JP7182585B2 - program - Google Patents

Description

The present invention relates to an information retrieval device and an information retrieval system, and more particularly to an information retrieval device and an information retrieval system for retrieving a retrieval object representing a specific point based on a retrieval phrase.

2. Description of the Related Art Conventionally, in a map search, there is known an information search device capable of performing a map search not only for data in which coordinate information is explicitly registered, but also for information contained in the content of atypical data (Patent Document 1).

In the information retrieval device of Patent Document 1, data including multiple words having attributes is input, multiple words included in the input data are extracted, and attribute-related words extracted for each attribute are converted. Search is performed using data with attribute information added.

For example, in the case of the word "Hokkaido" related to the location attribute, the information retrieval device of Patent Document 1 has the coordinate information of "Hokkaido" added to the atypical data. can be used to output predetermined search results.

JP 2014-191714 A

However, in the configuration of the information search device of Patent Document 1, search results cannot be output unless the word has an attribute, and a more accurate search target cannot be detected from the search phrase.

An object of the present invention is to detect search targets with higher accuracy.

An information search device according to a first aspect of the present disclosure includes a position information acquisition unit, a feature word generation unit, and a search unit. The location information acquisition unit acquires location information. The position information indicates the position on the map that is the starting point of the search. The feature word generation unit generates a feature word from the search phrase. The feature words are associated with words representing search targets. The search unit searches for the search target from the storage database. The accumulation database accumulates the feature words and the search targets. The search target corresponds to the feature word. The search target has point information. The location information indicates a specific location. The search unit searches for the search target corresponding to the feature word within a predetermined range from the position based on the position information and the feature word.

The information search device of the first aspect can detect search targets with higher accuracy.

In the information search device according to the second aspect of the present disclosure, the position information acquisition unit acquires the position information. The position information is estimated from words included in the search phrase by machine learning.

The information retrieval device of the second aspect can estimate position information by machine learning. The information search device of the second aspect can indirectly estimate the position information from the words included in the search phrase even if the search phrase does not include a word directly indicating the position information.

The information search device according to the third aspect of the present disclosure further includes a category determination section and a search storage device. The category discriminating unit discriminates the category of the search phrase. Words corresponding to the categories are stored in the search storage device. In the accumulation database, the feature words are classified according to the categories. The category determination unit has a category determination unit and a category estimation unit. The category determination unit determines the category based on whether or not words included in the search phrase match words stored in advance in the search storage device. The category estimating unit uses machine learning to estimate the category from words included in the search phrases that have not been determined by the category determining unit. The search unit searches for the search target corresponding to the feature word within a predetermined range from the position and in the determined category based on the position information, the category, and the feature word.

The information retrieval device of the third aspect can estimate the category by machine learning. The information search device of the third aspect can indirectly estimate the category from the words included in the search phrase even if the search phrase does not include a word that directly indicates the category.

In the information search device according to the fourth aspect of the present disclosure, the feature word generation unit selects words including nouns, adjectives, verbs, and adverbs in the search phrase that have a higher degree of similarity with words representing the search target. Generate feature words.

The information retrieval device of the fourth aspect can detect a retrieval target with higher accuracy by generating a feature word having a higher degree of similarity with the word representing the retrieval target.

An information retrieval system according to a fifth aspect of the present disclosure includes an information processing terminal, an information retrieval device, and an information storage device. The information retrieval device is configured to be communicable with the information processing terminal. The information storage device is configured to be accessible by the information retrieval device. The information processing terminal includes an input device, a terminal communication device, and a display device. The input device is configured to allow entry of a search phrase. The terminal communication device is configured to be able to transmit the search phrase to the information search device. The terminal communication device is configured to receive a search target searched based on the search phrase from the information search device. The display device is configured to display the search target.

The information search device includes a search communication device, a position information acquisition section, a feature word generation section, and a search section. The search communication device is configured to receive the search phrase from the information processing terminal. The search communication device is configured to be able to transmit the search target. The location information acquisition unit is configured to acquire location information. The position information indicates the position on the map that is the starting point of the search. The feature word generation unit generates a feature word from the search phrase. The feature word is associated with the word representing the search target. The search unit is configured to search the storage database for the search target.

The search target has point information. The location information indicates a specific location. The search unit is configured to search for the search target corresponding to the feature word within a predetermined range from the position based on the position information and the feature word. The information storage device includes an accumulation storage device. The accumulation storage device stores the accumulation database. The accumulation database accumulates the feature words and the search targets. The search target corresponds to the feature word.

The information search system of the fifth aspect can detect search targets with higher accuracy.

In the information retrieval system according to the sixth aspect of the present disclosure, the information processing terminal further includes a terminal storage device. Map information is stored in the terminal storage device. The display device displays the specific point indicating the search target on the map of the map information.

The information search system of the sixth aspect can display a more accurate search target point on a map.

In the information search system according to the seventh aspect of the present disclosure, the feature word generation unit calculates statistical data of the search phrases. The information storage device associates the search phrase received by the search communication device from the information processing terminal within a preset period with the statistical data, and stores them in the storage database.

The information search system of the seventh aspect can detect search targets with higher accuracy based on search phrases within a preset period.

In the information search system according to the eighth aspect of the present disclosure, the feature word generation unit generates the feature word having a higher degree of similarity with the word representing the search target from the words included in the search phrase. The feature word generation unit preferentially outputs the feature word with a higher frequency based on the statistical data.

The information search system of the eighth aspect can detect search targets with even higher accuracy.

According to the present invention, search targets can be detected with higher accuracy.

1 is a schematic configuration diagram of an information search system according to this embodiment; FIG. It is a flowchart figure which shows the processing operation of the information search system which concerns on this embodiment. FIG. 4 is a flow chart diagram showing an example of processing for searching for spots based on clusters. It is a figure which shows the specific example which determined the cluster from the search phrase. It is a figure which shows an example of the data table stored in a category database. It is a figure which shows an example of the data table stored in the spot database. FIG. 2 is a block diagram showing an example of the main configuration of an information storage device; FIG. 10 is a flow chart diagram showing an example of processing for generating a spot database; FIG. 10 is a flow chart diagram showing an example of processing for generating a category database;

The information retrieval system 10 of this embodiment will be described below. The information retrieval system 10 includes an information processing terminal 1, an information retrieval device 2, and an information storage device 3, as shown in FIG. The information retrieval device 2 is configured to be able to communicate with the information processing terminal 1 . The information storage device 3 is configured to be accessible by the information retrieval device 2 .

