JP7365849B2 - Summarization devices and programs - Google Patents

Description

The present invention relates to a summarization device and a program.

For example, when a large-scale disaster or the like occurs, many news articles are simultaneously created by news organizations and the like. Furthermore, it is necessary to create several forms of text regarding certain facts obtained through interviews, etc., such as a manuscript for a news program or a manuscript of a summary text. Here, the summarized text is, for example, text distributed using an L-shaped screen of a television, text provided by a smartphone application, or the like. On the other hand, in situations such as large-scale disasters, reporters who write news articles need to conduct reporting activities in a variety of areas, and there may be a temporary shortage of manpower.

For example, if a summary version of text can be automatically generated based on a news manuscript written by a person, it is possible to save labor and speed up content production.

Non-Patent Document 1 describes a technology for automatically generating news headlines using neural network technology.

Non-Patent Document 2 describes a technique for automatically generating summaries of news articles by imposing constraints on summary length, sentence length, and number of sentences.

Kazuma Murao, Ken Kobayashi, Hayato Kobayashi, Taichi Yatsuka, Takeshi Masuyama, Tatsuru Higurashi, Yoshimune Tabuchi, “A Case Study on Neural Headline Generation for Editing Support”, Proceedings of NAACL-HLT 2019, pages 73-82, Minneapolis, Minnesota, June 2 - June 7, 2019, Association for Computational Linguistics, URL: https://aclweb.org/anthology/papers/N/N19/N19-2010/ Shun Tanaka, Ryohei Sasano, Daiya Takamura, Manabu Okumura, “News article summaries with constraints on summary length, sentence length, and number of sentences,” Proceedings of the 22nd Annual Conference of the Language Processing Society, March 2016, URL: https //www.anlp.jp/proceedings/annual_meeting/2016/pdf_dir/D1-3.pdf

The technology described in Non-Patent Document 1 uses a neural network to generate news headlines. However, when using a neural network, there is a problem in that the relationship between input data and output data cannot necessarily be accurately determined in advance due to the nature of the technology. For example, if you know to some extent what kind of information is included in the input news manuscript and what kind of information you want to include in the output generated from it, you can use a neural network. It is more reliable to obtain the expected output by performing processing based on rules, etc. than by relying on it.

The technique described in Non-Patent Document 2 generates a summary of an article after giving constraints on, for example, the number of sentences to be output. However, for purposes such as reporting, it is desirable to generate a large number of summary sentences by extracting as much important information as possible from the input manuscript, rather than imposing restrictions on the number of output sentences.

The present invention has been made based on the above circumstances, and uses processing based on rules, etc., without relying on machine learning, based on the input text, without setting any restrictions on the number of output sentences. An object of the present invention is to provide a summary device and program that can automatically generate a summary that comprehensively includes necessary information contained in a document.

[1] In order to solve the above problems, a summarization device according to one aspect of the present invention includes a text acquisition unit that acquires text data, and performs morphological analysis processing or refers to place name information stored in advance. By doing so, a place name extraction unit extracts a place name from the text data, an influence range extraction unit extracts an expression including a numeral included in the text data as an expression of the influence range, and a verb included in the text data. a theme extraction unit that extracts the included expression as a theme, the place name, the expression of the range of influence, and the theme, wherein at least either the place name or the expression of the range of influence is missing. and a summary generation unit that generates a summary by applying a pre-stored generation rule to the set that allows .

[2] Further, in one aspect of the present invention, in the above summarization device, the place name includes a high-level place name, a middle-level place name, and a low-level place name as a hierarchical structure, and the place name information includes the high-level place name. further comprising a topic area identification unit that specifies a topic area based on the place name extracted by the place name extraction unit, the topic area identification unit including affiliation relationship information that is information of the intermediate place name belonging to the area name. , by estimating the top place names to which the text data is a topic based on the set of middle place name candidates extracted by the place name extraction unit, identifying the topic area represented by the top place names; It is something.

[3] Further, in one aspect of the present invention, in the above summarization device, the place name information includes adjacency relationship information indicating whether or not the pair of high-rank place names are in an adjacency relationship; The identification unit determines whether or not the plurality of estimated high-rank place names are adjacent to each other by referring to the adjacency relationship information, and transitively determines the relationship between pairs of high-rank place names having an adjacency relationship. The topic area is specified as a set of a plurality of high-rank place names obtained by applying the method to the above.

[4] Further, in one aspect of the present invention, in the above summarization device, the topic area specifying unit is based on a set of intermediate place name candidates extracted from the text data by referring to the place name information. In this method, the scores of the top place names are calculated and the top place names with relatively good scores are identified as the topic areas, and the scores of the top place names are determined based on the affiliation information. A value obtained by taking the sum of the reciprocal of the number of all high-rank place names to which the mid-rank place name candidate belongs for each of the mid-rank place name candidates to which the mid-rank place name candidate belongs, for all of the mid-rank place name candidates belonging to the high-rank place name; It is calculated as follows.

[5] Further, in one aspect of the present invention, in the above summarization device, the place name extraction unit selects one of the high-rank place names and the middle-rank place names among the place names extracted by performing the morphological analysis process. extracting a place name that was not specified in the lower place name as the lower place name, and placing the lower place name in the position closest to the lower place name among the intermediate place names that appear in the text data before the lower place name. The name is extracted as a lower place name belonging to the intermediate place name that appears.

[6] Further, in one aspect of the present invention, in the above summarization device, the place name extraction unit further extracts a noun accompanied by a specific keyword that makes it inferred to be a place name as the subordinate place name, and The lower place name is extracted as a lower place name belonging to the intermediate place name that appears in the position closest to the lower place name among the intermediate place names that appear in the text data prior to the lower place name. .

[7] In addition, one aspect of the present invention is to provide a computer with a text acquisition unit that acquires text data, and to perform morphological analysis processing or to refer to pre-stored place name information. a place name extraction unit that extracts a place name from data; an influence range extraction unit that extracts an expression containing a numeral included in the text data as an expression of an influence range; and an influence range extraction unit that extracts an expression containing a verb included in the text data as a subject. a set of a theme extraction unit to be extracted, the place name, the expression of the range of influence, and the theme, the set allowing for the lack of at least one of the place name or the expression of the range of influence; , and a summary generation unit that generates a summary by applying pre-stored generation rules.

According to the present invention, it is possible to automatically generate a summary from text data. For example, even if the text data contains a lot of information (for example, information about dozens of evacuation centers in the event of a disaster is listed), a summary that covers that information can be generated. . Further, according to the present invention, a summary can be generated without relying on machine learning results. That is, the processing results do not depend on machine learning data. In other words, according to the present invention, an unintended summary is less likely to be output.

1 is a block diagram showing a schematic functional configuration of a summarization device according to an embodiment of the present invention. FIG. FIG. 3 is a schematic diagram showing an example of text data acquired by the text acquisition unit according to the embodiment. It is a schematic diagram showing an example of composition of data of a place name dictionary stored in a place name dictionary storage unit according to the same embodiment. It is a schematic diagram showing the example of composition of the data of the adjacency dictionary which the adjacency relationship dictionary storage part by the same embodiment stores. It is a flowchart which shows the procedure of the process by the topic area identification part by the same embodiment. FIG. 3 is a schematic diagram showing place name data and score data for explaining an example of processing by the topic area identification unit according to the embodiment. FIG. 6 is a schematic diagram illustrating data for explaining an example of processing by a lower place name extraction unit according to the embodiment. FIG. 7 is a schematic diagram illustrating a specific example of the expression of the influence range extracted by the influence range extraction unit according to the embodiment. FIG. 7 is a schematic diagram showing another specific example of the expression of the influence range extracted by the influence range extraction unit according to the embodiment. FIG. 6 is a schematic diagram illustrating an example of a sentence dependency analysis result for explaining the processing of the subject extraction unit according to the embodiment. 7 is a flowchart showing a processing procedure of a summary generation unit according to the embodiment. FIG. 6 is a schematic diagram illustrating an example of rules for generating a summary expression, which are stored in a generation rule storage unit according to the embodiment. 3 is a flowchart showing the overall processing procedure of the summarization device according to the embodiment.

