JP6136702B2 - Location estimation method, location estimation apparatus, and location estimation program - Google Patents

Description

  The present invention relates to a location estimation method, a location estimation device, and a location estimation program.

  A terminal such as a mobile phone or a smartphone has generally been equipped with a positioning function such as GPS (Global Positioning System). The user of the terminal can measure the current location using the terminal. In addition, when a user of a terminal transmits a document from the terminal using social media, for example, information on the current location can be given to the document.

  In recent years, a dictionary-based location estimation method and a machine learning-based location estimation method have been disclosed as methods for estimating a location mentioned in a document transmitted from a terminal. The dictionary-based location estimation method is a method for estimating a location mentioned in a document by using a dictionary that is prepared in advance and that shows information on a combination of a keyword and a location corresponding to the keyword. . On the other hand, the machine learning-based location estimation method is a method of learning a tendency of keywords appearing in a document and estimating a mentioned location from a keyword (region related word) indirectly corresponding to the location.

JP 2010-128898 A Special table 2012-529717 gazette

  According to the dictionary-based location estimation method, a keyword representing a location is registered in the dictionary, and if the location for the keyword can be identified, the location can be estimated correctly. However, for example, when the same place name is used in a plurality of regions, ambiguity that cannot uniquely identify a place becomes a problem. For example, even if “Minato Ward” is extracted from a document, it is difficult to specify whether it means “Minato Ward in Osaka City”, “Minato Ward in Nagoya City”, or “Minato Ward in Tokyo”.

  Further, according to the machine learning-based location estimation method, an approximate location can be specified, but a detailed location may not be specified. For example, although it is possible to specify up to Osaka Prefecture, it may not be possible to specify the city or town in Osaka Prefecture. As described above, the machine learning-based location estimation method has a problem that the granularity of estimation is larger than that of the dictionary-based location estimation method.

  In one aspect of the present invention, a place estimation method, a place estimation apparatus, and a place estimation program capable of improving the accuracy of estimating a place referred to in a document transmitted from a user terminal are provided. With the goal.

According to one aspect of the invention, there is provided a location estimation method executed by a location estimation device , wherein a first word is extracted from user transmission information, and information related to an address corresponding to the first word is stored from a storage unit. A first index value corresponding to the first keyword including the first word and the address information is extracted and calculated for each location, and the transmission information indicates based on the first index value A location estimation method for estimating a location is provided.

  According to one embodiment, it is possible to provide a place estimation method, a place estimation apparatus, and a place estimation program capable of improving the accuracy of estimating a place referred to in a document transmitted from a user terminal. it can.

FIG. 1 is a diagram illustrating an example of a configuration of a location estimation system. FIG. 2 is a diagram illustrating an example of a hardware configuration of the location estimation apparatus. FIG. 3 is a flowchart illustrating an example of a location estimation method by the location estimation system. FIG. 4 is a flowchart illustrating an example of a learning phase process in S101. FIG. 5 is a diagram illustrating an example of document information to which regional information is added. FIG. 6 is a diagram illustrating an example of a list of words extracted from a document with regional information. FIG. 7 is a flowchart illustrating an example of processing for extracting information about an address from the address expression dictionary in S203. FIG. 8 is a diagram illustrating an example of a list of address expressions in the learning phase. FIG. 9 is a diagram illustrating an example of an address expression dictionary. FIG. 10 is an example of a list of information regarding addresses extracted from the address expression dictionary. FIG. 11 is a diagram illustrating an example of a keyword group in the learning phase. FIG. 12 is a diagram illustrating an example of a region classification rule. FIG. 13 is a flowchart illustrating an example of the process of the estimation phase in S102. FIG. 14 is a diagram illustrating an example of information of a target document. FIG. 15 is a flowchart illustrating an example of processing for extracting a character string constituting an address from the address expression dictionary in S403. FIG. 16 is a diagram illustrating an example of a list of address expressions in the estimation phase. FIG. 17 is an example of a list of address and coordinate information extracted from the address expression dictionary. FIG. 18 is a diagram illustrating an example of a keyword group in the estimation phase. FIG. 19 is an example of the calculation result of the affiliation score. FIG. 20 is an example of a list of regions where the affiliation score exceeds a predetermined threshold. FIG. 21 is a diagram illustrating an example of an estimation result. FIG. 22 is a diagram illustrating another example of estimating the location mentioned in the target document from the target document. FIG. 23 is a flowchart showing a modification of the learning phase. FIG. 24 is a diagram illustrating an example of a word list in the learning phase.