The information processing terminal 1 includes an input device 11 , a terminal communication device 12 and a display device 13 . The input device 11 is configured so that a search phrase can be input. The terminal communication device 12 is configured to be able to transmit the search phrase to the information search device 2 . The terminal communication device 12 is configured to be able to receive the search target searched based on the search phrase from the information search device 2 . The display device 13 is configured to be able to display search targets.

The information search device 2 includes a search communication device 21 , a position information acquisition section 22 , a feature word generation section 24 and a search section 25 . The search communication device 21 is configured to receive a search phrase from the information processing terminal 1 . The search communication device 21 is configured to be able to transmit search targets. The position information acquisition unit 22 is configured to be able to acquire position information. The position information indicates the position on the map that is the starting point of the search. The feature word generation unit 24 generates feature words from the search phrase. A feature word is associated with a word representing a search target. The search unit 25 is configured to search for a search target from the storage database.

The search target has point information. Point information indicates a specific point. The search unit 25 is configured to search for a search target corresponding to the feature word within a predetermined range from the position based on the position information and the feature word.

The information storage device 3 includes an accumulation storage device 31 . The accumulation storage device 31 stores an accumulation database. The accumulation database accumulates characteristic words and search targets. A search target corresponds to a feature word.

The information search system 10 of this embodiment can detect search targets with higher accuracy.

Hereinafter, each configuration in the information retrieval system 10 of this embodiment will be specifically described.

The information retrieval system 10 according to this embodiment is configured so that the information processing terminal 1 used by the user and the information retrieval device 2 can communicate with each other. The information processing terminal 1 and the information retrieval device 2 are connected by wireless communication via a public communication network 50, for example. In FIG. 1, the flow of information communicated via the public communication network 50 is exemplified by double-headed arrows.

The information processing terminal 1 includes an input device 11, a terminal communication device 12, and a display device 13, as well as a terminal control device 14 and a terminal storage device 15, as shown in FIG. The information processing terminal 1 is terminal equipment used by a user. Examples of terminal devices include mobile phones and smart phones.

The input device 11 is configured so that a search phrase can be input. A search phrase is a word for searching for a search target. Search phrases are also words arbitrarily entered by the user. A search phrase may be a sentence containing a word, a combination of words, or a subject and a predicate. The input device 11 is configured so that an input search phrase can be instructed to be transmitted from the terminal communication device 12 . The input device 11 is, for example, a touch panel integrated with the display device 13, a voice input device capable of voice input, or a physical key. The input device 11 is not limited to a configuration of a touch panel integrally configured with the display device 13, and may be configured only with an input unit.

The terminal communication device 12 is configured to be able to communicate with the information retrieval device 2 . The terminal communication device 12 transmits the information of the search phrase. The terminal communication device 12 receives information to be searched. Since the search target has information indicating a specific point, it is also point information. The point information is POI (Point Of Interest) information. Specifically, the terminal communication device 12 is configured to be able to communicate with the search communication device 21 of the information search device 2 using the public communication network 50 . The terminal communication device 12 may be configured to be capable of not only public communication using the public communication network 50 but also wired communication, for example. Further, the terminal communication device 12 may be configured to be capable of wireless communication conforming to communication standards such as infrared communication, Wi-Fi (registered trademark), Bluetooth (registered trademark), or BLE (Bluetooth Low Energy).

The display device 13 is configured to display various types of information. The display device 13 displays, for example, search phrases. The display device 13 may display a map. The display device 13 may display the points indicated by the search target point information searched by the information search device 2 so as to be superimposed on the map. The display device 13 can, for example, visualize a plurality of location information and display them superimposed on the map indicated by the map information stored in the terminal storage device 15 . The display device 13 is not limited to a configuration of a touch panel integrally configured with the input device 11, and may be configured only with a display unit.

The terminal control device 14 controls various information processing operations. The terminal control device 14 is electrically connected to the input device 11 , the terminal communication device 12 , the display device 13 and the terminal storage device 15 . The terminal control device 14 is configured to be able to control the input device 11 , the terminal communication device 12 and the display device 13 based on a program stored in the terminal storage device 15 . The terminal control device 14 is composed of, for example, a CPU (Central Processing Unit) or an MPU (Micro Processing Unit).

The terminal storage device 15 has a volatile storage device in which information disappears when power is turned off, and a non-volatile storage device in which information does not disappear even when power is turned off. The volatile storage temporarily stores information processed by the terminal controller 14 . The volatile storage device stores, for example, map information indicating a map and point information added to the map information. Volatile storage devices include, for example, RAM (Random Access Memory). A non-volatile memory device is a storage capable of storing various information processing programs and various types of information. Non-volatile storage devices include, for example, ROM (Read Only Memory). For ROM, for example, flash memory or HDD (Hard Disk Drive) is used.

The information retrieval device 2 includes a retrieval communication device 21 and a retrieval control device 27, as shown in FIG. The information retrieval device 2 includes a retrieval storage device 28 in addition to a retrieval communication device 21 and a retrieval control device 27 . The information retrieval device 2 is configured by, for example, a server. The search control device 27 has a position information acquisition section 22 , a category determination section 23 , a feature word generation section 24 , a search section 25 and a clock section 26 .

The search communication device 21 is configured to be able to communicate with the information processing terminal 1 . The search communication device 21 is configured to receive search phrase information from the information processing terminal 1 . The search communication device 21 is configured to be able to transmit search target information searched by the search control device 27 to the information processing terminal 1 . That is, the search communication device 21 receives search phrase information. The search communication device 21 transmits search target information. Specifically, the search communication device 21 is configured to be able to communicate with the terminal communication device 12 of the information processing terminal 1 using the public communication network 50 .

