JP2018092367A - Related word extracting device and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To enable an extraction of a related word regardless of a co-occurrence and a concept dictionary.SOLUTION: A time-series text storing unit 122 of a related word extracting device 10 stores a time-series text that is text data in association with time information, and a time-series frequency extracting unit 124 extracts an appearance frequency of each letter string in time series contained in the time-series text. A correlation value calculating unit 125 calculates, for each of the different letter strings, a correlation value that is a value indicating a strength of a causal relation between the different letter strings in a quantitative manner using the appearance frequency in time series extracted by the time-series frequency extracting unit 124, and determines whether or not those different letter strings are related to each other on the basis of the calculated correlation value. An output unit 13 outputs, as the related word, the other letter string determined as being related to the letter string of a keyword by the correlation value calculating unit 125.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a related word extraction device and a program.

  In conventional related word presentation techniques, a related word of a word is calculated using an index based on the co-occurrence of the word (see, for example, Patent Document 1). Related word presentation techniques based on co-occurrence of such words include dice coefficients and word2vec using a neural network. In this related word presentation technique, the degree of relevance represented by the index does not increase unless mutual words appear in the same document.

  On the other hand, there is a related word presentation technique using a concept dictionary (for example, see Patent Document 2). In this related word presentation technique, it is necessary to prepare a concept dictionary in advance. In the existing concept dictionary, words having some known relations such as upper and lower relations and causal relations are listed.

Japanese Patent No. 3598211 JP2015-130111A

  In related word presentation technology based on co-occurrence, words having similar meanings tend to be presented as related words in many cases. In reality, however, “pool” and “morning glory” have different meanings and fields, but each year is a characteristic word in summer, and “Kagamine” and “kotatsu” are also characteristic words in winter. There seems to be a relationship between these words. The related word presentation technology based on co-occurrence cannot be obtained as related words unless these words appear in the same document. In addition, there are many cases in which an explicit relationship cannot be expressed between these two words, and it is difficult to be included in an existing concept dictionary. Therefore, the related word presentation technique using the concept dictionary cannot be obtained as a related word.

  There is a DTM (Dynamic Topic Model) as a model for obtaining a word occurrence probability based on a topic in consideration of time series. However, since DTM does not correspond to a period such as this time of year, it is necessary to specify periodicity using a discrete Fourier transform or the like. Moreover, a certain word does not always appear in the same cycle. For example, there are some words that appear after an incident, and after that, the words appear more frequently, and the word that the trend of the topic ends at some point stops appearing.

  The present invention has been made in view of such circumstances, and provides a related word extraction apparatus and program capable of extracting related words without using word co-occurrence or concept dictionaries.

One aspect of the present invention is a related word extraction device that extracts related words of each character string included in text data, the input unit for inputting a predetermined keyword, and the text data associated with time information. A storage unit for storing a certain time series text, a time series frequency extraction unit for extracting a time series frequency that is an appearance frequency when each character string included in the time series text appears in a time series, and the time series Using the time series frequency extracted by the frequency extraction unit, a correlation value that is a value that quantitatively represents the strength of the causal relationship between the different character strings is calculated, and the correlation value is different based on the calculated correlation value A correlation value calculating unit that determines whether or not the character string is related; and an output unit that outputs the other character string that the correlation value calculating unit determines to be related to the character string of the keyword as a related word; Specially prepared It is a related term extraction apparatus to be.
According to this invention, the related word extracting device calculates the appearance frequency of the time series of each character string with reference to the time series text that is text data associated with the time information. The related word extraction device determines whether or not there is a causal relationship between character strings using the calculated time-series appearance frequency, and outputs a character string determined to have a causal relationship with a keyword as a related word.
As a result, the related word extraction device presents words related to events that have the same time as keywords and events that have a causal relationship with keywords as related words without using word co-occurrence or concept dictionaries. can do.

One aspect of the present invention is the related word extraction device described above, wherein a content information storage unit that stores content text information that is text data indicating information related to content, and the content text information is searched using the related words. The content search unit further includes a content search unit that outputs the content information specified by the content text information obtained as a result of the search by the content search unit.
According to this invention, the related word extraction device searches for content using the related word of the keyword, and presents information on the content specified by the search to the user.
According to the present embodiment, the related word extraction device can present content related to the keyword to the user.

One aspect of the present invention is the related word extraction device described above, in which the storage unit includes the first time-series text including the text data and the time information included in a web page on the Internet, and information related to the content. And a second time-series text including time information, and the related word extraction device further includes a keyword extraction unit that extracts a keyword from the first time-series text, and the time series The frequency extraction unit extracts a time series frequency of the keyword included in the first time series text and a time series frequency of each character string included in the second time series text, and the correlation value calculation unit Is a correlation between the keyword and the character string using the time series frequency extracted by the time series frequency extraction unit for each of the keyword and the character string. And determining whether or not the character string is related to the keyword based on the calculated correlation value, and the output unit is determined to be related to the keyword by the correlation value calculation unit A character string is output as a related word representing the theme of the content, and information on the content specified by the second time-series text including the output related word is output.
According to the present invention, the related word extracting device extracts words that are hot topics on the Internet as keywords, and each character string obtained from the time-series appearance frequency of the keywords on the Internet and the information related to the contents. The related word of the keyword is acquired from the information related to the content using the appearance frequency of the time series. The related word extraction device searches the content using the acquired related words, and presents to the user the content theme indicated by the related words and the content information specified by the time-series text from which the related words are obtained. To do.
Thereby, the related word extraction apparatus can present to the user content related to an event that is a topic on the Internet and the theme of the content.