  Hereinafter, embodiments of the present invention will be specifically described with reference to FIGS. 1 to 24.

  FIG. 1 is a diagram illustrating an example of a location estimation system. As shown in FIG. 1, the location estimation system includes a location estimation device 10 and a terminal device 30. The location estimation device 10 and the terminal device 30 are connected via a network 50 so that they can communicate with each other.

  The location estimation device 10 is a device that receives a document transmitted from the terminal device 30 and estimates a location mentioned in the document. As will be described later, the location estimation apparatus 10 performs estimation by combining, for example, a dictionary-based location estimation method and a machine learning-based location estimation method. The location estimation apparatus 10 is realized by, for example, a personal computer (PC) or a server. A method of processing executed by the place estimation device 10 will be described later.

  The terminal device 30 is a terminal owned by, for example, a user who transmits a document or a user who uses the location estimation device 10, and is, for example, a smartphone, a mobile phone, a notebook PC (Personal Computer), a desktop PC, or a tablet terminal. Here, the document is, for example, a news article, a sentence written on social media such as Twitter (registered trademark) or a blog. Note that the user includes a user who transmits a document assigned with a region in order to acquire a region classification rule in a learning phase to be described later.

  Next, the hardware configuration of the location estimation apparatus 10 will be described.

  FIG. 2 is a diagram illustrating an example of a hardware configuration of the location estimation apparatus 10. As shown in FIG. 2, the location estimation apparatus 10 includes a CPU (Central Processing Unit) 61, a ROM (Read Only Memory) 62, a RAM (Random Access Memory) 63, a storage device 64, a network interface 65, and a portable storage medium. Drive 66 and the like.

  Each component of the location estimation device 10 is connected to a bus 67. The storage device 64 is, for example, an HDD (Hard Disk Drive). In the location estimation device 10, a program (including a location estimation program) stored in the ROM 62 or the storage device 64, or a program (including a location estimation program) read from the portable storage medium 68 by the portable storage medium drive 66. Is executed by a processor such as the CPU 61, so that the function of the location estimation apparatus 10 is realized.

  Hereinafter, the function of each part which comprises the place estimation apparatus 10 is demonstrated.

  As illustrated in FIG. 1, the place estimation device 10 includes a first storage unit 11, a second storage unit 12, a reception unit 13, a keyword extraction unit 14, a dictionary search unit 15, a score calculation unit 16, An estimation unit 17, an output unit 18, and a transmission unit 19 are provided.

  The first storage unit 11 corresponds to, for example, the ROM 62, the storage device 64, the portable storage medium drive 66, or the portable storage medium 68 shown in FIG. The first storage unit 11 can store a place estimation program for estimating a place mentioned in a document transmitted from the terminal device 30.

  The second storage unit 12 corresponds to, for example, the ROM 62, the RAM 63, the storage device 64, the portable storage medium drive 66 or the portable storage medium 68 shown in FIG. The 2nd memory | storage part 12 is used as a database (DB; DataBase) for memorize | storing the various information used for each process performed in the place estimation apparatus 10. FIG.

  The receiving unit 13 receives a document assigned with a region from the terminal device 30. The receiving unit 13 is connected to the terminal device 30 so as to be able to communicate with each other, and is realized by the network interface 65 of FIG.

  The keyword extraction unit 14 can extract words from the document received by the reception unit 13. Moreover, the keyword extraction part 14 can extract an address expression from the extracted word. Here, the address expression is a name indicating a building, a park, a company, a school, or a store having an address, or the address itself.

  The dictionary search unit 15 searches the address expression dictionary stored in the second storage unit 12 and extracts information related to the address corresponding to each word extracted by the keyword extraction unit 14. The position information extraction unit 15 is realized by a processor such as the CPU 61 or MPU in FIG.

  The score calculation unit 16 is a region-related score indicating the degree of relevance with each region for each of the word extracted by the keyword extraction unit 14 and the address expression extracted by the dictionary search unit 15 in the learning phase described later. Is calculated for each region. The calculated region-related score is used as data for the region classification rule. The region-related score is an example of a second index value.