Also, the search communication device 21 is configured to be able to communicate with the information storage device 3 . The search communication device 21 transmits information for searching for a search target. The search communication device 21 receives the searched search target information from the information storage device 3 . Specifically, the search communication device 21
It is configured to be able to communicate with the storage communication device 33 of the information storage device 3 using the public communication network 50 . The search communication device 21 may be configured not only for public communication using the public communication network 50, but also for wired communication, for example. Further, the terminal communication device 12 may be configured to be capable of wireless communication corresponding to communication standards such as infrared communication, Wi-Fi (registered trademark), Bluetooth (registered trademark), or BLE.

The search control device 27 is electrically connected to the search communication device 21 and the search storage device 28 . The search control device 27 is configured such that a position information acquisition section 22, a category determination section 23, a feature word generation section 24, a search section 25, and a clock section 26 function. The search control device 27 drives, for example, a position information acquisition program, a category discrimination program, a feature word generation program, a search program, and a timer program stored in the search storage device 28 to operate. The search control device 27 is composed of, for example, a CPU or an MPU.

The position information acquisition program functions as the position information acquisition unit 22 that acquires position information by being executed on the computer. The category discrimination program functions as a category discrimination section 23 that discriminates the category of search phrases by being executed on a computer. The feature word generation program functions as a feature word generation unit 24 that generates feature words from search phrases by being executed on a computer. The search program functions as a search unit 25 that searches for a search target from the storage database stored in the information storage device 3 by being executed on the computer. The timekeeping program functions as a timekeeping unit 26 capable of timekeeping the year, day, and hour by being executed on the computer.

The position information acquisition unit 22 acquires position information. The location information is, for example, longitude information and latitude information. The location information may include altitude information in addition to longitude information and latitude information. The location information acquisition unit 22 can acquire location information transmitted from the information processing terminal 1, for example. Position information transmitted from the information processing terminal 1 includes, for example, position information input by the input device 11 of the information processing terminal 1 . Also, the position information acquisition unit 22 can acquire position information from a search phrase transmitted from the information processing terminal 1 .

A search phrase consists of text data. A search phrase contains a phrase. A phrase may be a single word or a group of words. Phrases may be written. The search phrase is converted into a format that allows text data to be handled mechanically. A search phrase is decomposed into a plurality of word units by morphological analysis. Here, morphological analysis can classify search phrases into the smallest meaningful units in natural language. Morphological analysis can also acquire part-of-speech information of classified morphemes. In morphological analysis, for example, it is also possible to obtain information as to whether words included in a search phrase are nouns, adjectives, verbs, or adverbs.

For morphological analysis, for example, a morphological analysis engine such as MeCab can be used. MeCab can be used with a named entity dictionary. Examples of named entity dictionaries include system dictionaries such as Neologd. The position information acquisition unit 22 morphologically analyzes the search phrase in order to obtain words included in the search phrase. The position information acquisition unit 22 is not limited to the configuration of morphologically analyzing the search phrase in order to obtain words included in the search phrase. The position information acquisition unit 22 may be configured to acquire words obtained by morphologically analyzing a search phrase by the category discrimination unit 23, the feature word generation unit 24, the search unit 25, or a morphological analysis unit (not shown), for example. The position information acquisition unit 22 may perform distributed representation in which the morphologically analyzed word is expressed numerically. The positional information acquiring unit 22 may collect the conjugations of the morphologically analyzed words. FastText, for example, may be used to summarize the conjugations of morphologically analyzed words.

When the word included in the search phrase is a place name, the position information acquisition unit 22 can acquire the position information from the place name database in which the place name and the position information are associated with each other. For example, the positional information acquiring unit 22 compares words obtained by breaking down the search phrase by morphological analysis with the place name database stored in the search storage device 28 . If a word obtained by decomposing a search phrase by morphological analysis matches a place name in the place name database, the position information acquisition unit 22 may acquire position information corresponding to the matching place name.

The geographical name database may consist of, for example, the names of prefectures, cities, wards, towns, villages, or facilities in association with location information. Here, the place name includes not only the name representing the municipality but also the name of the facility. In other words, a place name is a name indicating geographical information. In the place name database, for example, the name of the prefecture and the location of the prefectural office, the name of the city and the location of the city hall, the name of the ward and the location of the ward office, the name of the town and the location of the town office, the name of the village and the village office , the name of the facility and the location information of the facility are configured to correspond to each other. Hereinafter, when a word included in a search phrase is a place name, obtaining position information from a place name database in which place names and position information are directly associated is also referred to as obtaining position information by non-machine learning. When the position information acquisition unit 22 acquires position information by non-machine learning, it performs a hard search to determine whether or not it matches a word included in the database.

When a word included in a search phrase does not include a place name, the position information acquisition unit 22 can acquire position information from a database in which search words and position information are associated with each other by machine learning. For example, the position information acquisition unit 22 replaces the search phrase with words decomposed by morphological analysis and search words with a high degree of similarity by machine learning. The position information acquiring unit 22 can, for example, perform morphological analysis on the search phrase and select a desired search term based on the degree of similarity between the pre-stored word and the search term. The position information acquisition unit 22 compares the search word and the position information with a place name database associated with each other. If the search term matches a place name in the place name database, the position information acquisition unit 22 may acquire the position information corresponding to the matching place name. In other words, the location information acquisition unit 22 uses geocoding technology to convert a place name indicating geographical information into coordinate values of longitude and latitude.

The position information acquisition unit 22 uses at least one of the position information from the information processing terminal 1, the word of the place name included in the search phrase from the information processing terminal 1, and the search term based on the search phrase from the information processing terminal 1. Then, the position information indicating the position on the map can be acquired. The position information acquisition unit 22 uses information from the information processing terminal 1 to acquire position information indicating a position on the map. The information retrieval device 2 according to the present embodiment can indirectly acquire position information even when the position information cannot be acquired directly from the information processing terminal 1 . The information search device 2 can acquire position information with a lower processing load than a configuration that acquires position information only by machine learning.

That is, in the information retrieval device 2 according to the present embodiment, the position information acquisition section 22 acquires position information. Location information is estimated from words included in search phrases by machine learning.