  One embodiment of the present invention is a program for causing a computer to function as any one of the related word extraction devices described above.

  According to the present invention, related words can be extracted regardless of word co-occurrence or concept dictionary.

It is a functional block diagram which shows the structure of the related word extraction apparatus by the 1st Embodiment of this invention. It is a figure which shows the example of the time series character string table by the embodiment. It is a figure which shows the example of the related word table by the embodiment. It is a flowchart which shows operation | movement of the related word extraction apparatus by the same embodiment. It is a flowchart which shows the time series character string table preparation process of the time series frequency extraction part by the embodiment. It is a figure showing the example of the time series frequency of two character strings. It is a functional block diagram which shows the structure of the related word extraction apparatus by 2nd Embodiment. It is a flowchart which shows operation | movement of the related word extraction apparatus by the same embodiment. It is a functional block diagram which shows the structure of the related word extraction apparatus by 3rd Embodiment. It is a functional block diagram which shows the structure of the related word extraction apparatus by 4th Embodiment. It is a functional block diagram which shows the structure of the related word extraction apparatus by 5th Embodiment. It is a flowchart which shows operation | movement of the related word extraction apparatus by the same embodiment.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

[First Embodiment]
In the present embodiment, related words of keywords specified by the user are extracted.
FIG. 1 is a functional block diagram showing the configuration of the related word extraction apparatus 10 according to the first embodiment, and shows only functional blocks related to the present embodiment. The related word extraction device 10 includes an input unit 11, a related word extraction unit 12, and an output unit 13. The related word extraction unit 12 includes a time series text storage unit 122 and a time series character string table storage unit 123. Prepare.

  The related word extraction device 10 stores time series text that is text data associated with time information, and extracts the appearance frequency of each character string included in the time series text. Then, the related word extraction device 10 calculates a correlation value, which is a value that quantitatively represents the strength of the causal relationship between different character strings, using the appearance frequency of the time series for each of the different character strings, It is determined whether or not these different character strings are related based on the calculated correlation value, and different character strings determined to be related are used as related words. When a predetermined keyword is input, the related word extraction device 10 outputs another character string determined to be related to the character string of the keyword as a related word.

  The input unit 11 receives input data such as a keyword from a terminal connected via a network, for example. The input unit 11 may receive input data from an input device such as a keyboard provided in the related word extraction device 10 or may read input data from a computer-readable storage medium.

  The related word extraction unit 12 extracts a related word of the keyword input by the input unit 11. The related word extraction unit 12 includes a storage unit 121, a time series frequency extraction unit 124, and a correlation value calculation unit 125.

  The storage unit 121 includes a time-series text storage unit 122 and a time-series character string table storage unit 123, and stores various data. The time series text storage unit 122 may be provided in a server device connected to the related word extraction device 10 via a network.

  The time series text storage unit 122 stores time series text that is document data including text data associated with time information. The time series text includes, for example, news articles, web texts, program subtitles, an electronic program guide (EPG), and the like. For example, when the time series text is a news article, the text data includes the content of the news article, and the time information includes information on the date and time included in the content of the news article, or the creation date and release date and time of the news article. . In addition, when the time series text is a blog published on the web on the Internet, the text data includes a sentence of the posted content of the blog, and the time information includes the posted date / time or published date / time of the blog. Furthermore, when the time-series text is a program subtitle, the text data may include the content of the subtitle, and the time information may be the broadcast start date and time of the program or the subtitle presentation time. When the time series text is an electronic program guide, the text data includes a program name, a program subtitle, a program summary, and the like, and the time information includes the broadcast start date and time of the program.

  The time-series character string table storage unit 123 stores a time-series character string table that is data indicating the frequency of appearance of each character string (hereinafter also simply referred to as “frequency”). The frequency may be the number of time-series texts in which character strings appear, or may be the number of character strings that appear in all time-series texts. The user may set which frequency is used.

The time series frequency extraction unit 124 calculates the time series frequency of each character string from the time series text stored in the time series text storage unit 122, and creates a time series character string table.
The correlation value calculation unit 125 performs a correlation value calculation process for calculating a correlation value between character strings based on the time-series frequency of each character string indicated by the time-series character string table. Here, the correlation value is a value that quantitatively represents the strength of the causal relationship between the character strings.
The correlation value calculation unit 125 generates a related word table in which character strings determined to have a certain level of correlation based on the calculated correlation value are associated with each other. The correlation value calculation unit 125 extracts a related word of the input keyword from the related word table.

  The output unit 13 transmits data such as related words extracted by the correlation value calculation unit 125 to a terminal connected via a network and displays the data. The output unit 13 may display the output data on a display device such as a display provided in the related word extraction device 10 or may write the output data to a computer-readable storage medium.