Embodiments of the invention will be described below with reference to the drawings. The present embodiment relates to a technology for automatically generating a summary based on text data such as a news manuscript, and an apparatus realized using the technology. First, the concept used in this embodiment will be explained.

The "upper place name", "middle place name", and "lower place name" are place names at each level of a place name having a hierarchical structure. The region of the place represented by the high-rank place name includes the region of the place represented by the middle-rank place name. The region of the place represented by the intermediate place name includes the region of the place represented by the lower place name. For example, the high-level place name is a prefecture name. Further, for example, the intermediate place name is a city, ward, town, or village name. Further, for example, a lower place name is a place name lower than a city, ward, town, or village name. For example, "Seijo," "Kinuta," "Okura," etc. are lower place names belonging to Setagaya Ward, Tokyo. Note that upper, middle, and lower only represent a relative relationship, and a name other than the prefecture name may be used as the upper place name, and a name other than the city, ward, town, or village name may be used as the middle place name. Other names such as country, state, district, county, borough, etc. may also be used as place names having a hierarchical structure.

The "topic area" is the area targeted for the topic of texts such as news manuscripts. For example, a "hot topic area" can be expressed using a single or multiple high-level place names (prefecture names). "Scope of influence" is an expression that expresses the range or magnitude of the event that is the topic of the text. For example, the expression "23 people in 10 households" is an example of an expression that indicates the range of damage caused by a natural disaster. The "subject" is an expression representing the phenomenon itself that is the topic of the text. For example, the expression ``We have issued an evacuation advisory and are calling for people to evacuate'' is an example of a topic that can be discussed in a news manuscript about a natural disaster.

The summarization device according to this embodiment analyzes the text data that is the manuscript of a news article, and determines the location where the event is occurring, the range of influence of the event, and the extent of the event. Extract thematic information. The summarization device uses part-of-speech information and dictionary data obtained by performing morphological analysis of text data for the extraction process. The summarization device generates and presents a bulleted summary from the obtained information while removing duplicate content.

FIG. 1 is a block diagram showing a schematic functional configuration of a summarization device 1 according to this embodiment. As shown in the figure, the summarization device 1 includes a text acquisition section 21, a middle place name extraction section 22, a topic area identification section 23, a lower place name extraction section 24, an influence range extraction section 25, and a subject extraction section 26. , a summary generation section 27, a place name dictionary storage section 31, an adjacency relationship dictionary storage section 32, and a generation rule storage section 33. Note that each of the intermediate place name extraction section 22 and the lower place name extraction section 24 may be simply referred to as a "place name extraction section." Each of these functional units can be realized by, for example, a computer and a program. Furthermore, each functional section has a storage means, if necessary. The storage means is, for example, variables on the program or memory allocated by executing the program. Furthermore, if necessary, nonvolatile storage means such as a magnetic hard disk drive (HDD) or solid state drive (SSD) may be used. Furthermore, at least some of the functions of each functional unit may be realized as a dedicated electronic circuit instead of a program. The functions of each part are as explained below.

The text acquisition unit 21 acquires text data given from the outside. This text is, for example, the text of a news manuscript written by reporters. In this embodiment, the text of a news manuscript that describes events in a specific region is particularly targeted for processing. However, the content of the text may be other content and is arbitrary.

The medium place name extraction unit 22 extracts medium place names from the text data acquired by the text acquisition unit 21. The intermediate place name is, for example, a city, ward, town, or village name. The intermediate place name extraction unit 22 may refer to the place name dictionary storage unit 31 when extracting intermediate place names. Information stored in the place name dictionary storage section 31 will be explained later. For example, the intermediate place name extraction unit 22 extracts intermediate place names by performing morphological analysis using a dictionary that includes all city, ward, town, and village names. The morphological analysis process itself can be performed using existing technology. As an example, existing morphological analysis tools such as MeCab may be used. What is obtained through the processing of the intermediate place name extraction unit 22 is a candidate for a medium place name.

The hot topic area specifying unit 23 specifies the hot topic area in the original text based on the list of medium place names extracted by the medium place name extracting unit 22. Specifically, the topic area specifying unit 23 estimates where the top place names that are being talked about are based on the middle place name candidates extracted by the middle place name extracting unit 22. There may be a plurality of higher place names estimated here. The high-level place name is, for example, a prefecture name. After estimating the top place name (single or plural), the topic area identification unit 23 selects only the middle place names that belong to the top place names that are the estimation results from among the middle place name candidates extracted by the middle place name extraction unit 22. Output as a list of location names. Among the intermediate place name candidates extracted by the intermediate place name extraction unit 22, the intermediate place names that do not belong to the high-rank place names that are the estimation results are determined not to correspond to the area that is being talked about in the original text, abandoned.

In other words, the topic area specifying unit 23 estimates the top place names that the text data is a topic of, based on the set of middle place name candidates extracted by the middle place name extraction unit 22. Identify topical areas.

As one specific method, the topic area identification unit 23 refers to information (adjacency relationship information) stored in the adjacency relationship dictionary storage unit 32, and determines whether a plurality of estimated top place names are adjacent to each other. A set of a plurality of high-rank place names obtained by determining whether or not the above-mentioned place names are determined and transitively applying the relationship between pairs of high-rank place names having an adjacency relationship (an example of a specific procedure for obtaining this set will be explained later) This is to identify the topic area expressed as (). Here, "a set of a plurality of high-rank place names obtained by transitively applying the relationship between pairs of high-rank place names that have an adjacency relationship" means, for example, that high-rank place names A and B have an adjacent relationship, and This method is based on a method in which, when place names B and C are in an adjacent relationship, it is determined that higher place names A, B, and C (set) are in an adjacent relationship by applying the transitivity law.

As a more specific example, the topic area identification unit 23 calculates a score for each top place name based on the set of middle place name candidates extracted from the text data, and calculates a score for each top place name. A method of specifying place names as topic areas can be used. The score of the high-rank place name is calculated based on the affiliation information (information indicating which high-rank place name belongs to which middle-rank place name) for each of the mid-rank place name candidates that belong to the high-rank place name. It is calculated as the sum of the reciprocal of the number of all high-rank place names to which the above place name belongs, for all of the mid-rank place name candidates belonging to the high-rank place name. The specific procedure will be explained later.

For example, a city, town or village name (medium place name) that exists in multiple prefectures (higher place names), such as a place name such as "Asahi Town," is extracted as a middle place name candidate by the intermediate place name extraction unit 22. Names of cities, towns, and villages that exist in prefectures that do not fall under the topic area should be discarded. Further, a place name such as "Asahi Town", which can be both a medium place name and a low-level place name, may be extracted by the medium place name extracting unit 22 as a medium place name candidate. If a low-level place name in the text is mistakenly extracted as a middle-level place name candidate, that middle-level place name should be discarded. The topic area specifying unit 23 determines whether the intermediate place name extracted by the intermediate place name extracting unit 22 is truly the intermediate place name of the topic area by correctly estimating the upper rank place name. The specific processing procedure by the topic area identification unit 23 will be explained in more detail later.