  Further, the score calculation unit 16 is an affiliation score that is an index indicating which region the keyword group extracted from the target document by the keyword extraction unit 14 is likely to belong to (relate to) in the estimation phase described later. Is calculated for each region. Then, by comparing the affiliation scores calculated for each region, it is possible to estimate to which region the place mentioned in the document is likely to belong. Here, one of the features of the present embodiment is that not only the word but also the character string constituting the address is a target for calculating the affiliation score for each region. The score calculation unit 16 is realized by a processor such as the CPU 61 or MPU in FIG. The affiliation score is an example of a first index value.

  The estimation unit 17 estimates the place (region) mentioned in the input document by comparing the affiliation score calculated by the score calculation unit 16 with a predetermined threshold. The estimation unit 17 is realized by a processor such as the CPU 61 or MPU in FIG.

  The output unit 18 outputs the estimation result of the place mentioned in the document, which is executed by the estimation unit 17. The output unit 18 is a display device such as a liquid crystal display, a plasma display, or an organic EL display.

  The transmission unit 19 has a function of transmitting the estimation result of the place mentioned in the document, which is executed by the estimation unit 17, to the network 50. For example, the terminal device 30 of the user who uses the location estimation device 10 can receive the estimation result transmitted from the location estimation device 10 via the network 50.

  Next, the place estimation method by the place estimation apparatus 10 will be described.

  FIG. 3 is a flowchart illustrating an example of a location estimation method by the location estimation system.

  As shown in FIG. 3, the place estimation device 10 first performs learning based on a document to which area information (area information), which is an example of place information, is given, and acquires a region classification rule. Execute (S101). Subsequently, the location estimation apparatus 10 acquires the target document, and executes an estimation phase in which the location mentioned in the target document is estimated based on the regional classification rule (S102). Hereinafter, details of the learning phase and the estimation phase will be described.

  First, the details of the learning phase processing in S101 will be described.

  FIG. 4 is a flowchart illustrating an example of a learning phase process in S101.

  First, the receiving unit 13 receives information on a document to which regional information is assigned from the terminal device 30 (S201). The receiving unit 13 stores the received document information in the second storage unit 12.

  FIG. 5 is a diagram illustrating an example of document information to which regional information is added. As shown in FIG. 5, the document information to which the region information is added has a configuration in which the region and the document are associated with each other. The region is, for example, a prefecture name. In the example of FIG. 5, the region “○ prefecture” is assigned as the region information to the document “Going to the tower and eating the famous eel on the way home”. Further, the region “△ prefecture” is given as the region information to the document “Speaking of Δ, it will be Δ temple”.

  Returning to FIG. 4, the keyword extraction unit 14 extracts words from the document to which the regional information is assigned (S202). In S <b> 202, the keyword extraction unit 14 reads the document to which the regional information stored in the second storage unit 12 is added. Then, the keyword extracting unit 14 extracts words included in the document to which the regional information is assigned while being associated with the regional information. As a word extracted here, a noun and a compound noun are preferable.

  FIG. 6 is a diagram illustrating an example of a list of words extracted from a document with regional information. In the example of FIG. 6, “○ Tower”, “Specialties”, and “Eel” are extracted as words corresponding to the region “○ Prefecture”. In addition, “Δ” and “Δ temple” are extracted as words corresponding to the region “Δ prefecture”.

  Returning to FIG. 4, the dictionary search unit 15 searches the address expression dictionary stored in the second storage unit 12, and extracts a character string constituting an address corresponding to the address expression (S <b> 203). Details of the process of S203 will be described below.

  FIG. 7 is a flowchart illustrating an example of processing for extracting information about an address from the address expression dictionary in S203.

  First, the keyword extraction unit 14 extracts an address expression from a document to which regional information is assigned, and creates a list of address expressions (S301).

  FIG. 8 is a diagram illustrating an example of a list of address expressions in the learning phase. As shown in FIG. 8, the list of address expressions includes “○ Tower” extracted from the document “○ went to Tower and ate the famous eel on the way back”, and document “△” “△ Temple” extracted from “It will be a temple” is registered. According to the processing of S301, the address expression is extracted using a specific expression extraction technique or the like. Any known technique can be used for the specific expression extraction technique.

  Returning to FIG. 7, after the process of S301, the dictionary search unit 15 determines whether or not the address expression is registered in the list of address expressions (S302). When it is determined that information is registered in the list of address expressions (Yes in S302), the dictionary search unit 15 extracts one address expression from the list of address expressions. Then, the keyword extraction unit 14 deletes the extracted address expression from the list of address expressions (S303).