  FIG. 2 is a diagram illustrating an example of a time-series character string table. The time-series character string table shown in the figure shows the frequency of n (n is 2 or more) different character strings W1, W2,..., Wn in units of 1 day from January 1, 2016.

  FIG. 3 is a diagram illustrating an example of a related word table. The related word table is data in which a character string is associated with a related word that is another character string related to the character string. The character string and the related word are character strings that appear in the time series text stored in the time series text storage unit 122.

<Operation of Related Word Extraction Device 10>
Next, the operation of the related word extraction device 10 will be described.
FIG. 4 is a flowchart showing the operation of the related word extraction device 10.
First, in step S105, the input unit 11 of the related word extraction device 10 receives an input of a keyword. For example, the input unit 11 of the related word extraction device 10 receives a keyword input by the user from the terminal and outputs the keyword to the related word extraction unit 12.
In step S110, the time-series frequency extraction unit 124 of the related word extraction unit 12 performs a time-series character string table creation process shown in FIG.

FIG. 5 is a flowchart showing the time-series character string table creation process of the time-series frequency extraction unit 124.
In step S <b> 205, the time-series frequency extracting unit 124 selects and reads one unselected time-series text stored in the time-series text storage unit 122.
In step S210, the time series frequency extraction unit 124 acquires time information from the selected time series text.

  In step S215, the time-series frequency extraction unit 124 cuts text data described in the time-series text in association with the acquired time information into character strings. The character string is, for example, a morpheme of a predetermined part of speech such as a noun, a character gram, and a noun pair. A pair of nouns used as character strings is a pair of a noun and a word related to the noun, such as “summer feature poem”. Conventional techniques such as morphological analysis and character gramm analysis can be used for dividing into character strings.

  In step S220, the time-series frequency extraction unit 124 assumes that the character string obtained by segmenting the text data associated with the time information appears at the time indicated by the time information, and the frequency of each different character string. Calculate The time-series frequency extraction unit 124 may set the frequency of each character string that appears at least once in the text data associated with the time information to “1”, and the character string is included in the text data associated with the time information. The number of occurrences of may be used as the frequency of the character string.

  If the time-series character string table has not been created, the time-series frequency extraction unit 124 generates a new time-series character string table and writes it in the time-series character string table storage unit 123. The unit time set in the time-series character string table is stored in the storage unit 121 in advance, and is variable according to user settings such as one month, one week, one day, and one hour.

  In step S225, the time-series frequency extraction unit 124 updates the time-series character string table stored in the time-series character string table storage unit 123 based on the frequency of each character string obtained in step S220.

  Specifically, if there is a character string that has not been registered in the time-series character string table among the character strings obtained in step S215, the time-series frequency extraction unit 124 additionally registers the character string. The time series frequency extraction unit 124 calculates the frequency of the character string set in the time series character string table corresponding to the time zone including the time information obtained in step S210, for the character string calculated in step S220. Update to a value that adds the frequency.

  In step S230, the time-series frequency extracting unit 124 determines whether or not the selected time-series text is the last time-series text stored in the time-series text storage unit 122. When the time series frequency extraction unit 124 determines that the selected time series text is not the last time series text stored in the time series text storage unit 122 (step S230: NO), the processing from step S205 is repeated. When the time-series frequency extraction unit 124 determines that the selected time-series text is the last time-series text stored in the time-series text storage unit 122 (step S230: NO), the related word extraction apparatus 10 The process from step S115 shown in FIG. 4 is performed.

  4, the time-series frequency extraction unit 124 outputs the time-series character string table stored in the time-series character string table storage unit 123 to the correlation value calculation unit 125. At this time, the correlation value calculation unit 125 extracts only the time-series frequency of the character string satisfying the condition from the time-series frequencies of each character string set in the time-series character string table, and the correlation value calculation unit 125. May be output. For example, the correlation value calculation unit 125 excludes information about low-frequency character strings from the time-series character string table. The low frequency is, for example, when ranking is performed in descending order of the total frequency in a predetermined period where the total frequency in the predetermined period is less than or equal to a predetermined value, the average frequency in the predetermined period is less than or equal to the average frequency of all character strings Is determined to be less than or equal to a predetermined order. Alternatively, use a topic model such as DTM (Dynamic Topic Model) to extract only the frequency of the top m character strings with high probability values in each topic. Burst the character string whose frequency varies with time Processing such as extraction by detection is conceivable. However, the time-series frequency extraction unit 124 may not delete information on the frequency of the character string that matches the keyword input in step S105.

  In step S120, the correlation value calculation unit 125 performs a correlation determination process for determining whether a causal relationship is established between the character strings set in the time-series character string table input from the time-series frequency extraction unit 124. . The correlation value calculation unit 125 may perform correlation determination processing for all character string sets set in the time-series character string table. The character string that matches the keyword input in step S105 and the keyword A correlation determination process may be performed between each of all character strings different from the above.

  The correlation value calculation unit 125 performs the Granger causality test using the time-series frequencies of the two character strings to be subjected to correlation determination processing. The Granger causality test is an existing method that is often used in economics, and is used to determine the existence of a causal relationship only from numerical data such as changes in sales over time, without using background knowledge. It is done. For example, by using the Granger causality test, it is possible to test the causal relationship only with time series data, such as high sales of ice cream → increased water accidents. The causal relationship includes the direction of the causal relationship between the two terms. However, the original meaning of an ice cream and water accident is not causal.