The lower place name extraction unit 24 acquires lower place names from the text data acquired by the text acquisition unit 21. Specifically, the lower place name extraction unit 24 extracts lower place names from the text data by using morphological analysis processing and rule matching processing. Further, the lower place name extraction unit 24 determines which intermediate place name the extracted lower place name is included in. First, the low-level place name extraction unit 24 extracts, as a low-level place name, a word that is determined to be a place name as a result of the morphological analysis process, but is neither a high-level place name nor a middle-level place name. Next, when the text data includes special endings such as "district", "region", "surroundings", etc., the lower place name extraction unit 24 combines these special endings and the word immediately before them to create a lower place name. Extract as. These special endings indicate that it is a place name, and may be called place name endings. Further, the lower place name extraction unit 24 uses the rule that the extracted lower place name belongs to the intermediate place name immediately before the place where the lower place name appears (the intermediate place name that is closest in the forward direction than the lower place name). , determine the intermediate place name to which the lower place name belongs. Here, the intermediate place name is, for example, a city, ward, town, or village name, and the lower place name is a character name, etc. lower than the city, ward, town, or village name. An example of the procedure for extracting lower place names will be explained later.

The influence range extracting unit 25 extracts expressions such as the magnitude of the degree or the magnitude of damage (referred to as “the influence range”) expressed by the sentence included in the text data acquired by the text acquisition unit 21. Specifically, the influence range extraction unit 25 extracts an expression including a numeral included in the text data as an expression of the influence range. More specifically, the influence range extraction unit 25 extracts the expression of the influence range according to a predetermined rule based on a number (numerical word) and the parts of speech of the words before and after the number. For example, the rules are as follows.

Rule 1: Columns that match the pattern below are expressions of influence ranges. However, in the following pattern, the part enclosed in parentheses is an optional part.
Pattern: {noun or prefix} + number word (+conjunctive particle + number word) (+{{noun, but not adjective stem} or parallel particle or adjective}
Note that the parallel particle is a word used to connect two or more words in an equal relationship by being attached to various words.

Rule 2: When expressions of a plurality of influence ranges appear consecutively, they are combined into one influence range expression.

Rule 3: For example, expressions that indicate date and time, such as "2 days," "4 o'clock," and "10 minutes," are excluded as not expressing the range of influence, even if they match the pattern specified in Rule 1. do.

The theme extraction unit 26 extracts the theme of the sentence included in the text data acquired by the text acquisition unit 21. The theme extraction unit 26 extracts expressions containing verbs included in the text data as themes. More specifically, for example, the subject extraction unit 26 first performs a dependency analysis process on the text data. The process itself of analyzing syntactic dependencies can be implemented using existing technology. For example, the theme extraction unit 26 may use CaboCha, which is an existing Japanese dependency analyzer. Based on the result of the dependency analysis, the theme extraction unit 26 extracts a theme by connecting a clause that directly relates to a clause that includes a verb. However, at this time, even if the clause is directly related to a clause that includes a verb, the theme extraction unit 26 uses expressions of the range of influence (expressions extracted by the range of influence extraction unit 25) and place names (the intermediate place name extraction unit 22 or extracted by the lower place name extraction unit 24) are excluded. Further, if the above-mentioned verb is a verb that is easy to be used as a denotative form, the subject extraction unit 26 removes the verb and the particle or auxiliary verb immediately before it, and converts it into a denotative form. Examples of verbs that are easy to use as formal expressions include verbs such as say, do, become, and put out. Incidentally, information on whether a certain verb is a verb that is easy to be used as a verbal stop may be stored in the dictionary data in advance, and the dictionary may be referred to.

The processing procedure by the theme extraction unit 26 is summarized as follows. That is, the theme extracting unit 26 extracts a theme and formats the extracted theme by the procedure listed below.
First stage of theme extraction: The theme extraction unit 26 refers to the result of sentence dependency analysis. If a certain clause is directly related to the next clause, the theme extraction unit 26 combines those clauses into one.
Second stage of theme extraction: The theme extraction unit 26 extracts a group that includes a verb from among the above groups as a theme. However, at this time, the theme extraction unit 26 excludes from the target groups that include specific verbs that are unlikely to be the theme.
Third stage of theme extraction: The theme extraction unit 26 connects clauses directly related to clauses containing the above-mentioned verbs. However, the theme extraction unit 26 selects connecting clauses only within a range that does not exceed clauses that include phrases extracted as expressions of place names or ranges of influence.
First stage of formatting: If the extracted verb is a specific verb, the subject extracting unit 26 removes the verb and formats it into a formal form.
Second stage of formatting: The theme extraction unit 26 removes expressions (conjunctions, auxiliary verb "masu", etc.) that are inappropriate for summary data.
Third stage of formatting: The theme extraction unit 26 removes the word following the verb, or if the verb is followed by an auxiliary verb expressing negation, the word following the auxiliary verb. Here, the auxiliary verb expressing negation is, for example, "nai" or "nu".

The summary generation unit 27 generates at least part of the information on the topic area specified by the topic area identification unit 23, the information on the influence range extracted by the influence range extraction unit 25, and the theme information extracted by the theme extraction unit 26. to generate a summary. Specifically, the summary generation unit 27 generates a set of a place name, an expression of the range of influence, and a subject (a set that allows for the omission of at least either the place name or the expression of the range of influence). , generates a summary by applying pre-stored generation rules. The summary generation unit 27 presents the generated summary data to the user, for example, by displaying it on a screen.

The place name dictionary storage unit 31 stores dictionary data regarding place names. The place name dictionary storage unit 31 may store information representing the affiliation relationship between relatively higher place names and lower place names. As an example, the place name dictionary storage unit 31 stores geographical affiliation relationships (inclusion/coverage of areas represented by those place names) between high-level place names (for example, prefecture names) and middle-level place names (for example, city, ward, town, and village names). Inclusion relationship) information (also referred to as "belonging relationship information") may be stored. The "affiliation relationship information" mentioned here is information about middle-rank place names that belong to high-rank place names.

The adjacency relationship dictionary storage unit 32 stores information indicating whether or not place names are in an adjacency relationship (also referred to as "adjacency information"). In particular, in this embodiment, the adjacency relationship information described above is information indicating whether or not a pair of high-rank place names are in an adjacency relationship.

Note that the information stored in the place name dictionary storage section 31 and the information stored in the adjacency relationship dictionary storage section 32 may be referred to as "place name information."

The generation rule storage unit 33 stores rules for generating a summary based on information extracted from the original text data. Specifically, these rules are for generating summary data based on place names, expressions of influence areas, and themes. Examples of generation rules will be explained later.

FIG. 2 is a schematic diagram showing an example of text data acquired by the text acquisition unit 21. Figure (A) shows the first example, ``Murakami City, Niigata Prefecture, has issued ``information on evacuation preparations,'' and the three locations of the former Shiodamachi Elementary School, Shimokita Yurika Hall, and Senpoku Hall have been We opened an evacuation center. ” shows the text of the news manuscript. Figure (B) shows the second example, ``Due to this heavy rain, an evacuation advisory has been issued to eight people from three households in Asahi-cho, Takahashi City, and information on evacuation preparations has been issued to some areas in Okayama City.'' .'' is the text of the news manuscript. In this way, the text acquisition unit 21 acquires, for example, text data of news related to disasters in a specific area. However, the content of the news is not limited to disasters. These news text data are, for example, written by news reporters and input into the summarization device.