  For example, as shown in FIG. 8, it is assumed that the dictionary search unit 15 extracts “○ tower” in a state where “○ tower” and “Δ temple” are registered in the list of address expressions. At this time, the dictionary search unit 15 deletes the extracted “○ tower” from the list of address expressions. As a result, only “△ temple” remains in the list of address expressions.

  Subsequently, the dictionary search unit 15 searches the address expression dictionary and determines whether or not the address expression extracted in S303 is registered in the address expression dictionary (S304).

  FIG. 9 is a diagram illustrating an example of an address expression dictionary. The address expression dictionary is a list in which address expressions are associated with information on places corresponding to the address expressions. The address expression dictionary has columns of “address expression”, “address”, and “coordinates (latitude, longitude)” as shown in FIG. 9, for example. Information registered in the column of “coordinates (latitude, longitude)” is information used in the estimation phase. The dictionary search unit 15 performs the determination of S304 in FIG. 7 by searching whether or not the extracted address expression is registered in the “address expression” column.

  When searching for an address expression, the character string of the address expression extracted by the dictionary search unit 15 and the character string of the address expression in the address expression dictionary do not need to be completely identical. For example, any known method such as a search method using approximate matching or a method of searching for an exact match after normalizing the extracted character string can be used.

  When it is determined that the address expression extracted in S303 is included in the address expression dictionary (Yes in S304), the dictionary search unit 15 searches the address expression dictionary for a character string that constitutes an address corresponding to the address expression. Extract. Then, the dictionary search unit 15 stores the extracted result in the second storage unit 12 (S305).

  For example, referring to the address expression dictionary of FIG. 9, it can be seen that “O tower” is registered in the “address expression” column. Therefore, the dictionary search unit 15 refers to the address column corresponding to “○ Tower”, and uses “○ prefecture”, “○ city”, and “○ prefecture” as character strings constituting the address “○ prefecture, ○ city, ○ town”. Extract “○ town”. After the process of S305, the process proceeds to S302, and the processes after S302 are executed again.

  On the other hand, when it is determined that the extracted address expression is not registered in the address expression dictionary (No in S304), the process proceeds to S302, and the processes after S302 are executed again. For example, if “Δ temple” is registered in the address expression list in the second processing of S302, it is determined again as S302 affirmative. Then, the process proceeds to S304 through the process of S303. In the process of S304, the dictionary search unit 15 refers to the address field corresponding to “△ temple” and sets “△ prefecture”, “△ city” as character strings constituting the address “△ prefecture, Δ city, Δ town”. "And" △ town "are extracted.

  In S302, when it is determined that the address expression is not registered in the list of address expressions (No in S302), the process proceeds to S204. In S302, for example, if “○ tower” and “Δ temple” are deleted from the list of address expressions, the address expressions registered in the list of address expressions disappear, so in this case, a negative determination is made in S302. And a series of processes which extract the character string which comprises an address are complete | finished.

  FIG. 10 is an example of a list of character strings constituting the address extracted from the address expression dictionary.

  As shown in FIG. 10, the character string constituting the address is registered in association with the area. Thus, by registering in association with the area, the work of merging the word and the character string constituting the address in the subsequent processing becomes easy.

  As described above, the place estimation apparatus 10 can acquire a character string that constitutes an address from the address expression dictionary.

  Returning to FIG. 4, after the process of S203, the score calculation unit 16 constructs a keyword group by merging the word extracted in S202 and the character string constituting the address extracted in S203 (S204). ).

  FIG. 11 is a diagram illustrating an example of a keyword group in the learning phase. When the list of words shown in FIG. 6 and the list of character strings constituting the address shown in FIG. 10 are merged, the keyword group shown in FIG. 11 can be obtained. As shown in FIG. 11, in the keyword group list, regional information and information related to a word or address corresponding to the region are registered in association with each other. For example, in the region “○ prefecture”, the corresponding words “○ tower”, “specialties” and “eel” and the corresponding addresses “○ prefecture”, “○ city” and “○ town” It is registered. In addition, in the region “△ prefecture”, the corresponding words “△” and “△ temple” and the character strings constituting the corresponding address “△ prefecture”, “△ city”, and “△ town” Is registered.