The frequency at time t of the character string Wi (i = 1, 2,..., N) is x _t , and the frequency at time t of the character string Wj (i ≠ j, j = 1, 2,..., N) is z _t . Then, consider the vector autoregressive model (VAR) model of the following equations (1) and (2).

_{x t = a 1 x t-} 1 + ... + a p x t-p + b 1 z t-1 + ... + b p z t-p + u 1t ... (1)

_{z t = c 1 x t-} 1 + ... + c p x t-p + d 1 z t-1 + ... + d p z t-p + u 2t ... (2)

a _i , b _i , c _i , d _i (i = 1, 2,..., p) are constants, and u _1t and u _2t are noises. Equation (1), the frequency _{x t} string Wi at time t, and frequency _{x t-1 ~x t-p} of the past string Wi, frequency of past string Wj _{z t-1} ~z _t It indicates that described in _-p, b 1 ~b _p is no causal relationship strings Wi and string Wj if 0. Equation (2), the frequency _{z t} string Wj at time t, and frequency _{z t-1 ~z t-p} past string Wj, frequency _{x t-1} ~x _t past string Wi It indicates that described in _-p, c 1 _{~c p} is not causal relationship string Wj and string Wi if 0.

The correlation value calculation unit 125 generates all sets of two different character strings from the character strings W1 to Wn, and performs the following processing with the character strings of each set as Wi and Wj. Incidentally, the correlation value calculation unit 125 obtains a frequency _{x t ~x t-p} the time series of strings Wi, the frequency _{z t ~z t-p} time series string Wj, from the time series string table To do.

The correlation value calculation unit 125 estimates Equation (1) by least square regression using the time series frequencies of the character strings Wi and Wj, and obtains the residual sum of squares USS _x of the estimated Equation (1). Further, the correlation value calculation unit 125 assumes b ₁ =... = B _p = 0, estimates Equation (1) by least square regression using the frequency of the character string Wi, and calculates the residual square of the estimated Equation (1). The sum RSS _x is calculated. Correlation value calculating section 125, the following equation (3) and the test statistic _{F x.}

F _x = ((RSS _x −USS _x ) / p) / (USS _x / (n−2p)) (3)

The correlation value calculation unit 125 performs a χ ² (chi-square) test that compares pF _x with the 95% point of χ ² (p). If pF _x is larger, the causal from the character string Wj to the character string Wi If the relationship exists and is small, it is determined that there is no causal relationship from the character string Wj to the character string Wi.

Similarly, correlation value calculation section 125 estimates equation (2) by least square regression using the frequencies of character strings Wi and Wj, and obtains residual square sum USS _y of estimated equation (2). Furthermore, correlation value calculation section 125 assumes c ₁ =... = C _p = 0, estimates equation (2) by least square regression using the frequency of character string Wj, and calculates the residual square of estimated equation (2). The sum RSS _y is calculated. The correlation value calculation unit 125 sets the following equation (4) as the test statistic _Fy .

_{F y = ((RSS y -USS} y) / p) / (USS y / (n-2p)) ... (4)

The correlation value calculation unit 125 performs a χ ² test that compares pF _y with the 95% point of χ ² (p). If pF _y is larger, there is a causal relationship from the character string Wi to the character string Wj. If it is smaller, it is determined that there is no causal relationship from the character string Wi to the character string Wj.

  When at least one of the causal relationship from the character string Wi to the character string Wj and the causal relationship from the character string Wj to the character string Wi exist, the correlation value calculation unit 125 converts the character string Wj into a related word of the character string Wi. And the character string Wi is a related word of the character string Wj. In the above description, the correlation value calculation unit 125 determines that the significance level is 95% and character string pairs that satisfy the level are causal, but the significance level is stored in the storage unit 121 in advance and is determined by the user. Set to variable.

In step S125, the correlation value calculation unit 125 creates, for each character string, a related word table that associates the character string with another character string that has been determined to have a causal relationship. At this time, the correlation value calculation unit 125 includes information on the correlation value expressed as a test statistic F _x or F _y that is a value that quantitatively represents the strength of the causal relationship between the character strings, and the direction of the causal relationship. And the related word is set in the related word table. Further, the correlation value calculation unit 125 may arrange related words in descending order of correlation values and set them in the related word table.

  The correlation value calculation unit 125 generates a keyword-related word table by extracting the keyword input in step S105 and the part where the related word corresponding to the keyword is set from the generated related word table, and outputs the keyword-related word table. 13 is output.

  In step S130, the output unit 13 presents the keyword-related word table extracted by the correlation value calculation unit 125. Specifically, the output unit 13 transmits the keyword-related word table to the keyword transmission source terminal, and the terminal displays the received keyword-related word table on the display.

  Instead of keywords, text data such as sentences may be input to the related word extraction device 10. The time series frequency extraction unit 124 or a keyword extraction unit (not shown) acquires a character string from the input text data, and uses the acquired character string as a keyword, as in the process of step S215 in FIG.