The information search device 2 according to the present embodiment can indirectly estimate position information from words included in the search phrase by machine learning even if the search phrase does not include a word directly indicating the position information. can.

The category determination unit 23 is configured to be able to determine the category of search phrases. Category determination unit 23 determines the category of the search phrase transmitted from information processing terminal 1 . Categories include, for example, hot springs, parks, museums, shrines, live performances, shopping, amusement, outdoors, events, gourmet, ramen, and mountains. The categories are divided into appropriate types. The types of categories can be, for example, 300 types.

The category discrimination unit 23 morphologically analyzes the search phrase. The category determining unit 23 is not limited to the configuration of morphologically analyzing a search phrase in order to obtain words included in the search phrase. The category determination unit 23 may acquire words obtained by morphologically analyzing a search phrase by the position information acquisition unit 22, the feature word generation unit 24, the search unit 25, or a morphological analysis unit (not shown), for example. The category discriminating unit 23 may perform distributed representation in which the words subjected to morphological analysis are represented numerically. The category discrimination unit 23 may group conjugations of words subjected to morphological analysis.

When a word included in a search phrase is a category name representing a category, the category discriminating section 23 can directly discriminate the category from the category database in which the category name and the word are associated with each other. The category discrimination unit 23 compares words obtained by decomposing a search phrase by morphological analysis with the category database stored in the search storage device 28, for example. If the words obtained by decomposing the search phrase by morphological analysis match the category name of the category database, the category discrimination unit 23 may store the matching category name in the search storage device 28 in association with the search phrase. Hereinafter, when a word included in a search phrase is a category name, determining the category from the category database is also referred to as category determination. That is, it can be said that the category determination unit 23 includes the category determination unit 231 that determines the category from the category database by non-machine learning when the word included in the search phrase is the category name.

When a word included in a search phrase does not include a category name, the category determination unit 23 can use machine learning to estimate a category from a category database in which search words and category names are associated with each other. For example, the category determination unit 23 replaces a word obtained by decomposing a search phrase by morphological analysis by machine learning with a search word having a high degree of similarity with the word. In this case, the category determination unit 23 may replace one of the words obtained by decomposing the search phrase by morphological analysis with a search word highly similar to that word, or replace two or more words with those words. may be replaced with a search term having a high degree of similarity with In addition, the category determination unit 23 may, for example, extract co-occurring words that co-occur from two or more words, and replace the co-occurring words with search words that are highly similar to words clustered in advance for each category. . Here, co-occurrence means that a predetermined word included in a search phrase appears simultaneously with another word in the search phrase. A co-occurring word is another word that appears at the same time as a given word. Since the categories often use different words depending on the classification, the category discrimination unit 23 preferably performs category discrimination by machine learning using teacher data whose inputs and corresponding outputs are known in advance.

The category discrimination unit 23 can, for example, perform morphological analysis on the search phrase, and select a desired search word based on the degree of similarity between pre-stored words and co-occurring words. The desired search term is set to match the category name. The category discriminator 23 compares search words that match the category name with the category database. If the search word matches the category name of the category database, the category determination unit 23 may estimate that the matching category name is the category of the search phrase. That is, the category determination unit 23 includes a category estimation unit 232 that performs category estimation by machine learning from the category database when a search word similar to a word included in a search phrase is a category name. I can say.

The category discriminating unit 23 discriminates the category by using a word indicating the category included in the search phrase from the information processing terminal 1 or a search term based on the word included in the search phrase from the information processing terminal 1. can be done.

In the information retrieval device 2 according to the present embodiment, even if the category cannot be determined directly from the search phrase, the category can be determined indirectly based on the words included in the search phrase. When estimating a category by machine learning, the information retrieval device 2 can determine the category with higher accuracy than when determining the category by non-machine learning.

In other words, the information retrieval device 2 according to this embodiment further includes the category determination section 23 and the retrieval storage device 28 . The category discriminator 23 discriminates the category of the search phrase. Words corresponding to the categories are stored in the search storage device 28 . The category determination section 23 has a category determination section 231 and a category estimation section 232 . The category determination unit 231 determines the category based on whether or not the words included in the search phrase match the words stored in the search storage device 28 in advance. The category estimating unit 232 uses machine learning to estimate a category from words included in search phrases that have not been determined by the category determining unit 231 .

The information search device 2 according to the present embodiment can indirectly estimate the category from the words included in the search phrase by machine learning even if the search phrase does not include a word directly indicating the category. Since the information retrieval device 2 includes the category determining unit 231 and the category estimating unit 232, the category can be determined with a lower processing load than a configuration that determines the category only by machine learning.

The feature word generation unit 24 is configured to be able to generate feature words related to words representing search targets. The feature word generation unit 24 generates feature words from the search phrases transmitted from the information processing terminal 1 . The feature word generation unit 24 morphologically analyzes the search phrase in order to generate feature words from the search phrase. The feature word generation unit 24 obtains words obtained by morphologically analyzing a search phrase by the position information acquisition unit 22, the category determination unit 23, the search unit 25, or a morphological analysis unit (not shown), and generates feature words. may be configured. The feature word generation unit 24 performs distributed representation in which the morphologically analyzed words are represented numerically. The characteristic word generation unit 24 summarizes the conjugations of the morphologically analyzed words. The "feature word" may be one word or a group of words. Also, "generating a feature word" includes determining a feature word or selecting a feature word.

The feature word generation unit 24, for example, by machine learning, replaces a word obtained by decomposing a search phrase by morphological analysis with a feature word having a high degree of similarity with the word. The feature word generation unit 24 may replace one word with a feature word having a high degree of similarity among the words obtained by decomposing the search phrase by the morphological analysis, or replace two or more words with a degree of similarity with those words. It may be replaced with a high feature word. The feature word generation unit 24 uses machine learning to generate feature words having a higher degree of similarity to words representing a search target from words included in a search phrase. The feature word generator 24 can, for example, perform morphological analysis on a search phrase and select a desired feature word based on the degree of similarity between a pre-stored word and the feature word. More specifically, the feature word generation unit 24 can determine how far apart the feature vectors of the feature words are by measuring the cosine similarity. That is, the feature word generation unit 24 can determine the degree of similarity by measuring the feature vector of the feature word by cosine similarity.