  Subsequently, the score calculation unit 16 calculates a region-related score for each region for each of the extracted word and the character string constituting the address (S205). As a method for calculating the region-related score and the affiliation score described later, various known score calculation methods in machine learning such as a method using a linear classifier learned by a support vector machine (Support Vector Machine, SVM) are used. Can do. By calculating the region-related score, the region classification rule can be acquired.

  FIG. 12 is a diagram illustrating an example of a region classification rule. As shown in FIG. 12, for each word, a region-related score for each region is registered. In the example of FIG. 12, the region is represented by a prefecture name. It shows that the relevance with the said area is so high that the value of an area related score is large. According to FIG. 12, for example, in the case of the word “○ tower”, the score of “○ prefecture” is 0.75, the score of “Δ prefecture” is 0.03, and the score of “× prefecture” is 0.04. Therefore, it can be seen that the three regions have the highest relevance to the prefecture. The acquired area classification rule is stored in the second storage unit 12 and used in the subsequent estimation phase.

  As described above, the place estimation device 10 executes the learning phase for acquiring the area classification rule.

  Next, the details of the estimation phase processing in S102 will be described.

  FIG. 13 is a flowchart illustrating an example of the process of the estimation phase in S102.

  First, the receiving unit 13 receives a target document whose location is to be estimated from the terminal device 30 (S401). The receiving unit 13 stores the received information on the target document in the second storage unit 12.

  FIG. 14 is a diagram illustrating an example of information of a target document. In the example of FIG. 14, “Take a vacation and have seen a temple or a shrine” is the target document. Unlike the learning document shown in FIG. 5, the information of the target document does not include regional information. For this reason, the correspondence between the target document and the region is unknown. The location estimation apparatus 10 performs each process after S401 in order to estimate the location mentioned in the target document.

  Returning to FIG. 13, after the processing of S401, the keyword extraction unit 14 extracts words from the target document (S402). In S <b> 402, the keyword extraction unit 14 reads the target document stored in the second storage unit 12. Then, the keyword extraction unit 14 extracts words included in the target document. Here, “vacation”, “Δ temple”, and “Δ shrine” are extracted as words from the target document “I took a vacation and have seen Δ temples and shrines”.

  Subsequently, the keyword extraction unit 14 extracts an address expression from the target document. And the dictionary search part 15 searches the address expression dictionary stored in the 2nd memory | storage part 12, and extracts the character string which comprises the address corresponding to an address expression (S403).

  FIG. 15 is a flowchart illustrating an example of processing for extracting a character string constituting an address from the address expression dictionary in S403.

  First, the keyword extraction unit 14 extracts an address expression from the target document and creates a list of address expressions (S501).

  FIG. 16 is a diagram illustrating an example of a list of address expressions in the estimation phase. As shown in FIG. 16, the keyword extraction unit 14 extracts “△ temple” and “△ shrine” as address expressions from the extracted words “vacation”, “△ temple”, and “△ shrine”. To do.

  Subsequently, the dictionary search unit 15 determines whether or not information is registered in the address expression list (S502). When it is determined that information is registered in the list of address expressions (Yes in S502), the dictionary search unit 15 extracts one address expression from the list of address expressions. Then, the dictionary search unit 15 deletes the extracted address expression from the list of address expressions (S503). On the other hand, when it is determined that the information is not registered in the address expression list (No in S502), the process proceeds to S404.

  For example, as shown in FIG. 16, it is assumed that the dictionary search unit 15 extracts “Δ temple” from the list of address expressions when “Δ temple” and “Δ shrine” are registered in the list of address expressions. At this time, the dictionary search unit 15 extracts “Δ temple” and then deletes “Δ temple” from the list of address expressions. As a result, only “△ shrine” remains in the list of address expressions.

  Subsequently, the dictionary search unit 15 determines whether the address expression extracted in S503 is registered in the address expression dictionary (S504). When it is determined that the address expression extracted in S503 is registered in the address expression dictionary (Yes in S504), the dictionary search unit 15 uses the address and the location information corresponding to the extracted address expression from the address expression dictionary. The coordinate information is extracted (S505).

  FIG. 17 is an example of a list of address and coordinate information extracted from the address expression dictionary. As shown in FIG. 17, corresponding address and coordinate information is extracted for each address expression.

  For example, if it is determined in S504 that the address expression “Δ temple” is registered in the address expression dictionary, the dictionary search unit 15 in S505 uses “Δ prefecture, Δ city, “Δ Town 34.xxxx, 138.xxxx” ”is extracted. After the process of S505, the process returns to S502, and the processes after S502 are executed again.