  When there are two or more keywords, the output unit 13 may increase the priority of related words with more related keywords, and display related words in order of priority. The color of the word may be changed and displayed, or the number of related keywords may be added to display the related word. Further, the output unit 13 may display the related words with information such as characters and figures representing the direction of the causal relationship between the keywords.

  In addition, the correlation value calculation unit 125 may set only the related words whose direction of the causal relationship with the keyword is in the predetermined direction in the related word table, and the related word whose direction of the causal relationship with the keyword is the predetermined direction from the related word table. May be extracted and set in the keyword-related word table.

  Note that when the process of FIG. 4 is repeated with different keywords, a time-series character string table may already be generated. In this case, in step S205 in FIG. 5, the time series frequency extraction unit 124 selects and generates one time series text newly registered in the time series text storage unit 122 after the previous processing in FIG. What is necessary is just to update the completed time series character string table.

  According to the related word extraction device 10 of the present embodiment, words (character strings) that become topics at the same time can be extracted as related words. Further, the related word extraction apparatus 10 can extract a word (character string) related to a certain event and a word (character string) related to an event that occurs at different times as a result of the occurrence of the event as related words. This will be described with reference to FIG.

  FIG. 6 is a diagram illustrating an example of time-series frequencies of two character strings Wi and Wj. When the event represented by the character string Wj occurs due to the occurrence of the event represented by the character string Wi, the frequency of the character string Wi and the frequency of the character string Wj appear at different timings. For example, when the economy recovers due to the rise in stock prices and the number of people going to travel abroad increases, the rise in stock prices and overseas travel are causally related, but they are conceptually different. In addition, frequency peaks occur at different times. In addition, since the frequency peaks are different, it is considered that there are few opportunities to co-occur in the same document. The related word extraction device 10 determines a causal relationship using the Granger causality test, thereby associating a word related to a causal event that has occurred in the past with respect to an event related to a keyword. Can be obtained as a word.

[Second Embodiment]
In this embodiment, the content searched using the related word extracted by the process similar to 1st Embodiment is recommended to a user. In the following, the difference from the first embodiment will be mainly described.

  FIG. 7 is a functional block diagram showing the configuration of the related word extracting apparatus 20 according to the second embodiment, and only the functional blocks related to this embodiment are extracted and shown. In this figure, the same parts as those in the related word extracting apparatus 10 according to the first embodiment shown in FIG. The difference between the related word extraction device 20 and the related word extraction device 10 shown in FIG. 1 is that a content information storage unit 21 and a content search unit 22 are further provided.

  The content information storage unit 21 stores the content ID and content text information of each piece of content data (hereinafter also referred to as “content”) to be introduced to the user in association with each other. The content ID is identification information that uniquely identifies the content, and the content text information is text data indicating information about the content. For example, the content text information may be text data indicating the title or description of the content, the content data itself including the text data, or all or part of the text data included in the content data.

  For example, if the content is a program, text data describing a program subtitle, a program summary sentence, or the like may be used as the content text information, or subtitle text included in the program may be used. If the content is music, electronic books, etc., text data describing the content introduction or summary is used as the content text information. Further, when the content is a web page or an electronic book containing text data, the content data itself or the whole or part of the text data included in the content may be used as the content text information.

  The content information storage unit 21 further stores content introduction information in association with the content ID. The content introduction information is information presented to introduce content to the user. For example, when the content is a program, the content introduction information includes a program name, broadcast date / time, channel, program overview, thumbnail image, URL (Universal Resource Locator) of a program viewing site, and the like. If the content is music, electronic books, etc., the content introduction information includes information such as the content title, introduction text, price, URL of the purchase site or download site, and if the content is a web page, the content introduction information The information includes information such as the URL of the web page. Part or all of the content text information may be used as the content introduction information.

  The content information storage unit 21 may be provided in a server connected to the related word extraction device 20 via a network.

<Operation of Related Word Extraction Device 20>
FIG. 8 is a flowchart showing the operation of the related word extraction device 20. In this figure, the same processes as those in the flowchart shown in FIG. The related word extraction device 20 performs the same processing as steps S105 to S125 of FIG.

  In step S305, the correlation value calculation unit 125 generates a keyword-related word table by extracting the keyword input in step S105 and the part where the related word corresponding to the keyword is set from the generated related word table. And output to the content search unit 22.

  In step S310, the content search unit 22 acquires a keyword from the keyword-related word table, and searches the content text information stored in the content information storage unit 21 using the acquired keyword as a search key. The content search unit 22 reads content introduction information associated with the same content ID as the content text information obtained as a result of the search from the content information storage unit 21.

  Subsequently, in step S315, the content search unit 22 acquires a related word from the keyword related word table, and searches the content text information stored in the content information storage unit 21 using the acquired related word as a search key. . The content search unit 22 reads content introduction information associated with the same content ID as the content text information obtained as a result of the search from the content information storage unit 21.

  The content search unit 22 outputs the content introduction information read in step S310 and the content introduction information read in step S315 to the output unit 13.

  In step S320, the output unit 13 presents the content introduction information received from the content search unit 22. Specifically, the output unit 13 transmits the content introduction information to the keyword transmission source terminal, and the terminal displays the received content introduction information on the display.