A feature vector can be obtained, for example, by using an algorithm such as word2vec, using as input data the data obtained by numerically expressing a word subjected to morphological analysis. Algorithms such as word2vec used here are machine-learned so that feature vectors of input words can be output. For cosine similarity, the cosine of two vectors is calculated and the cosine value is used as the degree of similarity. Cosine similarity represents the closeness of angles between vectors, the closer to 1 the more similar. For cosine similarity, the closer it is to 0, the less similar it is. The feature word generation unit 24 may calculate statistical data of search phrases according to similarities. Preferably, the feature word generator 24 is configured to preferentially output a feature word with a higher frequency based on statistical data among the plurality of feature words.

The feature word generator 24 extracts co-occurring words from words included in the search phrase. In this case, the feature word generation unit 24 may use the co-occurring word as the feature word, and the search unit 25 may search for a search target from the feature word. The use of co-occurrence words by the characteristic word generation unit 24 enables effective derivation of search targets even from uncorrelated words that seem meaningless to represent search targets. In other words, the feature word generation unit 24 generates feature words, which are feature vectors that more easily represent a search target, from feature vectors by machine learning.

That is, the feature word generation unit 24 generates feature words related to words representing a search target from a search phrase. In the information retrieval device 2 according to the present embodiment, the feature word generation unit 24 generates feature words having a higher degree of similarity with words representing a search target from words including nouns, adjectives, verbs, and adverbs in a search phrase. do.

The information search device 2 can detect a search target with higher accuracy by generating a feature word having a higher degree of similarity with a word representing the search target. Especially when the words included in the search phrase are adjectives, the information retrieval device 2 generates characteristic words from words including at least adjectives, so it is possible to retrieve search targets that include subjective or emotional meanings. .

The search unit 25 is configured to search for a search target from the storage database. Feature words are classified by category in the accumulation database. The search target has point information indicating a specific point. A specific point represents a search target position on the map. The point information is, for example, longitude information and latitude information. The location information may include altitude information in addition to longitude information and latitude information. The search unit 25 acquires position information from the position information acquisition unit 22 . The search unit 25 acquires category information from the category determination unit 23 . The search unit 25 acquires feature words from the feature word generation unit 24 . The search unit 25 searches for a search target from the storage database based on the position information, category, and feature word. The search target searched by the search unit 25 corresponds to the feature word within a predetermined range from the position indicated by the position information and in the determined category. In other words, based on the position information, the category, and the feature word, the search unit 25 searches for a search target corresponding to the feature word within a predetermined range from the position indicated by the position information and in the determined category. configured to be able to

The clock unit 26 is configured to be able to clock the year, day, and hour. The timekeeping unit 26 may be configured to measure time by, for example, counting the number of clocks of the CPU. The clock unit 26 can specify, for example, the date and time when the search communication device 21 receives the search phrase from the information processing terminal 1 . The clock unit 26 can discriminate the search phrases received by the search communication device 21 for each predetermined period. The clock unit 26 can store the date and time when the search phrase was received in the search storage device 28 in association with the statistical data.

The search memory device 28 has a volatile memory device whose information disappears when the power is turned off, and a non-volatile memory device whose information does not disappear even when the power is turned off. Volatile storage temporarily stores information processed by search controller 27 . An example of a volatile memory device is, for example, RAM.

The nonvolatile storage device is a storage capable of storing a position information acquisition program, a category discrimination program, a feature word generation program, a search program, and a timer program. An example of a nonvolatile memory device is, for example, ROM. Also, the nonvolatile storage device may store a place name database, a category database, statistical data, or a system dictionary.

In the information retrieval system 10 of the present embodiment, for example, when the content of the search phrase is "restaurant with a beautiful night view", the feature word generation unit 24 uses morphological analysis to generate "night view", "beautiful", and "restaurant". to get the word The feature word generation unit 24 determines the degree of similarity of feature words co-occurring with "night view", "beautiful", or "restaurant" by machine learning. You can search for location information of "restaurants" whose search target is ambiguous.

The information storage device 3 includes a storage control device 32 and a storage communication device 33 in addition to the storage storage device 31 . The information storage device 3 is configured by, for example, a database server.

The accumulation storage device 31 stores an accumulation database in which search targets to be searched by the information search device 2 are accumulated. In the storage database, search targets are classified for each category. In the storage database, search targets are stored in association with feature words.

The storage storage device 31 has a volatile storage device in which information disappears when power is turned off, and a non-volatile storage device in which information does not disappear when power is turned off. The volatile storage temporarily stores information processed by the storage controller 32 . Volatile storage devices include, for example, RAM. A non-volatile storage device is a storage capable of storing various information processing programs and various types of information. Non-volatile storage devices include, for example, ROMs. A non-volatile storage stores an accumulation database. The non-volatile storage device may store map information indicating a map and point information to be added to the map information.

The accumulation control device 32 is electrically connected to the accumulation storage device 31 and the accumulation communication device 33 . The accumulation control device 32 is configured to be able to control the accumulation storage device 31 and the accumulation communication device 33 . The accumulation control device 32 can be configured using a CPU or an MPU. The accumulation control device 32 is configured, for example, to be able to implement a predetermined program based on a program stored in advance in the accumulation storage device 31 .

The storage communication device 33 is configured so that the information retrieval device 2 can access it. The storage communication device 33 is configured to receive information from the information search device 2 for searching for a search target. The storage communication device 33 is configured to be able to transmit information to be retrieved to the information retrieval device 2 .

Further, the information storage device 3 may be configured so as to construct a storage database from an SNS (Social Networking Service) provided by an external information processing device (not shown). The information storage device 3 constructs a storage database from the generated feature words in the same manner as the information retrieval device 2 generates feature words from search phrases based on the posted article information that the external information processing device has made public. can do. Posted article information is, for example, information of a posted article posted by a user to an external information processing device.

Next, the overall operation of the information retrieval system 10 according to this embodiment will be described.