  In the process of S502 executed again, “Δ shrine” is registered in the address expression list. In this case, it is determined again as S502 affirmative. Then, the process proceeds to S504 through the process of S503. In the process of S504, the dictionary search unit 15 uses two types of area information, that is, “△ prefecture, △ city, ▲ town 34.xxxx, 137.xxxx”, as address and coordinate information corresponding to “△ shrine”. Extract “□ prefecture, □ city, □ town 37.xxxx, 131.xxxx”. Thus, the reason why the information of two types of regions is extracted is that “Δ shrine” exists in a plurality of regions, and information on the two regions is registered in the address expression dictionary. After the process of S505, the process returns to S502.

  In the process of S502 executed for the third time, since all address expressions have been extracted, there is no address expression registered in the list of address expressions. Therefore, the dictionary search unit 15 determines that the information is not registered in the address expression list (No in S502), and proceeds to S404.

  Returning to FIG. 13, the keyword extracting unit 14 and the score calculating unit 16 extract a keyword group by merging the word extracted in S402 and the character string constituting the address extracted in S403 (S404). ). In S404, the keyword extraction unit 14 extracts a character string that constitutes an address from the list of address and coordinate information extracted from the address expression dictionary in S505. For example, as a result, “Δ prefecture”, “Δ city”, “▲ town”, “□ prefecture”, “□ city”, and “□ town” are extracted. When these character strings are merged with the words “vacation”, “Δ temple”, and “Δ shrine” extracted in S402, a keyword group can be obtained.

  FIG. 18 is a diagram illustrating an example of a keyword group in the estimation phase. As described above, a keyword group used for location estimation is extracted.

  Returning to FIG. 13, after the process of S404, the score calculation unit 16 calculates the affiliation score for each region in order to estimate which region the keyword group belongs to (relevant) is likely (S405). . The score calculation unit 16 calculates the affiliation score for each region based on the keyword group and the region classification rule acquired in S101.

  FIG. 19 is an example of the calculation result of the affiliation score. In the example of FIG. 19 as well, the region is represented by the prefecture name as in FIG. Moreover, it shows that the relevance with the said area is so strong that the value of an affiliation score is large.

  Returning to FIG. 13, after the process of S405, the estimation unit 17 creates a list of regions where the affiliation score exceeds a predetermined threshold (S406). Specifically, the estimation unit 17 refers to the affiliation score for each region calculated in S405, and extracts a region that exceeds a predetermined threshold. For example, referring to FIG. 17 with a threshold value of 0, only “Δ prefecture” having an affiliation score of 1.88 is extracted. Therefore, the estimation unit 17 creates a list of regions on which “Δ prefecture” is placed. Note that it is preferable to set the threshold value so that the estimation accuracy becomes the highest while repeating an experiment for measuring the estimation accuracy using a document in which a place already mentioned is known. By this method, it is possible to improve the estimation accuracy.

  FIG. 20 is an example of a list of regions where the affiliation score exceeds a predetermined threshold. In the example of FIG. 20, only “Δ prefecture” is registered as a region in the region list. However, when there are a plurality of regions exceeding the threshold, a plurality of regions are registered.

  Returning to FIG. 13, after the processing of S405, the estimation unit 17 determines whether or not regional information is registered in the regional list (S407).

  If it is determined that the area information is not registered in the area list (No at S407), the process of the estimation phase ends.

  On the other hand, when it is determined that the area information is registered in the area list (Yes in S407), the estimation unit 17 extracts one area from the area list. And the estimation part 17 deletes the said area | region from the area | region list, after extracting an area | region (S408).

  For example, the estimation unit 17 extracts “Δ prefecture” in a state where only “Δ prefecture”, which is area information, is registered in the area list. Thereafter, the dictionary search unit 15 deletes “Δ prefecture” from the list of regions. As a result, the area information is not registered in the area list.

  Then, the estimation part 17 extracts the information regarding the address relevant to the extracted area. Then, the output unit 18 outputs the extraction result (S409). For example, from the list of address and coordinate information shown in FIG. 17, the estimating unit 17 obtains “Δ temple Δ prefecture, Δ city, Δ town 34.xxxx, 138.xxxx” as a series of information including “Δ prefecture”. And “△ shrine △ prefecture, △ city, ▲ town 34.xxxx, 137.xxxx” are extracted. Then, the output unit 18 outputs the extracted information as an estimation result.