FIG. 3 is a schematic diagram showing a configuration example of data of a place name dictionary stored in the place name dictionary storage unit 31. As shown in FIG. As shown in the figure, the place name dictionary storage unit 31 stores a place name dictionary as table-format data having items such as high-rank place names, middle-rank place names, and affiliations, for example. In addition, in the figure, a line number is attached to each line of data for convenience. The high-level place name is, for example, a prefecture name. Further, the intermediate place name is, for example, a city, ward, town, or village name. This data also represents the inclusion relationship between high-rank place names and middle-rank place names. That is, when looking at a certain row, place names that correspond to high-level place names include place names that correspond to intermediate place names. As a specific example, referring to line number 17, Tokyo refers to Okutama Town. The same applies to other lines. The place name dictionary storage unit 31 holds, for example, data regarding the names of all prefectures and the names of all cities, wards, towns, and villages in Japan in this format. Furthermore, in the illustrated example, the affiliation item has information on the name of a higher-level place to which the "ward" belongs. Usually, a ward may belong directly to a prefecture (in the case of a special ward) or to a city. For example, referring to line number 4, the top place name to which Shibuya Ward belongs is Tokyo. Furthermore, the top place name to which Nishi Ward belongs in line number 21 is Yokohama City. Furthermore, the top place name to which Saiwai Ward belongs in line number 33 is Kawasaki City. For cities, towns, and villages that are mid-level place names other than wards, there is no need to use the data item of affiliation because they always belong directly to prefectures. By referring to this place name dictionary storage unit 31, it can be known to which high-level place name the intermediate place name belongs.

FIG. 4 is a schematic diagram showing an example of the structure of data of an adjacent dictionary stored in the adjacent relationship dictionary storage unit 32. As shown in FIG. As shown in the figure, the adjacency dictionary storage unit 32 stores tabular data having a set of pairs of high-level place names. In addition, in the figure, a line number is attached to each line of data for convenience. The high-level place name is, for example, a prefecture name. As an example, the line with line number 3 has paired data for Tokyo and Yamanashi prefectures. This means that Tokyo and Yamanashi Prefecture are adjacent to each other. The same applies to data in other rows. Here, the adjacency of top place names does not necessarily mean that their areas are in contact with each other. For example, the line with line number 15 indicates that Nara Prefecture and Shiga Prefecture are in an adjoining relationship, but the areas of these two prefectures are not in contact with each other. However, when dealing with news such as natural disasters, for example, it is meaningful to treat Nara Prefecture and Shiga Prefecture as neighboring cases, so the adjacency relationship dictionary storage unit 32 stores information on the adjacency relationship between these two prefectures. That is useful.

Below, the functions of each part will be explained in more detail using specific examples of data.

A specific example of the process performed by the intermediate place name extraction unit 22 is as follows. For example, from the text data in FIG. 2A, the medium place name extracting unit 22 extracts medium place names (city, ward, town, and village names) such as "Murakami City" and "Shioda Town." Further, for example, from the text data in FIG. 2(B), the medium place name extraction unit 22 extracts medium place names (city names) such as "Takahashi City," "Asahi Town," and "Okayama City."

FIG. 5 is a flowchart showing the procedure of processing by the topic area identification unit 23. The processing procedure will be explained below along with this flowchart.

In step S11, the topic area identification unit 23 obtains a list of intermediate place name candidates. In this embodiment, the intermediate place name is a city, ward, town, or village name. This list of intermediate place name candidates is a list of intermediate place names extracted from the input text by the intermediate place name extraction unit 22. The intermediate place name extraction unit 22 extracts intermediate place names from the text by referring to the place name dictionary storage unit 31. At this time, the intermediate place name extracting unit 22, for example, also identifies a place name that is used as a middle place name in the input text, but is not used as a middle place name, but is the same as a middle place name held in the place name dictionary storage unit 31. There is also a possibility that place names with the font characters may be extracted. In this step, the topic area specifying unit 23 receives a list that may include all such intermediate place name candidates from the medium place name extracting unit 22.

In step S12, the topic area identification unit 23 calculates a top place name score. The high-ranking place name score is, for example, a prefecture score. The topic area specifying unit 23 calculates the top place name score by referring to the place name dictionary storage unit 31. For example, the calculation formula for calculating the top place name score is the following formula (1).

Top place name score = Σ (1/Number of top place names with the corresponding middle place name) ... (1)

A high-ranking place name score is calculated for each high-ranking place name using equation (1). "Σ" in formula (1) means that the high-rank place name to be calculated is for all the mid-rank place name candidates that belong to the high-rank place name among the mid-rank place name candidates extracted by the mid-rank place name extraction unit 22 for a certain input text. represents the calculation of the sum of .

When the high-level place name is a prefecture name and the intermediate place name is a city, ward, town, or village name, the above equation (1) is equivalent to the below equation (2).

Prefecture score = Σ (1/number of prefectures with the name of the city, ward, town, village) ... (2)

A prefecture score is calculated for each prefecture using equation (2). "Σ" in formula (2) is the prefecture that is the calculation target, which is the city, ward, town, and village name candidate that belongs to the prefecture among the city, ward, town, and village name candidates extracted by the intermediate place name extraction unit 22 for a certain input text. Represents a calculation that takes the sum of all.

In step S13, the topic area identification unit 23 identifies the top place name (for example, prefecture) as the topic area based on the score calculated in step S12.

Next, in step S14, the topic area identification unit 23 determines that the top place name with the highest score among the remaining top place names (for example, prefectures) is the top place name (multiple place names) that have already been determined to be a topic area. In this case, a process is performed to determine whether or not it is adjacent to any one of them. Note that the topic area specifying unit 23 can determine whether a certain high-ranking place name and another high-ranking place name are in an adjacent relationship by referring to the adjacency relationship dictionary storage unit 32.

In step S15, the process branches depending on the result of the determination process in step S14. If the high-ranking place name is adjacent to the high-ranking place name (if there are multiple places, any one of them) that has already been determined to be a hot topic area (step S15: YES), the process proceeds to the next step S16. If they are not adjacent (step S15: NO), the process advances to step S17.

In step S16, the topic area specifying unit 23 adds the top place name with the highest score among the remaining top place names (top place names not yet included in the topic area) to the topic area. After processing this step, the process returns to step S14 in order to determine whether there are any higher ranking place names to be added to the topic area. In other words, through a loop of processing in steps S14, S15, and S16, the topic area specifying unit 23 repeatedly adds adjacent top place names to the topic area. In other words, the topic area identifying unit 23 may identify a list of multiple top place names as a topic area.

If the process proceeds to step S17, in the same step, the topic area identification unit 23 selects a list of top place names determined to be topic areas, and a list of top place names that are middle-rank place name candidates acquired in step S11. Outputs a list of intermediate place names belonging to each category. That is, the topic area specifying unit 23 specifies the topic area targeted by the original text by outputting a list of high-rank place names and a list of middle-rank place names.

FIG. 6 is a schematic diagram showing place name data and score data for explaining an example of processing by the topic area identification unit 23.

FIG. 6A shows an example of a list of intermediate place names extracted by the intermediate place name extraction unit 22. As shown in the figure, in this example, the intermediate place names extracted by the intermediate place name extraction unit 22 are Okayama City, Takahashi City, and Asahi Town. FIG. 6(B) shows information on top place names obtained by the topic area identification unit 23 by referring to the place name dictionary storage unit 31, corresponding to each of the middle place names included in FIG. 6(A). . As shown in the figure, Okayama City only corresponds to Okayama Prefecture. This indicates that it has been determined from the data stored in the place name dictionary storage unit 31 that the only high-rank place name that includes the middle place name Okayama City is Okayama Prefecture. Additionally, Takahashi City only corresponds to Okayama Prefecture. This indicates that it has been determined from the data stored in the place name dictionary storage unit 31 that the only high-rank place name that includes the intermediate place name Takahashi City is Okayama Prefecture. In addition, Asahi Town corresponds to six top place names: Yamagata Prefecture, Toyama Prefecture, Mie Prefecture, Hokkaido, Niigata Prefecture, and Fukui Prefecture. This is because the place name dictionary storage unit 31 stores that each of these six high-rank place names includes the middle-rank place name Asahi-cho, and that these are the only six high-rank place names that include the middle-rank place name Asahi-cho. Expresses what was determined from the available data.