  When the number of keywords is two or more, the content search unit 22 may assign priorities so that the content related to more keywords is higher in order and display the content presentation information in order of priority. In addition, the content search unit 22 may prioritize the content so that the content related to more related words is higher in order and display the content presentation information in the order of priority. Further, when there are a plurality of contents related to the same number of related words, the content search unit 22 may determine the priority based on the result of summing or multiplying the correlation values between the related words and the keywords.

  In the present embodiment, for example, it is possible to acquire a keyword-related word using a vast amount of information published on the Internet and use it for content recommendation. In addition, it is possible to recommend to the user content related to an event that causes an event related to a keyword or content related to an event that is expected to occur next to an event related to a keyword.

[Third Embodiment]
In the second embodiment, the time-series text and the content text information are different data, but in this embodiment, the time-series text and the content text information are the same data. Hereinafter, differences from the second embodiment will be mainly described.

  FIG. 9 is a functional block diagram showing the configuration of the related word extracting device 30 according to the third embodiment, and only functional blocks related to the present embodiment are extracted and shown. In this figure, the same parts as those in the related word extracting apparatus 20 according to the second embodiment shown in FIG. The related word extracting device 30 and the related word extracting device 20 shown in FIG. 7 are different from the related word extracting unit 12, the content information storage unit 21, and the content search unit 22 in that the related word extracting unit 32, the content information It is a point provided with the memory | storage part 33 and the content search part 34. FIG.

  The related word extraction unit 32 is different from the related word extraction unit 12 illustrated in FIG. 7 in that a storage unit 321 is provided instead of the storage unit 121. The storage unit 321 includes a time series text storage unit 322 and a time series character string table storage unit 123. The time series text storage unit 322 stores the time series text associated with the content ID. The time-series text is content text information including or adding time information.

  The content information storage unit 33 stores content introduction information associated with the content ID. When the content text information and the content introduction information are the same information, the related word extraction device 30 may not provide the content information storage unit 33.

  The content search unit 34 is the same as the content search unit 22 of the second embodiment except that the search target based on keywords and the search target based on related terms are time-series text information stored in the time-series text storage unit 322. Perform the process.

  The related word extraction device 30 performs the same process as the process shown in FIG. However, in step S310, the content search unit 34 of the related word extraction device 30 searches the time series text stored in the time series text storage unit 322 using the keyword acquired from the keyword related word table as a search key. The content search unit 34 reads content introduction information associated with the same content ID as the time-series text obtained as a result of the search from the content information storage unit 33. In step S315, the content search unit 34 searches the time series text stored in the time series text storage unit 322 using the related words acquired from the keyword related word table as a search key. The content search unit 34 reads content introduction information associated with the same content ID as the time-series text obtained as a result of the search from the content information storage unit 33. In step S315, the output unit 13 presents the content introduction information read by the content search unit 34.

  In this embodiment, it is possible to acquire a keyword-related word using content that can be introduced to the user and information related to the content, and use it for content recommendation.

[Fourth Embodiment]
In the present embodiment, recently used words are used as keywords, and related words are extracted and contents are recommended. Hereinafter, differences from the second embodiment will be mainly described.

  FIG. 10 is a functional block diagram showing the configuration of the related word extracting device 40 according to the fourth embodiment, and only functional blocks related to the present embodiment are extracted and shown. In this figure, the same parts as those in the related word extracting apparatus 20 according to the second embodiment shown in FIG. The related word extraction device 40 is different from the related word extraction device 20 shown in FIG. 7 in that it further includes a time-series text storage unit 41 and a keyword extraction unit 42. The related word extraction device 40 may be configured to further include a time-series text storage unit 41 and a keyword extraction unit 42 in the related word extraction device 30 according to the third embodiment shown in FIG.

  The time-series text storage unit 41 is time-series text such as web text, program subtitles, news articles, and posted texts of social network services that are published near the time when the user searches. The time series text storage unit 41 stores, for example, time series text of a relatively short period such as one day, one week, and one month from the current time. The period is stored in advance in the storage unit 121 and can be variable.

  When the search start is input by the input unit 11, the keyword extraction unit 42 converts each time-series text stored in the time-series text storage unit 41 into a character string by the same process as step S 215 in the first embodiment. Carve out. The keyword extraction unit 42 extracts a character string as a keyword from the time series text by an existing arbitrary burst detection process. Alternatively, the keyword extraction unit 42 may extract a character string whose frequency is higher than a predetermined value than other character strings, and extracts a character string whose frequency is higher by a predetermined ratio or a predetermined number than a past average as a keyword. May be. The related word extraction device 40 performs the processing after step S110 in FIG. 8 using the extracted keyword.

  Thereby, the related word extraction apparatus 40 extracts the keyword according to the present time, and the related word of the keyword. The related word extraction device 40 can present to the user the content retrieved by each of the extracted keyword and the related word. For example, during the summer season, keywords such as “pool” and “fireworks display” are selected from a short time ago to the present, and have the same frequency distribution as these keywords (that is, the frequency in summer It is possible to recommend content related to (higher) words to the user. Alternatively, when “a rise in stock prices” is currently a topic, it is possible to recommend content related to “overseas travel” that is predicted to be a topic in the future based on past results to the user.

