document, the sentence of sentence number 11 is not recognized as an important sentence. Therefore, the summary sentence obtained by the conventional method is as follows. “Here, consider the case of extracting the abstract / abstract extraction from the document in order to reduce the reference effort by reducing the sentence volume of each document. In this case, the document volume is reduced. However, it is necessary to use a method that does not damage the important contents of the original document.In this method, the format and context of the sentence are assumed in advance, and the sentence is analyzed according to that assumption. And abstracts are extracted, and abstractions are extracted by defining the importance of the content by using the coarseness of the dependency of sentences. ”

On the other hand, in the method of this embodiment, the characteristic vector in the field of natural language processing using the vector space method is previously used by utilizing the patent specification data of the natural language processing using the vector space method. Creating.
Therefore, in the field-specific feature vector dictionary 18, words such as “appearance frequency”, “emphasis”, and “generation” that frequently appear in natural language processing using the vector space method are used as feature vectors in the natural language processing field. Corresponding word TFIDF data is also included. As a result, the sentence of sentence number 7 is no longer treated as an important sentence in the natural language processing field, and the sentence of sentence number 11 is recognized as an important sentence instead.

As a result, the summary output from the summary output unit 13 is as follows. "In this case, it is necessary to use a method that reduces the text volume of the document so that the important contents contained in the original document are not impaired. A method that determines the important points and extracts words that are included in the original document to generate a summary sentence, and a method that extracts the emphasized / claimed sentence in the sentence by the sentence end expression and the idiom This method is based on the assumption that the form or context of the sentence is preliminarily analyzed and the sentence is analyzed according to the hypothesis to extract a summary, or the sensitivity of sentence dependency is used. It includes defining content importance and extracting abstracts. "

As described above, according to the present embodiment, the summary sentence is created in consideration of even the words characteristic of the field to which the sentence to be summarized belongs, so that a more accurate summary sentence can be easily created. It becomes possible to do.

A modified example of this embodiment will be described below. In the above description, the case where the inner product of sentence vectors is used as the sentence similarity is described, but it is also possible to use the Euclidean distance of the sentence vector as the sentence similarity. As described above, the vector D1 of the sentence d1 in the database storing a plurality of sentences can be expressed by the following equation. D1 (d1) ＝ (TF (d1, t1) * IDF (t1), TF (d1, t2) * IDF (t2), TF (d1, t3) * IDF (t3), ......, TF (d1, tn ) * IDF (tn))

Similarly, the vector D2 of the sentence d2 can be expressed by the following equation. D2 (d2) ＝ (TF (d2, t1) * IDF (t1), TF (d2, t2) * IDF (t2), TF (d2, t3) * IDF (t3), ……, TF (d2, tn ) * IDF (tn)) where t1, t2, t3, ..., Tn are words that appear in the database, and t1 to tn correspond to all words in the database. Based on this, the Euclidean distance DE, which is the similarity between the document vector D1 (d1) corresponding to the sentence d1 and the document vector D2 (d2) corresponding to the sentence d2, can be calculated by the following formula. DE = | D1 (d1) -D2 (d2) |

[0030]

According to the present invention, when automatically generating a summary sentence for a sentence to be summarized, the summary sentence is generated in consideration of the field to which the sentence to be summarized belongs. Can be easily created.

[Brief description of drawings]

FIG. 1 is a schematic functional configuration block diagram of a document summarizing device according to an embodiment.

FIG. 2 is an explanatory diagram of a plurality of sentence vectors corresponding to a plurality of learning sentences belonging to each field.

FIG. 3 is an explanatory diagram of a database in a field-specific feature vector dictionary.

FIG. 4 is a processing flowchart of a sentence analysis unit.

FIG. 5 is a processing flowchart of a sentence spectrum generation unit.

FIG. 6 is a processing flowchart of a sentence vector generation unit.

FIG. 7 is a processing flowchart of a sentence vector comparison unit.