FIG. 6(C) shows the scores of each of the top place names calculated by the topic area identification unit 23. As described above, this score is calculated using equation (1) or equation (2) equivalent thereto. For example, Okayama Prefecture has Okayama City and Takahashi City as its members among the middle-ranked place names listed in FIG. 6(A). For each of Okayama City and Takahashi City, the "number of high-rank place names having the corresponding middle-rank place name" is 1. In other words, Okayama City exists only in Okayama Prefecture, and Takahashi City only exists in Okayama Prefecture. Additionally, Asahi Town does not exist in Okayama Prefecture. Therefore, according to equation (1), the score of Okayama Prefecture is calculated as (1/1+1/1), which is 2.000. Additionally, Yamagata Prefecture has Asahi Town as a member of the middle place names listed in Figure 6(A). Regarding this Asahi town, the "number of high-rank place names having the middle-rank place name" is 6. Furthermore, Okayama City and Takahashi City are not members of Yamagata Prefecture. Therefore, according to equation (1), the score for Yamagata Prefecture is (1/6), which is calculated as 0.167 (rounded to the fourth decimal place). The scores for the top place names of Toyama Prefecture, Mie Prefecture, Hokkaido, Niigata Prefecture, and Fukui Prefecture are also calculated as 0.167 using the same calculation procedure as the score for Yamagata Prefecture above.

As a result of the data processing shown in FIG. 6, the topic area specifying unit 23 specifies only Okayama Prefecture as a topic area. Toyama Prefecture, Mie Prefecture, Hokkaido, Niigata Prefecture, and Fukui Prefecture each have the intermediate place name ``Asahi Town'' as a member, but none of them have an adjacency relationship with Okayama Prefecture. The presence or absence of an adjacency relationship is determined by referring to the adjacency relationship dictionary storage unit 32, as described above. In addition, Okayama Prefecture, which is a hot topic area, has Okayama City and Takahashi City as members, out of Okayama City, Takahashi City, and Asahi Town, which are intermediate place name candidates. Therefore, the topic area specifying unit 23 outputs Okayama Prefecture as the top place name of the specified topic area, and outputs Okayama City and Takahashi City as the middle place name of the topic area. In this case, Asahi Town will be treated as not being an intermediate place name in subsequent processing.

FIG. 7 is a schematic diagram showing data for explaining an example of lower place name extraction processing by the lower place name extraction unit 24. The figure shows the text of one news article. This news article consists of three sentences. Here, symbols S1, S2, and S3 are assigned to each sentence for convenience. The sentence in S1 is ``Due to this heavy rain, an evacuation advisory has been issued to 8 people from 3 households in Asahi-cho, Takahashi City.'' The sentence in S2 is ``Also, information on evacuation preparations has been issued for some areas of Okayama City.'' The sentence in S3 is ``According to the fire department, there are some houses in the Yamada district in the mountains of Takahashi City that are submerged in water up to the floor level.'' By performing the morphological analysis process, the lower place name extraction unit 24 extracts "Takahashi City" and "Asahi Town" included in the sentence S1, "Okayama City" included in the sentence S2, and "Okayama City" included in the sentence S3. Extract "Takahashi City" as a place name. Among these, "Takahashi City" and "Okayama City" have already been identified by the process of the topic area specifying unit 23 as intermediate place names in the topic area. Therefore, the lower place name extracting unit 24 extracts only "Asahi Town" from among the place names listed above as a lower place name. Further, the lower place name extracting unit 24 extracts lower place names through rule matching processing. That is, the lower place name extraction unit 24 extracts the word "Yamada district" as a lower place name by matching with the word "district".

Note that the lower place name extraction unit 24 extracts the union of the set of lower place names extracted by the morphological analysis process and the set of lower place names extracted by the rule matching process as a lower place name.

Through the above processing, the lower place name extraction unit 24 extracts the next place name string from the text consisting of the three sentences S1, S2, and S3. The place name sequence is "Takahashi City" (medium place name) - "Asahi Town" (lower place name) - "Okayama City" (medium place name) - "Takahashi City" (medium place name) - "Yamada District" (subordinate place name). The order of place names in this place name string matches the order of appearance in the original text data. Then, the lower place name extracting unit 24 determines which intermediate place name the extracted lower place name belongs to. As mentioned above, the intermediate place name immediately before the appearance of the lower place name is the intermediate place name to which the lower place name belongs. That is, in this example, the lower place name extraction unit 24 determines that the lower place name "Asahi Town" belongs to the intermediate place name "Takahashi City" that appears immediately before. Further, the lower place name extraction unit 24 determines that the lower place name "Yamada District" also belongs to the intermediate place name "Takahashi City" that appears immediately before. In other words, when processing is performed based on the example text data in FIG. The relationship of district (subordinate place name) is extracted.

FIG. 8 is a schematic diagram showing a specific example of the expression of the influence range extracted by the influence range extraction unit 25. FIG. 8(A) shows part of the text data acquired by the text acquisition unit 21. The text data shown here is only the middle part of a certain sentence. The middle part says, ``A total of 11 people in three households were advised to evacuate.'' FIG. 8(B) shows the result of morphological analysis processing corresponding to the text of FIG. 8(A). As shown in the figure, the result of the morphological analysis process has information on each morpheme extracted from the original text and the corresponding part of speech. Note that in FIG. 8(B), numbers are assigned to each morpheme (each row of this table) for convenience. The influence range extracting unit 25 finds an expression of the influence range based on the above-described rule for extracting the influence range expression while referring to the data representing the result of the morphological analysis. As shown in Figure 8(B), the result of the morphological analysis process for this text example is (number 1, morpheme: , part of speech: particle) - (number 2, morpheme: 3, part of speech: numeral) - (number 3, Morpheme: household, part of speech: noun) - (number 4, morpheme: no, part of speech: particle) - (number 5, morpheme: combination, part of speech: verb) - (number 6, morpheme: te, part of speech: particle) - (Number 7, morpheme: 11, part of speech: number) - (number 8, morpheme: person, part of speech: noun) - (number 9, morpheme: ni, part of speech: particle) - (number 10, morpheme: evacuation advisory, part of speech: (noun) - (number 11, morpheme: is, part of speech: particle).

The influence range extraction unit 25 matches the above-mentioned morphological analysis processing result with the pattern prescribed by the above-mentioned "Rule 1", and as a result, two influence ranges of "3 households" and "11 people" are extracted. Extract the expression. Furthermore, in this example, the aforementioned "Rule 2" and "Rule 3" are not applied.

FIG. 9 is a schematic diagram showing another specific example of the expression of the influence range extracted by the influence range extraction unit 25. FIG. 9(A) shows part of the text data acquired by the text acquisition unit 21. The illustrated text data is only the middle part of a certain sentence, ``And there are two houses.'' FIG. 9(B) shows the result of morphological analysis processing corresponding to the text in FIG. 9(A). As shown in Figure 9(B), the result of the morphological analysis process for this text example is (number 1, morpheme:, part of speech: particle) - (number 2, morpheme: 2, part of speech: numeral) - (number 3, morpheme: ridge, part of speech: noun) - (number 4, morpheme: of, part of speech: particle) - (number 5, morpheme: house, part of speech: noun) - (number 6, morpheme: ga, part of speech: particle) be.

The influence range extraction unit 25 performs matching between the above morphological analysis processing result and the pattern prescribed by the above-mentioned "Rule 1", and as a result, extracts the expression of the influence range "2 buildings". Furthermore, in this example, the aforementioned "Rule 2" and "Rule 3" are not applied.