First, the relationship between the information processing terminal 1 and the information retrieval device 2 will be described. In the information retrieval system 10 according to this embodiment, a user inputs a search phrase with the input device 11 of the information processing terminal 1 . When the input device 11 receives an instruction to transmit a search phrase, the information processing terminal 1 transmits the search phrase from the terminal communication device 12 to the information search device 2 .

Further, when the user selects a position on the map displayed on the display device 13, the information processing terminal 1 transmits position information indicating the selected position from the terminal communication device 12 to the information search device 2. may be configured. When receiving the position information from the information processing terminal 1, the information search device 2 can use the position information as a starting point on the map for search.

Next, the operation of information retrieval device 2 that receives a search phrase from information processing terminal 1 will be described with reference to FIG.

In the flowchart shown in FIG. 2, among the steps performed by the information retrieval device 2, the information retrieval device 2 receives a search phrase from the information processing terminal 1 (S101). The information search device 2 stores the received search phrase in the search storage device 28 .

The information search device 2 performs a position information acquisition step of acquiring position information by a position information acquisition program controlled by the search control device 27 (S102). A position information acquisition program acquires position information, when position information is transmitted from the information processing terminal 1 in a position information acquisition process. The location information acquisition program acquires location information from the search phrase when there is no acquirable location information in the location information acquisition step.

Next, the information retrieval device 2 performs a category determination step of determining the category of the search phrase by means of a category determination program controlled by the search control device 27 (S103). A category determination program determines a category name from a category database when a word indicating a place name is included in a search phrase. If the search phrase does not contain a word indicating a place name, the category discrimination program uses machine learning to generate a search term from the search phrase. The category discrimination program estimates the category name from the category database based on search words with a high degree of similarity among the generated search words.

Subsequently, the information retrieval device 2 performs a feature word generation step of generating feature words related to words representing a search target from the search phrase by means of a feature word generation program controlled by the search control device 27 (S104).

Next, the information retrieval device 2 performs a retrieval execution step of retrieving a retrieval target from the storage database by a retrieval program controlled by the retrieval control device 27 (S105).

Finally, the information retrieval device 2 may terminate after retrieving the retrieval target.

According to the information retrieval system 10 of the present embodiment, even if the user inputs a search phrase based on an ambiguous image into the information processing terminal 1, the information retrieval device 2 performs a flexible search using the user's subjective words or emotional words. It can be carried out. For example, even if a search phrase such as "super spicy ramen" is input, the information retrieval device 2 displays the location of a restaurant where you can eat super spicy ramen on the display device 13 of the information processing terminal 1. It can be displayed with a marker of the color corresponding to the category of ramen on the map.

That is, in the information retrieval system 10 of this embodiment, the information processing terminal 1 further includes a terminal storage device 15 . The terminal storage device 15 stores map information. The display device 13 displays a specific point indicating the search target on the map of the map information.

The information search system 10 of the present embodiment can display a more accurate search target point on a map.

In addition, in the information retrieval system 10 of the present embodiment, the feature word generator 24 calculates statistical data of search phrases. The information storage device 3 associates the search phrases received by the search communication device 21 from the information processing terminal 1 within a preset period with the statistical data, and stores them in the storage database.

The information search system 10 of this embodiment can detect search targets with higher accuracy based on search phrases within a preset period.

Furthermore, in the information retrieval system 10 of the present embodiment, the feature word generation unit 24 generates feature words having a higher degree of similarity with words representing a search target from words included in a search phrase. The feature word generator 24 is configured to preferentially output feature words with higher frequencies based on statistical data.

The information search system 10 of this embodiment can detect search targets with even higher accuracy.

In the information retrieval system 10 of the above-described embodiment, for example, when the information processing terminal 1 is configured to have a touch panel, the map displayed on the display device 13 is enlarged or reduced, and the user selects an area to be searched. The location information can be input to the input device 11 by tapping the place to be specified.

Further, the information processing terminal 1 displays the search target as a marker on the map on the display device 13 . The marker is, for example, a balloon in which a pictogram representing the category is displayed in the center. For example, the information processing terminal 1 may change the color or shade of the marker for each category and cause the display device 13 to display a map including the marker. The information processing terminal 1 may change the size of the marker according to the degree of similarity between the word included in the search phrase and the feature word, and cause the display device 13 to display a map including the marker.

Furthermore, in the information retrieval system 10 of the above embodiment, the information processing terminal 1 may have a navigation function. By having a navigation function, the information processing terminal 1 can guide the user to point information on the map displayed on the display device 13 . The navigation function is stored in the terminal storage device 15 of the information processing terminal 1 and controlled by the terminal control device 14 . Navigation functions may be stored in the search storage device 28 of the information search device 2 and controlled by the search control device 27 . In the information retrieval system 10, either the information processing terminal 1 or the information retrieval device 2 may be equipped with a navigation function as long as the display device 13 of the information processing terminal 1 can display a map and navigation.

Specifically, the navigation function may be stored in one of the information retrieval device 2 and the information processing terminal 1, and the navigation function may be controlled by the other device. Alternatively, the navigation function may be divided and stored in both the information retrieval device 2 and the information processing terminal 1, and either one of the devices or both devices may cooperate to control the navigation function. good.

Each processing or each function of the above embodiments may be realized by centralized processing by a single device or single system, or distributed processing by multiple devices or multiple systems. may be implemented with Further, each component of the above embodiment may be configured by dedicated hardware. Each component of the above embodiments may be implemented by executing a program for components that can be implemented by software.

Each component of the above-described embodiment may be implemented by, for example, a CPU executing a software program recorded on a recording medium. The program may be executed by being downloaded from a server, or may be executed by reading a program recorded on a predetermined recording medium. Also, the number of computers that execute the programs may be singular or plural. The above embodiments may be configured for centralized processing or may be configured for distributed processing.

However, the invention may be embodied in various other forms without departing from its spirit, scope or essential characteristics. The above-described embodiments are merely examples in all respects, and should not be construed in a restrictive manner. That is, the present invention is an information retrieval device or an information retrieval system, using the above embodiment as an example.