  The user may input time information using the input unit 11. In this case, the time-series text storage unit 41 stores time-series text for a searchable period. The keyword extraction unit 42 reads the time series text of a relatively short period such as one day, one week, one month or the like from the input time information from the time series text storage unit 41 and extracts the keywords. In this case, the time series text storage unit 41 may not be provided, and the keyword extraction unit 42 may extract keywords using the time series text read from the time series text storage unit 122 based on the input time.

[Fifth Embodiment]
In the present embodiment, words that have recently become a hot topic on the Internet are used as keywords, and related words are extracted and contents are recommended.

  FIG. 11 is a functional block diagram showing the configuration of the related word extracting device 50 according to the fifth embodiment, and only the functional blocks related to the present embodiment are extracted and shown. In this figure, the same parts as those in the related word extracting device 30 according to the third embodiment shown in FIG. The related word extraction device 50 includes an input unit 11, a keyword extraction unit 51, a related word extraction unit 52, a content information storage unit 33, a content search unit 54, and an output unit 13. The related word extraction unit 52 includes storage units 521-1 and 521-2, time-series frequency extraction units 524-1 and 524-2, and a correlation value calculation unit 525. The storage unit 521-1 includes a time series text storage unit 522-1 and a time series character string table storage unit 523-1, and the storage unit 521-2 includes a time series text storage unit 522-2 and a time series character string. It has a table storage unit 523-2.

  The keyword extraction unit 51 extracts keywords using the time series text stored in the time series text storage unit 522-1. The time-series text storage unit 522-1 stores an SNS web page such as a blog published as a time-series text on the Internet. The time-series character string table storage unit 523-1 stores the time-series character string table created by the time-series frequency extraction unit 524-1. The time series frequency extraction unit 524-1 creates a time series character string table of keywords extracted using the time series text stored in the time series text storage unit 522-1.

  The time series text storage unit 522-2 stores the time series text associated with the content ID. The time-series text is content text information including or adding time information. For example, the time-series text storage unit 522-2 stores time-series text relating to programs such as subtitles. The time-series character string table storage unit 523-2 stores the time-series character string table created by the time-series frequency extraction unit 524-2. The time series frequency extraction unit 524-2 uses the time series text stored in the time series text storage unit 522-2 and performs the same processing as the time series frequency extraction unit 124 of the first embodiment shown in FIG. Create a time series string table.

  The correlation value calculation unit 525 is set in the time series frequency of each keyword set in the time series character string table created by the time series frequency extraction unit 524-1 and the time series frequency extraction unit 524-2. The Granger causality test is performed using the time series frequency of each character string. Correlation value calculation section 525 extracts the related words of each keyword based on the test result, and outputs them to content search section 54.

The content information storage unit 33 stores content introduction information of programs in which time-series text is stored in the time-series text storage unit 522-2.
The content search unit 54 reads the content ID associated with the time-series text from which the related word is obtained from the time-series text storage unit 522-2, and stores the content introduction information of the program associated with the read content ID as the content. Read from the information storage unit 33. The content search unit 54 associates the read content introduction information with related words included in the time-series text associated with the same content ID as the content introduction information, and outputs the related information to the output unit 13. This related term represents the theme of the content.

<Operation of Related Word Extraction Device 50>
FIG. 12 is a flowchart showing the operation of the related word extraction device 50.
In step S505, when the search start is input by the input unit 11, the keyword extraction unit 51 includes the time series text stored in the time series text storage unit 522-1 for one day, one week, one month from the present time. Identify time-series text for relatively long periods. The keyword extraction unit 51 uses the specified time series text to perform the same processing as the keyword extraction unit 42 of the third embodiment, and the characters that become the keywords Wa1, Wa2,..., Wam (m is an integer of 1 or more) A column is extracted and output to the time-series frequency extraction unit 524-1.

  In step S510, the time series frequency extraction unit 524-1 calculates the time series frequency of each keyword Wa1, Wa2,..., Wam using the time series text stored in the time series text storage unit 522-1. A time series character string table T1 is created.

  Subsequently, in step S515, the time-series frequency extraction unit 524-2 uses the time-series text stored in the time-series text storage unit 522-2 to create a time-series character string table similar to step S115 in FIG. By processing, a time-series character string table T2 in which the time-series frequency of the character strings Wb1, Wb2,..., Wbn (n is an integer of 1 or more) is set is created. Character strings Wb1, Wb2,..., Wbn are keyword related word candidates.

  In step S520, the correlation value calculation unit 525 sets the time series frequency of the keywords Wa1, Wa2,..., Wam set in the time series character string table T1 and the character strings set in the time series character string table T2. All combinations of the keyword Wai (i = 1, 2,..., M) and the character string Wbj (j = 1, 2,..., N) using the time series frequency of Wb1, Wb2,. Granger causality test is performed.

In step S525, the correlation value calculation unit 525 extracts a related word of each keyword Wai from the character strings Wb1, Wb2,..., Wbn based on the test result. The extracted related term _{Wi1, Wi2, ..., Wik i} (k i is an integer from 1 to n) it is. Correlation value calculating section 525, a keyword Wai (i = 1,2, ..., m) and its associated word Wi1, Wi2, ..., and generates a related word table that associates Wik _i.
In step S530, the correlation value calculation unit 525 outputs the generated related word table to the content search unit 54.