FIG. 10 is a schematic diagram showing an example of a sentence dependency analysis result for explaining the processing of the subject extraction unit 26. The data shown here is the dependency analysis processing for the text (example sentence): ``Due to this heavy rain, an evacuation advisory has been issued to 8 people in 3 households in Asahi-cho, Takahashi City./EOS.'' represents the result. Note that "EOS" is a symbol indicating the end of a sentence. The data of the dependency analysis results can be obtained, for example, in the form of plain text, but in FIG. 10, for convenience, line numbers are attached and horizontal lines are drawn to separate sections.

In FIG. 10, the number following the asterisk at the beginning of each clause is an integer value (clause number) that identifies the clause. The first clause has a clause number of 0, followed by 1, 2, 3, and so on. The integer value followed by "D" next to the clause number of each clause is the clause number to which the clause relates. If the clause number of the relevant clause is "-1", there is no clause to which the clause applies. Furthermore, for each word that makes up each clause, information about the word (part of speech information, reading, etc.) is described. The dependency relationships represented by this data are as follows. The clause with clause number 0 is "kono" and relates to the clause with clause number 1. The verse with verse number 1 is ``due to heavy rain,'' and relates to the verse with verse number 6. The clause with clause number 2 is "in Asahi-cho, Takahashi City" and relates to the clause with clause number 3. The clause with clause number 3 is "3 households" and relates to the clause with clause number 4. The verse with verse number 4 is "to eight people" and relates to the verse with verse number 6. The clause with clause number 5 is "Evacuation Recommendation" and relates to the clause with clause number 6. The clause with clause number 6 is "It is being issued." and there is no dependent clause ("-1" is written as the clause number).

Then, the theme extraction unit 26 extracts the theme according to the following procedure.
First step: First, by referring to the results of the dependency analysis, the theme extraction unit 26 combines the clauses into one if a clause is directly related to the next clause. In other words, in the example shown in Figure 10, there are four locations that directly relate to the next clause: clause numbers 0 to 1, clause numbers 2 to 3, clause numbers 3 to 4, and clause numbers 5 to 6. be. To summarize these clauses related to the next clause, the sentence is: ``Due to this heavy rain, an evacuation advisory has been issued to eight people in three households in Asahi-cho, Takahashi City.'' location). This group may be called a "chunk".

Second stage: Next, the subject extraction unit 26 extracts groups that include verbs from among these groups. However, we exclude verbs that are unlikely to be the subject, such as ``yoru'' and ``awasu''. Note that information on whether a certain verb is a verb that is unlikely to be a subject may be stored in the dictionary data in advance, and the dictionary may be referred to. As a result, only "an evacuation advisory has been issued" is extracted as a group that includes verbs that can be the subject.

Third stage: Next, the theme extraction unit 26 connects clauses that directly include clauses containing verbs. However, connecting clauses are selected only within a range that does not exceed clauses that include phrases extracted as expressions of place names or ranges of influence. In the example shown in FIG. 10, the clause "to 8 people" is directly related to "is issued" which includes a verb, but since it is an expression that has already been extracted by the influence range extraction unit 25, the theme is extracted. Section 26 does not connect to a verb. In addition, the other parts, ``Due to this heavy rain,'' and ``Three households in Asahi-cho, Takahashi City,'' are not directly related to the verb ``Issuedemasu'', so the theme extraction unit 26 extracts these. not extracted.

In addition, in the case of other sentence examples, the process by which the subject extraction unit 26 extracts the subject is as follows. For example, if the original sentence is "We have set up evacuation shelters in a total of 12 locations and are calling for people to evacuate," the theme extraction unit 26 follows the above procedure and selects the theme "We have set up evacuation shelters in 12 locations and are calling for people to evacuate." , We are asking people to evacuate.” In addition, if the original sentence is "We are distributing blue sheets for use in first aid to residents whose homes have been damaged," the subject extraction unit 26 selects "Blue Sheets to be used for first aid to residents whose homes have been damaged" as the subject according to the above procedure. Two items are extracted: ``Used for first aid'' and ``We are distributing blue sheets for use in first aid.''

Then, the theme extraction unit 26 performs processing to format the extracted theme according to the following procedure.
First stage: If the extracted verb is a specific one, the verb is removed and reformatted into a formal form. Specific verbs include, for example, "say,""naru,""send," and the like. Information on verbs to be removed in the formatting stage may be stored in advance as dictionary data, and the subject extraction unit 26 may refer to the dictionary data to determine whether or not the verb is to be removed. For example, when the expression "An evacuation advisory has been issued." is extracted as a theme, the verb "to issue" should be removed in order to reformat it into a formal expression, so the theme extraction unit 26 Remove all words after the noun "recommendation" immediately before the verb "put out". That is, the subject extraction unit 26 formats the expression "An evacuation advisory has been issued." into the expression "evacuation advisory."

Second stage: Remove expressions that are unsuitable for summarized data, such as conjunctions and the auxiliary verb "masu". When the extracted theme expression includes a specific conjunction or auxiliary verb, the theme extraction unit 26 leaves the verb immediately before it and removes all the sentences after it. For example, the subject extraction unit 26 formats the expression ``We are setting up an evacuation center and calling on people to evacuate'' into the expression ``We are setting up a shelter and calling on people to evacuate.''

Stage 3: Remove the word following the verb. However, if the verb is followed by an auxiliary verb expressing negation, the words up to the auxiliary verb are left and the words following the auxiliary verb are removed. Examples of auxiliary verbs expressing negation include "nai" and "nu". For example, if the extracted theme is "There is no hope for recovery," the theme extraction unit 26 leaves the auxiliary verb "nai" (negation) following the verb "to stand," and then Remove the word. In other words, the subject extraction unit 26 formats the expression "There is no hope for recovery" into "There is no hope for recovery."

FIG. 11 is a flowchart showing the procedure of processing by the summary generation unit 27. The process will be explained below along with this flowchart.

In step S21, the summary generation unit 27 removes duplicate content from among multiple themes extracted from one sentence. In the process of the subject extraction unit 26 described above, a plurality of subjects may be extracted from one sentence. The summary generation unit 27 may remove some of these themes depending on the degree of word overlap between the themes. Specifically, the summary generation unit 27 calculates the degree of word overlap between the extracted themes. The summary generation unit 27 calculates the ratio of words included in another subject to the words included in a certain subject as the degree of overlap. If there is a pair of themes with a degree of overlap of 60% or more, the summary generation unit 27 retains only the theme that includes more words from the pair and removes the other theme.

In step S22, the summary generation unit 27 associates place names and influence ranges. Here, the place name is a place name extracted by the intermediate place name extraction unit 22 or the lower place name extraction unit 24. Further, the influence range is an expression of the influence range extracted by the influence range extracting unit 25. In the original sentence, if single or multiple expressions of influence area appear after one place name appears and before another place name appears, those expressions of influence area are It can be assumed that the event occurred at the location of the place name. Thereby, the summary generation unit 27 associates the place name with the expression of the influence range.

In step S23, the summary generation unit 27 associates the subject with (place name+range of influence). In a sentence, when (place name + range of influence) appears and the next theme appears after that, the summary generation unit 27 associates (place name + range of influence) with that theme.

Through the processing from steps S21 to S23, the summary generation unit 27 identified a set of (place name, range of influence, subject). The place name, influence range, and subject in a certain set are associated in the processing of steps S22 and S23. Note that the summary generation unit 27 may specify a plurality of pairs (place name, range of influence, subject) based on one sentence. Furthermore, in the set (place name, range of influence, subject), there may be cases where a place name is not included or an expression of the range of influence is not included.

In step S24, the summary generation unit 27 generates a summary expression by applying the rules read from the generation rule storage unit 33 to each of the pairs (place name, range of influence, subject). Note that the rules for generating the summary will be explained next.