[Summary of embodiment]
(1: About feature words)
According to the information search device 2 described above, the function of each unit included in the search control device 27 allows the search phrase input by the user performing the search to be performed even if the search phrase does not directly indicate the spot the user wants to search. can also detect valid spots and present them to the user.

Specifically, the search control device 27 includes a feature word generation unit 24 and a search unit 25 . Of these, the feature word generation unit 24 generates feature words from words obtained by morphologically analyzing the search phrase. Hereinafter, words obtained by morphologically analyzing a search phrase may be referred to as input words. In this specification, the term "generate" is used with respect to feature words to mean "specify," "determine," "detect," or "select."

Here, the feature word generated by the feature word generation unit 24 may be a word highly similar to the input word, a co-occurring word of the input word, or both of them. good. Hereinafter, a word whose similarity to the input word is higher than a predetermined threshold may be referred to as a similar word.

When similar words are used as feature words, the feature word generation unit 24 generates feature information indicating features of the input word. This feature information may indicate the feature of the input word, and for example, the feature vector of the input word may be used as the feature information. The feature vector can also be calculated by using, for example, the GloVe, WordNet, or fastText algorithms, in addition to the word2vec described above. Then, the feature word generation unit 24 calculates the degree of similarity between the feature vector of each feature word associated with the spot in the accumulation database and the feature vector calculated above, and specifies words whose degree of similarity is equal to or greater than the threshold. . As a result, the feature word generation unit 24 can identify similar words similar to the input word among the feature words associated with the spots in the accumulation database. Note that the method of calculating the degree of similarity is not particularly limited, and can be calculated by cosine similarity, for example.

On the other hand, when co-occurring words are used as feature words, the feature word generation unit 24 identifies the co-occurring words of the input word. Co-occurring words can be extracted by, for example, TF-IDF (Term Frequency-Inverse Document Frequency). Further, the feature word generation unit 24 may use the co-occurring words of the above-described similar words as feature words.

Then, the search unit 25 searches the storage database stored in the storage storage device 31 using the generated feature word, and detects spots associated with the feature word in the storage database. Note that the "spot" means, for example, a point, a place, or a building installed at a specific point or place. In the accumulation database, each spot is associated with position information indicating the geographical position of the spot. As described above, the feature words used for searching are similar words or co-occurring words of the input word, so even if the search phrase does not directly indicate the spot that the user wants to search, It can be detected and presented to the user.

Of the words obtained by morphologically analyzing the search phrase, the feature word generation unit 24 may target words of a predetermined part of speech for feature word generation. The predetermined parts of speech may be, for example, nouns, adjectives, verbs and adverbs. This makes it possible to generate highly appropriate feature words by excluding parts of speech such as conjunctions, interjections, particles, and auxiliary verbs that are difficult to generate significant feature words.

(2: Location information)
The search control device 27 may include the position information acquisition section 22 . The position information acquisition unit 22 acquires position information that serves as a reference for determining a spot search range based on information received from the information processing terminal 1 used by the user. Then, the search unit 25 sets a predetermined range based on the position indicated by the acquired position information, and detects spots within the range. This makes it possible for the searching user to search for spots focused on the area of interest.

The position information acquisition unit 22 may acquire the position information from words obtained by morphologically analyzing the search phrase. For example, if words obtained by morphologically analyzing a search phrase include words indicating locations, the location information acquisition unit 22 uses a place name database in which words indicating locations and location information are associated with each other. location information. Note that, for example, a place name can be given as a word indicating a position.

Further, the position information acquisition unit 22 may acquire position information using a position information database for specifying position information. The positional information database is a database in which words are associated with positional information indicating positions corresponding to the words. When acquiring the position information using the position information database, the position information acquisition unit 22 uses at least one of a word obtained by morphological analysis of the search phrase, a similar word of the word, and a co-occurring word of the word. to search the location information database. As a result, the positional information acquiring unit 22 can acquire the positional information related to the word obtained by morphologically analyzing the search phrase. Similar words and co-occurring words may be identified using a machine-learned model. In this case, the position information acquisition unit 22 acquires position information estimated by machine learning from words included in the search phrase.

Information processing terminal 1 may acquire position information indicating its own position using, for example, a GPS (Global Positioning System) and transmit the acquired position information to information search device 2 . In this case, the positional information acquisition unit 22 may acquire the positional information transmitted from the information processing terminal 1 . By using such position information, it is also possible to detect spots around the current location of information processing terminal 1 .

(3: Categories)
The search control device 27 may include the category discrimination section 23 . In addition, category determination section 23 may include category determination section 231 and category estimation section 232 . The category determination unit 23 determines the category of the word obtained by morphologically analyzing the search phrase. In this case, each spot is associated with its category in the accumulation database. This enables the search unit 25 to detect spots of the category determined by the category determination unit 23 among the spots corresponding to the search phrase.

The category determination unit 231 can determine the category corresponding to the word obtained by morphologically analyzing the search phrase by using, for example, a category database that associates the word with the category corresponding to the word.

The category estimation unit 232 uses a machine-learned algorithm to specify a category corresponding to a word obtained by morphologically analyzing a search phrase. The above algorithm may be obtained by machine-learning the correspondence between word feature information and categories, for example. The word feature information is, for example, a feature vector. When using an algorithm that machine-learns the correspondence between word feature information and categories, the category estimation unit 232 acquires word feature information obtained by morphological analysis of a search phrase, and applies the acquired feature information to the above algorithm. A category can be specified by inputting. In addition, the category estimation unit 232 may specify at least one of similar words of words obtained by morphological analysis of the search phrase and co-occurrence words of words obtained by morphological analysis of the search phrase. Then, the category estimation unit 232 may identify the category by searching the category database using at least one of the identified similar words and co-occurring words. Similar words and co-occurring words may be identified using a machine-learned model. In this case, the category estimation unit 232 estimates the category from the words included in the search phrase by machine learning. Further, for example, the category estimation unit 232 may calculate the degree of similarity between a feature vector of a word belonging to a certain category and a feature vector of a word obtained by morphologically analyzing a search phrase. In this case, if the calculated similarity is equal to or greater than the threshold, the category estimation unit 232 determines that the word obtained by morphologically analyzing the search phrase belongs to the certain category.