  FIG. 21 is a diagram illustrating an example of an estimation result. As shown in FIG. 21, the location estimation result can be output for each address expression included in the target document.

  Returning to FIG. 13, after the process of S409, the process returns to S407. And the process after S407 is performed again. In this embodiment, since only “Δ prefecture” is registered in the area list, it is determined that S407 is negative in the second processing of S407 executed after “Δ prefecture” is deleted. As a result, the estimation phase ends.

  As described above, the location estimation apparatus 10 can estimate the location mentioned in the received document.

  Hereinafter, the effect by this embodiment is demonstrated.

  FIG. 22 is a diagram illustrating another example of estimating the location mentioned in the target document from the target document. FIG. 22A shows the target document. FIG. 22B shows a list of address and coordinate information extracted from the address expression dictionary corresponding to the address expression in the target document. FIG. 22C shows the calculation result of the affiliation score for each area calculated based on the address expression in the target document.

  As shown in FIG. 22 (a), in the estimation phase, when the target document is “Used Musashikosugi Station yesterday”, “Musashikosugi Station” is extracted as an address expression. Then, referring to FIG. 22B, since “Musashi Kosugi Station” is registered in the address expression dictionary, information on the address corresponding to “Musashi Kosugi Station” from the address expression dictionary “Kanagawa Prefecture, Kawasaki City” , Nakahara-ku ”and coordinate information“ 35.5766666667, 139.65944444 ”can be extracted.

  However, in the learning phase, the keyword group used when creating the regional classification rule does not necessarily cover all address expressions in the dictionary. For example, in the learning phase, when “Musashi Kosugi Station” is not included in the keyword group used when calculating the region-related score, it is not learned that “Musashi Kosugi Station” is in Kanagawa Prefecture. For this reason, as shown in the example of FIG. 22C, when the affiliation score for Kanagawa Prefecture is calculated, the affiliation score may be lower than a threshold value (set to 0.3). As a result, in spite of the correct estimation by the address expression dictionary in the estimation by the dictionary, the estimation result by the dictionary is rejected by the estimation by the machine learning, and it is determined that the target document does not mention the place. possible.

  On the other hand, according to the present embodiment, in the learning phase and the estimation phase, the keyword group is constructed by adding not only the word extracted from the target document but also the character string constituting the address extracted from the address expression dictionary. . As a result, since the region-related score and the affiliation score can be calculated more accurately, as described above, the estimation result is based on the estimation by the machine learning despite the correct estimation by the address expression dictionary. It can be prevented from being rejected.

  The process of constructing a keyword group by adding not only words but also character strings constituting an address can be expected to have a corresponding effect even when it is performed only in the estimation phase. However, it is preferable to perform the above-described processing in both the learning phase and the estimation phase.

(Modification)
Next, a modified example in the present embodiment will be described. In addition, since the information processing system for implement | achieving this modification can use the structure of the place estimation system illustrated by FIG. 1 and FIG. 2, description about an overlapping part is abbreviate | omitted.

  In the embodiment shown in FIG. 23, in the learning phase, when extracting a character string that constitutes an address from the address expression dictionary, the address expression is extracted from the document, and then information related to the address corresponding to each address expression is extracted. doing. On the other hand, this modification is characterized in that the character string constituting the address corresponding to the word is extracted by searching the address expression dictionary without extracting the address expression.

  FIG. 23 is a flowchart showing a modification of the learning phase. Note that the processing from the start of the learning phase processing to S202 is the same as the processing described so far, and a description thereof will be omitted.

  After the process of S202, the keyword extraction unit 14 creates a word list based on the words extracted in the process of S202 (S301a).

  FIG. 24 is a diagram illustrating an example of a word list in the learning phase. As shown in FIG. 24, in the word list, it can be said that “○ Tower”, “Specialties”, “Eel” and the document “△” extracted from the document “○ went to the tower and ate the famous eel on the way back”. For example, “△” and “△ temple” extracted from “it will be △ temple” are registered.

  Returning to FIG. 23, after the processing of S301a, the dictionary search unit 15 searches the words registered in the word list word by word whether there is information regarding the address corresponding to the address expression dictionary.

  First, the dictionary search unit 15 determines whether or not a word is registered in the word list (S302a). When it is determined that a word is registered in the word list (Yes in S302a), the dictionary search unit 15 extracts one word from the word list. Then, the keyword extraction unit 14 deletes the extracted word from the word list (S303a).