FIG. 12 is a schematic diagram showing an example of rules for generating a summary expression, which are stored in the generation rule storage unit 33. Here, although the figure shows three rules, rules 1 to 3, the number of rules stored in the generated rule storage section 33 is arbitrary. Each rule is described as a pair of a condition that is a prerequisite for application and a summary expression that is generated when the condition is met.

The condition of Rule 1 is that the set includes a middle place name. This rule 1 stipulates that when this condition is satisfied, the summary generation unit 27 should generate the summary expression "<subject><median place name><range of influence>".
The condition of Rule 2 is that the set includes a medium place name and a low place name. Rule 2 specifies that when this condition is satisfied, the summary generation unit 27 should generate a summary expression of "<subject><medium place name><list of low order place names>".
The condition of Rule 3 is that the set does not include a place name. Rule 3 specifies that when this condition is satisfied, the summary generation unit 27 should generate the summary expression "<subject>".

Examples of summary sentences generated by applying each of the above rules 1 to 3 are as follows.
An example of the summary generated by the summary generation unit 27 by applying Rule 1 is "Evacuation Recommendation: Tamano City, 2398 households, 8961 people." Another example of the summary that the summary generation unit 27 generates by applying Rule 1 is "Tamano City, where the risk of landslides is increasing." Note that this summary does not include an expression of the scope of influence.
An example of the summary generated by the summary generation unit 27 by applying Rule 2 is "evacuation advisory Tamano City Hibichi District Wada District Shibukawa District". Another example of the summary generated by the summary generation unit 27 by applying Rule 2 is "Murakami City, Former Shiota-cho Elementary School, Shimokita Yurika Hall, Senpoku Hall."
An example of a summary that the summary generation unit 27 generates by applying Rule 3 is "Establish an evacuation shelter and call for evacuation." Another example of the summary that the summary generation unit 27 generates by applying Rule 3 is "There is no hope for recovery due to the risk of secondary disaster."

FIG. 13 is a flowchart showing the overall processing procedure of the summarization device 1. The following will explain along this flowchart.

In step S31, the text acquisition unit 21 acquires text data from the outside. This text data is, for example, a news manuscript for a news program.
In step S32, the intermediate place name extraction unit 22 extracts intermediate place name candidates from the text data acquired by the text acquisition unit 21.
In step S33, the topic area identification unit 23 identifies a topic area based on the list of intermediate place name candidates extracted by the intermediate place name extraction unit 22. The topic area specifying unit 23 outputs information on top place names and middle place names for the specified topic areas.
In step S34, the lower stagnation extraction unit 24 extracts lower place names from the text data acquired by the text acquisition unit 21.
In step S35, the influence range extraction unit 25 extracts the expression of the influence range from the text data acquired by the text acquisition unit 21.
In step S36, the subject extraction unit 26 extracts a subject from the text data acquired by the text acquisition unit 21.
In step S37, the summary generation unit 33, based on the combination of the place name, expression of influence area, and subject of the topic area specified by the topic area identification unit (some information in the group may be missing), A summary is generated by applying the rules read from the generation rule storage unit 33.

Note that the text of the summary generated by the summarization device 1 can be used, for example, as data for broadcasting or web distribution. The text of the summary generated by the summarization device 1 may be automatically distributed, or only the text appropriately selected by a person or the like from among the texts may be distributed.

The present embodiment and its modified examples are summarized as follows.

(1) The summarization device 1 includes at least a text acquisition section 21, a place name extraction section, an influence range extraction section 25, a subject extraction section 26, and a summary generation section 27. The text acquisition unit 21 acquires text data. The place name extraction unit extracts place names from the text data by performing morphological analysis processing or by referring to place name information stored in advance (for example, place name dictionary storage unit 31 or adjacency relationship dictionary storage unit 32). Extract. The influence range extracting unit 25 extracts expressions including numerals included in the text data as expressions of the influence range. The theme extraction unit 26 extracts expressions containing verbs included in the text data as themes. The summary generation unit 27 generates a summary by applying a pre-stored generation rule to the set of the place name, the expression of the influence range, and the subject. However, the above set allows for the omission of at least either the place name or the expression of the range of influence.

Note that, as a modification, the place names do not have to be structured as a hierarchy of upper-middle-lower. Even in that case, the place name extraction unit can extract place names from the text data by performing morphological analysis processing or by referring to place name information. The summary generation unit 27 may generate a summary based on the place names extracted in this way.

(2) The place name may include a high-level place name, a middle-level place name, and a low-level place name as a hierarchical structure. In this case, the place name information includes affiliation information that is information about the intermediate place name that belongs to the higher place name. Further, the summarization device 1 further includes a topic area identifying unit 23 that identifies a topic area based on the place name extracted by the place name extracting unit (medium place name extracting unit 22, etc.). The topic area specifying unit 23 estimates the top place names that the text data is a topic of, based on the set of middle-rank place name candidates extracted by the place name extraction unit, so that the area represented by the top place name is This is to identify the hot topic area. That is, based on the affiliation relationship information, the topic area identification unit 23 determines, from a set of middle-rank place names (candidates) that appear in the text data, to which high-rank place name the middle-rank place name that is an element of the set tends to belong. , and thereby identify top place names as topical areas. Here, the number of higher place names may be one or more. In other words, even if a mid-rank place name happens to belong to multiple high-rank place names (for example, different municipalities with the same name belong to multiple prefectures), the above mid-rank place name (candidate) By statistically understanding in which top place names the set is distributed, the topic area specifying unit 23 can specify the topic area. An example of the calculation method for this purpose has already been described using equation (1) or equation (2), but the topic area may be specified using other calculation methods.

(3) The place name information may include adjacency relationship information indicating whether or not the pair of high-rank place names are in an adjacency relationship. The topic area specifying unit 23 determines whether the plurality of estimated high-ranking place names are adjacent to each other by referring to the adjacency relationship information. Thereby, the topic area identifying unit 23 can also specify the topic area expressed as a set of a plurality of high-rank place names obtained by transitively applying the relationship between pairs of high-rank place names having an adjacency relationship.

(4) As a specific example, the topic area identification unit 23 calculates the score of the top place name based on the set of middle place name candidates extracted from the text data by referring to the place name information. Then, the top place names with relatively good scores are specified as the topic areas. Based on the affiliation relationship information, the score of this high-ranking place name is the reciprocal of the number of all high-ranking place names to which the mid-ranking place name candidate belongs for each of the mid-ranking place name candidates belonging to the high-ranking place name. It is calculated as the sum of all of the intermediate place name candidates belonging to the place name.

As a modification, the likelihood that the top place name is a hot topic area may be calculated using another statistical calculation method.

(5) As already specifically explained, the lower place name extraction unit 24 (also simply called the “place name extraction unit”) extracts the upper place names and the above place names from among the place names extracted by performing the morphological analysis process. A place name that is not specified in any of the intermediate place names is extracted as the lower place name, and the lower place name is used as the lower place name among the intermediate place names that appear in the text data before the lower place name. The place name may be extracted as a lower place name belonging to the intermediate place name that appears at the position closest to .

(6) As already explained in detail, this lower place name extraction unit 24 further extracts nouns (for example, “Yamada district”) accompanied by specific keywords (“district”, “region”, etc.) that make it possible to infer that it is a place name. '') is further extracted as the lower-level place name, and the lower-level place name appears in the position closest to the lower-level place name among the middle-level place names that appear in the text data before the lower-level place name. The place name may be extracted as a lower place name belonging to the intermediate place name.