[Explanation of symbols]

10 ... Document summarizing device 11 ... Morphological analysis unit 12 …… Sentence analysis section 13 ... Summary output section 14 ... Sentence vector generation unit 15 …… Sentence vector generator 16: Sentence vector comparison unit (important sentence extraction unit) 17 ... Field-specific feature vector generator 18: Feature vector dictionary by field

Claims (15)

[Claims] 1. A sentence vector generation unit that generates a sentence vector of a sentence that constitutes a sentence to be summarized, a sentence vector generation unit that generates a sentence vector corresponding to the sentence to be summarized based on the sentence vector, A sentence vector, an area-specific feature vector that characterizes a field to which the summary target sentence belongs, and an important sentence extraction unit that extracts an important sentence from the summary target sentence based on the sentence vector. Document summarization device. 2. The document summarization apparatus according to claim 1, wherein the sentence vector generation unit sets an average vector of sentence vectors of sentences constituting the sentence to be summarized as the sentence vector. apparatus. 3. The document summarizing device according to claim 1, wherein the important sentence extraction unit calculates a first inner product which is an inner product of the sentence vector and the sentence, and the sentence vector and the field-specific feature vector. A second inner product that is an inner product of the first inner product and the second inner product is defined as a similarity, and the important sentence is extracted by comparing the similarity with a predetermined reference similarity. A document summarizing device characterized by: 4. The document summarizing apparatus according to claim 1, further comprising a summary sentence creating unit that creates a summary sentence based on the extracted important sentence. 5. The document summarizing apparatus according to claim 1, further comprising a field-specific feature vector generation unit that generates the field-specific feature vector based on a plurality of learning sentences for each field. Summarization device. 6. The document summarizing apparatus according to claim 5, wherein the field-specific feature vector generation unit generates a sentence vector corresponding to each of the learning sentences, and generates a sentence vector corresponding to the plurality of learning sentences. A document summarization device, wherein an average vector is generated as the field-specific feature vector. 7. A sentence vector generation process for generating a sentence vector of a sentence constituting a summary target sentence; a sentence vector generation process for generating a sentence vector corresponding to the summary target sentence based on the sentence vector; A sentence vector, a field-specific feature vector that characterizes a field to which the sentence to be summarized belongs, and an important sentence extraction process of extracting an important sentence from the sentence to be summarized based on the sentence vector. Control method for document summarization device. 8. The control method of the document summarizing apparatus according to claim 7, wherein the sentence vector generating step includes a step of setting an average vector of sentence vectors of sentences constituting the sentence to be summarized as the sentence vector. A method of controlling a document summarizing device, comprising: 9. The control method of the document summarizing apparatus according to claim 7, wherein the important sentence extracting step includes a step of calculating a first inner product which is an inner product of the sentence vector and the sentence; Calculating a second inner product that is an inner product with the field-specific feature vector; using a sum of the first inner product and the second inner product as a similarity; and using the similarity as a predetermined reference similarity. And a step of extracting the important sentence by comparing with the above. 10. The control method of the document summarizing device according to claim 7, further comprising a summary sentence creating step of creating a summary sentence based on the extracted important sentence. . 11. A control program of a document summarization device for causing a computer to function as a document summarization device for generating summary text data based on input summarization target text data, which corresponds to the summarization target text data. Generate a sentence vector of a sentence that constitutes a summary target sentence, generate a sentence vector corresponding to the summary target sentence based on the sentence vector, and characterize the sentence vector, the field to which the summary target sentence belongs A control program for a document summarizing device, wherein an important sentence is extracted from the sentence to be summarized based on another feature vector and the generated sentence vector. 12. The control program for the document summarizing apparatus according to claim 11, wherein an average vector of sentence vectors of sentences constituting the sentence to be summarized is calculated and used as the sentence vector. Device control program. 13. The control program of the document summarizing apparatus according to claim 11, wherein a first inner product which is an inner product of the sentence vector and the sentence is calculated, and a first inner product of the sentence vector and the field-specific feature vector is calculated. Calculating a certain second inner product, making the sum of the first inner product and the second inner product a similarity, and extracting the important sentence by comparing the similarity with a predetermined reference similarity. A control program for a document summarizing device. 14. The control program of the document summarizing device according to claim 11, wherein the control program of the document summarizing device is configured to create a summary sentence based on the extracted important sentence. 15. A recording medium on which the control program of the document summarizing device according to claim 11 is recorded.