  Subsequently, the dictionary search unit 15 searches the address expression dictionary and determines whether or not the word extracted in S303a is registered in the address expression dictionary (S304a).

  When it is determined that the extracted word is included in the address expression dictionary (Yes in S304a), the dictionary search unit 15 extracts a character string constituting the address corresponding to the extracted word from the address expression dictionary. Then, the dictionary search unit 15 stores the extracted result in the second storage unit 12 (S305a).

  On the other hand, when it is determined that the extracted word is not registered in the address expression dictionary (No in S304a), the process proceeds to S302a, and the processes after S302a are executed. If this process is repeated, the words registered in the word list will eventually disappear. When there is no registered word, it is determined in S302a that the address expression is not registered in the list of address expressions (No in S302a), and the process proceeds to S204. Thereby, a series of processes for extracting the character string constituting the address ends. Since the processing after S204 is the same as the processing described so far, description thereof is omitted.

  According to this modification, the process of extracting the address expression from the word can be omitted, so that the process can be simplified. In addition, it is possible to dispel the concern that omission of address expression extraction from words occurs.

  The preferred embodiments of the present invention have been described in detail above, but the present invention is not limited to specific embodiments, and various modifications and changes can be made. For example, the processing of the location estimation apparatus 10 described so far not only performs estimation by combining a dictionary-based location estimation method and a machine learning-based location estimation method, but also includes only a dictionary-based location estimation method or machine learning. It is also applicable in the case of only the base location estimation method. For example, when performing estimation using only the machine learning-based location estimation method, after extracting information related to the address corresponding to the extracted word, the frequency of the location included in the information related to the address is counted, and the count number is It is also possible to estimate where the document is referring by comparing each time. Here, the count number is another example of the first index value.

  Further, for example, in the above-described modification, the process of searching the address expression dictionary and extracting the character string that forms the address corresponding to the word in the learning phase has been described. On the other hand, this process can also be applied to a process of extracting address and coordinate information corresponding to a word using an address expression dictionary in the estimation phase.

10: location estimation device 11: first storage unit 12: second storage unit 13: reception unit 14: keyword extraction unit 15: dictionary search unit 16: score calculation unit 17: estimation unit 18: output unit 19: transmission unit 30: Terminal device 50: Network 61: CPU
62: ROM
63: RAM
64: Storage device 65: Network interface 66: Portable storage medium drive 67: Bus 68: Portable storage medium

Claims (6)

A location estimation method executed by a location estimation device,
Extracting the first word from the user's outgoing information,
Extracting information about the address corresponding to the first word from the storage unit;
Calculating a first index value corresponding to a first keyword including the first word and information about the address for each location;
Based on the first index value, the location indicated by the transmission information is estimated.
A location estimation method characterized by that. The process of estimating the location is:
A location corresponding to an index value larger than a predetermined threshold value among the first index values is estimated as a location indicated by the transmission information,
When there is no index value larger than the predetermined threshold in the first index value, it is estimated that there is no place indicated by the transmission information.
The location estimation method according to claim 1, further comprising a process. Using the second keyword including the second word extracted from the document information to which the location information is assigned and the information related to the address corresponding to the second word, the second keyword and the location A process of calculating a second index value indicating the strength of relevance to each;
The location estimation method according to claim 1, wherein the first index value is calculated based on the second index value. The process of extracting from the storage unit includes extracting a plurality of candidates related to the location corresponding to the first word,
The process of estimating the location indicated by the transmission information includes narrowing down the plurality of candidates based on the first index value calculated for each location.
The location estimation method according to claim 1 or 2, characterized in that: A keyword extraction unit for extracting the first word from the user's transmission information;
A dictionary search unit for extracting information on an address corresponding to the first word from a storage unit;
A score calculation unit that calculates, for each place, a first index value corresponding to a first keyword including the first word and information about the address;
An estimation unit that estimates a place indicated by the transmission information based on the first index value;
A location estimation apparatus characterized by comprising: In place estimation device,
A process of extracting the first word from the user's transmission information;
A process of extracting information about an address corresponding to the first word from a storage unit;
A process of calculating, for each location, a first index value corresponding to a first keyword including the first word and information about the address;
A process of estimating a location indicated by the transmission information based on the first index value;
Location estimation program for running.

Images