According to the embodiment described above, the summarization device 1 can output a summary based on text data such as a news manuscript. Further, the summarization device 1 does not output a summary using a model obtained by machine learning, but performs a process of generating a summary based on rules and algorithms. In other words, the summary output by the summarization device 1 is not obtained depending on machine learning data, but is highly predictable. In addition, the summarization device 1 is not limited to the number of sentences to output, and can output multiple (in some cases, many) summary sentences, that is, it can output a summary sentence that covers all the information contained in the input text data. Summary data can be output.

In addition, the summarization device 1 has information on a set of middle-rank place names that belong to high-rank place names in the place name dictionary storage unit 31 in advance, and processes the set of middle-rank place names that appear in the text by referring to the place name dictionary storage unit 31. Based on the set of place name candidates, it is possible to appropriately identify topic areas (top place names that are the subject of topics). In addition, the summarization device 1 stores information on whether or not the top place names are adjacent to each other in the adjacency relationship dictionary storage unit 32 in advance, and performs processing while referring to the adjacency relationship dictionary storage unit 32. Place names can be identified as a group of topical areas.

By using the summarization device 1, there is no need to manually create text exclusively for summary data, and summaries automatically generated based on text data with news manuscripts can be transmitted over communication networks, on TV broadcast screens, or with subtitles. It will be possible to distribute via text or digital signage installed in various locations. For example, even in situations where a wealth of summary data is required, such as during a disaster, there is no need to allocate personnel for this purpose.

Note that at least part of the functions of the summarization device in the above-described embodiments (including modified examples) can be realized by a computer. In that case, a program for realizing this function may be recorded on a computer-readable recording medium, and the program recorded on the recording medium may be read into a computer system and executed. Note that the "computer system" herein includes hardware such as an OS and peripheral devices. Furthermore, "computer-readable recording media" refers to portable media such as flexible disks, magneto-optical disks, ROM, CD-ROM, DVD-ROM, and USB memory, and storage devices such as hard disks built into computer systems. Say something. Furthermore, a "computer-readable recording medium" refers to a medium that temporarily and dynamically stores a program, such as a communication line when transmitting a program via a network such as the Internet or a communication line such as a telephone line. In that case, it may also include something that retains a program for a certain period of time, such as a volatile memory inside a computer system that is a server or client. Further, the program may be one for realizing a part of the above-mentioned functions, or may be one that can realize the above-mentioned functions in combination with a program already recorded in the computer system.

Although a plurality of embodiments have been described above, the present invention can be further implemented in the following modifications.
For example, the above embodiment can be implemented in such a way that each functional unit performs morphological analysis processing. On the other hand, the morphological analysis process may be performed on the original text data only once, and each part may refer to the result.

Although the embodiments of the present invention have been described above in detail with reference to the drawings, the specific configuration is not limited to these embodiments, and includes designs within the scope of the gist of the present invention.

INDUSTRIAL APPLICATION This invention can be utilized, for example in a broadcasting business, a content distribution business, etc. However, the scope of use of the present invention is not limited to what is exemplified here.

1 Summarization device 21 Text acquisition unit 22 Medium place name extraction unit (place name extraction unit)
23 Topic area identification unit 24 Lower place name extraction unit (place name extraction unit)
25 Influence range extraction unit 26 Theme extraction unit 27 Summary generation unit 31 Place name dictionary storage unit (storage unit that stores place name information)
32 Adjacency dictionary storage unit 33 Generation rule storage unit

Claims (6)

a text acquisition unit that acquires text data;
a place name extraction unit that extracts place names from the text data by performing morphological analysis processing or by referring to place name information stored in advance;
The pattern of "{noun or prefix} + numeral (+conjunctive particle + numeral) (+{{noun, but not an adjective stem} or parallel particle or adjective})" included in the text data (however, the pattern enclosed in parentheses is an influence range extraction unit that extracts an expression (the part that is displayed can be omitted) as an expression of the influence range;
a theme extraction unit that extracts, as a theme, an expression obtained by connecting clauses directly related to clauses containing verbs included in the text data;
A set of the place name, the expression of the range of influence, and the subject, which is stored in advance in the set that allows for the omission of at least either the place name or the expression of the range of influence. a summary generator that generates a summary by applying a generation rule;
a topic area identification unit that identifies a topic area based on the place name extracted by the place name extraction unit;
Equipped with
The place name includes a high-level place name, a middle-level place name, and a low-level place name as a hierarchical structure,
The place name information includes affiliation information that is information about the intermediate place name that belongs to the high-level place name,
The topic area identification unit estimates the top place names that the text data is a topic of, based on the set of intermediate place name candidates extracted by the place name extraction unit, so that the text data is represented by the top place name. It specifies the topic area,
The summary generation unit generates the summary using the topic area identified by the topic area identification unit.
Summarization device. The place name information includes adjacency relationship information indicating whether or not the pair of high-rank place names are in an adjacency relationship;
The topic area identification unit refers to the adjacency relationship information to determine whether or not the plurality of estimated high-rank place names are adjacent to each other, and determines the relationship between pairs of high-rank place names that have an adjacency relationship. The topic area is specified as a set of a plurality of top place names obtained by transitively applying
A summarization device according to claim 1 . The topic area identification unit calculates scores of top place names based on a set of middle place name candidates extracted from the text data by referring to the place name information, and selects top place name candidates with relatively good scores. The top place names are identified as the topic areas,
The score of the above-mentioned high-rank place name is calculated based on the above-mentioned affiliation relationship information, based on the above-mentioned affiliation relationship information, the score of the above-mentioned high-rank place name is the reciprocal of the number of all high-rank place names to which the mid-rank place name candidate belongs for each of the mid-rank place name candidates belonging to the high-rank place name. The value is calculated as the sum of all of the intermediate place name candidates belonging to the place name.
A summarization device according to claim 1 or 2 . The place name extracting unit extracts, as the lower place name, a place name that is not specified in either the upper place name or the middle place name from among the place names extracted by performing the morphological analysis process, and Extracting the lower place name as a lower place name belonging to the intermediate place name that appears in the position closest to the lower place name among the intermediate place names that appear in the text data prior to the lower place name ,
The summary generation unit generates the summary of "<Subject><Medium place name><List of low order place names>" when the set includes a middle place name and a low order place name.
A summarization device according to any one of claims 1 to 3 . The place name extraction unit further extracts a noun accompanied by a specific keyword that makes it presumed to be a place name as the lower place name, and the lower place name appears in the text data before the lower place name. Extracting as a lower place name belonging to the intermediate place name that appears in the position closest to the lower place name among the intermediate place names,
A summarization device according to claim 4 . computer,
a text acquisition unit that acquires text data;
a place name extraction unit that extracts place names from the text data by performing morphological analysis processing or by referring to place name information stored in advance;
The pattern of "{noun or prefix} + numeral (+conjunctive particle + numeral) (+{{noun, but not an adjective stem} or parallel particle or adjective})" included in the text data (however, the pattern enclosed in parentheses is an influence range extraction unit that extracts an expression (the part that is displayed can be omitted) as an expression of the influence range;
a theme extraction unit that extracts, as a theme, an expression obtained by connecting clauses directly related to clauses containing verbs included in the text data;
A set of the place name, the expression of the range of influence, and the subject, which is stored in advance in the set that allows for the omission of at least either the place name or the expression of the range of influence. a summary generator that generates a summary by applying a generation rule;
a topic area identification unit that identifies a topic area based on the place name extracted by the place name extraction unit;
Equipped with
The place name includes a high-level place name, a middle-level place name, and a low-level place name as a hierarchical structure,
The place name information includes affiliation information that is information about the intermediate place name that belongs to the high-level place name,
The topic area identification unit estimates the top place names that the text data is a topic of, based on the set of intermediate place name candidates extracted by the place name extraction unit, so that the text data is represented by the top place name. It specifies the topic area,
The summary generation unit generates the summary using the topic area identified by the topic area identification unit.
A program to function as a summarization device.

Images