In step S535, the content search unit 54 uses the related words Wik (i = 1, 2,..., M, k = 1, 2,..., K _i ) as search keys, and the time-series text storage unit 522. Search for time-series text stored in -2. The content search unit 54 reads content introduction information associated with the same content ID as the time-series text obtained as a result of the search from the content information storage unit 33. The content search unit 54 uses the related word Wik used as a search key when content introduction information is obtained as a content theme. In other words, the content introduced to the user by the content introduction information is past content that uses the related word Wik as a theme. The content search unit 54 outputs the content introduction information and the content theme to the output unit 13.

  In step S540, the output unit 13 presents content introduction information and a content theme. The output unit 13 may present the content introduction information corresponding to more related words Wik with a higher priority.

  Note that the content search unit 54 searches the content introduction information using the related word Wik that has a causal relationship in the direction from the keyword Wai to the related word Wik, so that the content search unit 54 is expected to occur next to the current topical event. Introductory information on content with the theme of events to be acquired can be acquired. In addition, the content search unit 54 searches the content introduction information using the related word Wik that has a causal relationship in the direction from the related word Wik to the keyword Wai, and thus considers that it is the cause of the current topic. Introductory information on content with the theme of events can be acquired.

  According to this embodiment, the related word extraction device 50 can present related words of character strings that frequently appear in recent SNSs as a theme. At this time, the related word extracting device 50 may simultaneously present information on past programs dealing with the theme. By calculating the time series correlation between the past SNS and the program, the theme of the program to be picked up next can be presented from the topic of the recent SNS.

  According to the embodiment described above, the related word extraction device can extract related words without using co-occurrence or a concept dictionary.

  Note that the related word extraction devices 10, 20, 30, 40, 50 described above have a computer system therein. The operation process of the related word extraction device 10, 20, 30, 40, 50 is stored in a computer-readable recording medium in the form of a program, and this program is read and executed by the computer system. The above processing is performed. The computer system here includes a CPU, various memories, an OS, and hardware such as peripheral devices.

Further, the “computer system” includes a homepage providing environment (or display environment) if a WWW system is used.
The “computer-readable recording medium” refers to a storage device such as a flexible medium, a magneto-optical disk, a portable medium such as a ROM and a CD-ROM, and a hard disk incorporated in a computer system. Furthermore, the “computer-readable recording medium” dynamically holds a program for a short time like a communication line when transmitting a program via a network such as the Internet or a communication line such as a telephone line. In this case, a volatile memory in a computer system serving as a server or a client in that case, and a program that holds a program for a certain period of time are also included. The program may be a program for realizing a part of the functions described above, and may be a program capable of realizing the functions described above in combination with a program already recorded in a computer system.

10, 20, 30, 40, 50 ... related word extraction device 11 ... input unit 12, 32, 52 ... related word extraction unit 13 ... output unit 21, 33 ... content information storage unit 22, 34, 54 ... content search unit 41 ... time-series text storage units 42, 51 ... keyword extraction units 121, 321, 521-1, 521-2 ... storage units 122, 322, 522-1, 522-2 ... time-series text storage units 123, 523-1, 523-2 ... Time-series character string table storage unit 124, 524-1, 524-2 ... Time-series frequency extraction unit 125, 525 ... Correlation value calculation unit

Claims (4)

A related word extraction device that extracts related words of each character string included in text data,
An input unit for inputting a predetermined keyword;
A storage unit for storing time-series text that is the text data associated with time information;
A time series frequency extraction unit that extracts a time series frequency that is an appearance frequency when each of the character strings included in the time series text appears in a time series;
Using the time series frequency extracted by the time series frequency extraction unit, a correlation value that is a value that quantitatively represents the strength of the causal relationship between the different character strings is calculated, and the calculated correlation value A correlation value calculating unit that determines whether or not the different character strings are related based on each other;
An output unit that outputs, as a related word, another character string that the correlation value calculation unit determines to be related to the character string of the keyword;
A related word extraction device comprising: The related word extraction device according to claim 1,
A content information storage unit that stores content text information that is text data indicating information about the content;
A content search unit that searches the content text information using the related terms;
The output unit is
Outputting the information of the content specified by the content text information obtained as a result of the search by the content search unit;
The related word extraction device characterized by this. The related word extraction device according to claim 1,
The storage unit
Storing the first time series text including the text data and the time information included in the web page on the Internet, and the second time series text including the text data indicating the information about the content and the time information;
The related word extraction device is
A keyword extraction unit for extracting a keyword from the first time-series text;
The time series frequency extraction unit
Extracting the time series frequency of the keyword included in the first time series text and the time series frequency of each character string included in the second time series text;
The correlation value calculation unit
Using the time-series frequency extracted by the time-series frequency extraction unit for each of the keyword and the character string, a correlation value between the keyword and the character string is calculated, and based on the calculated correlation value Determine whether the string is related to the keyword;
The output unit is
The character string determined to be related to the keyword by the correlation value calculation unit is output as a related word representing the theme of the content, and the content of the content specified by the second time-series text including the output related word Output information,
The related word extraction device characterized by this.   The program for functioning a computer as a related word extraction apparatus as described in any one of Claims 1-3.

Images