The determined category may be used for weighting, which will be described later, or may be used for narrowing down the search range of the search unit 25 . In the latter case, the search unit 25 searches for spots belonging to the category determined by the category determination unit 23 among the spots stored in the storage database, and the feature words generated by the feature word generation unit 24 are associated with the spots. Detects spots where For example, consider a case where the category determination unit 23 determines that the category of the word "landscape" included in the search phrase is "sightseeing spot". In this case, the search unit 25 searches spots associated with the category of “sightseeing spot” among the spots accumulated in the accumulation database using the feature word generated from “landscape”. This makes it possible to appropriately narrow down the search range and quickly obtain appropriate search results.

(4: Priority in search)
If a plurality of feature words can be generated, the feature word generation unit 24 may generate a feature word with a high priority among the plurality of feature words. For example, in the information retrieval system 10, the priority is a ranking calculated based on the cumulative number of searches used in the past, the frequency of searches during a predetermined period in the past, and both or one of the cumulative number of times and the frequency of use. may be at least one of For example, when the characteristic word generating unit 24 can generate "scenery" and "restaurant" as characteristic words, the search frequency for "restaurant" in the most recent week is higher than the search frequency for "scenery" in the most recent week. If it is expensive, "restaurant" may be used as a feature word.

Note that the feature word generation unit 24 should calculate information indicating the number of times and frequency as described above for the search phrases received from the information processing terminal 1 within a preset period, and store the information as statistical data. Just do it. Also, the stored statistical data is associated with the search phrase received from the information processing terminal 1 .

Also, when the feature word generation unit 24 generates a plurality of feature words, the search unit 25 may perform a search using the feature word with the highest priority among the plurality of feature words. Further, the search unit 25 may preferentially output a search result using a feature word with a high priority over a search result using a feature word with a low priority. For example, the search unit 25 may output search results in descending order of priority using feature words. Also, the above statistical data can be used for weighting, which will be described later.

(5: Construction of accumulation database)
The information storage device 3 can acquire various types of information related to spots to be searched and construct an accumulation database. For example, the information storage device 3 may generate feature words from articles posted on SNS as described above, or articles posted on websites introducing spots and sites posting word-of-mouth information on spots. A feature word may be generated from Hereinafter, an article posted on an SNS or an article posted on a site that posts word-of-mouth information about a spot is also referred to as a comment. Moreover, a website may be abbreviated as a site.

For example, when the information storage device 3 acquires a comment "The view is the best" posted on a word-of-mouth site about the spot "Observatory X", the information storage device 3 morphologically analyzes the comment to add the words "View" and "Best". get the word Next, the information storage device 3 calculates feature vectors of these words, and uses the calculated feature vectors to generate feature words whose feature vectors are similar to each word. Then, the information storage device 3 adds the spot "Observatory X" to the storage database, and records the words obtained by the morphological analysis in association with the feature words generated from the words. If the spot "Observatory X" is already recorded in the accumulation database, the word obtained by the morphological analysis and the feature word generated from the word are additionally recorded in association with the already recorded spot. .

Also, instead of or in addition to the feature words, co-occurring words or clusters described below may be associated and recorded. In addition, on the above-mentioned sites, spots may be evaluated by points, stars, or the like. In that case, the information storage device 3 may also record the spot evaluation in the storage database. In addition, for example, the number of accesses to the site, the number of accesses to the page of the spot, the number of comments on the spot, the date and time of comment posting, and the name of the site may be recorded in the accumulation database. Note that the site name may be, for example, a service name. These pieces of information can be used, for example, for weighting spots, thereby enabling, for example, high-rated spots to be preferentially searched.

As described above, the information storage device 3 can expand the storage database based on the word-of-mouth information, thereby making it possible to search for new spots, and based on the impressions of people who have actually visited the spots. It is also possible to enable searching. For example, if a feature word is recorded based on the comment "The view is the best" for the spot "Observatory X", the vague search phrases such as "The feeling of liberation is the best" can also be detected.

Also, by constructing an accumulation database based on information related to spots, it is possible to detect spots that match the user's intentions even when an ambiguous search phrase is entered, as in the above example. It is possible. On the other hand, general search sites must include various targets other than spots in their search targets, so it is usually not possible to find spots that match the user's intentions with ambiguous search phrases. For example, even if you search for the phrase "the feeling of liberation is the best" on a general search site, the spot "Observatory X" may not be detected, or even if it is detected, it may be difficult to recognize because it is buried in other miscellaneous search results. state.

[Search using clusters]
As described in the above embodiment, the feature word generation unit 24 may set one word as a "feature word", or may set a word group consisting of a plurality of words as a "feature word". The latter case will be described in more detail below with reference to FIG.

When a word group consisting of a plurality of words is used as a feature word, in the accumulation database, the feature word group consisting of a plurality of feature words having similar feature information indicating the feature of the word is associated with the spot to be searched. . As mentioned above, the feature information is, for example, a feature vector. Also, hereinafter, the feature word group is called a cluster. When a search phrase is input, the feature word generator 24 determines a cluster corresponding to the word based on feature information of the word included in the search phrase. Then, the search unit 25 detects spots associated with the clusters determined by the feature word generation unit 24 from the accumulated database.

According to this configuration, even when a search phrase that is not directly associated with a spot is input, it is possible to detect a suitable spot that matches the intention of the user performing the search. For example, in the accumulation database, when a cluster A consisting of similar feature words such as "landscape" and "landscape" is associated with a spot A, it is assumed that a search phrase including the word "landscape" is entered. do. In this case, since "landscape" is similar to "landscape" and "landscape", it is determined that "landscape" belongs to cluster A, so spot A can be detected from the accumulation database. Whether or not the feature information is similar can be determined, for example, by determining whether or not the degree of similarity is equal to or greater than a predetermined threshold.

Further, when a group of words consisting of a plurality of words is used as a feature word, a category database may be used in which clusters consisting of a plurality of feature words having similar feature information indicating characteristics of words are associated with categories. In this case, the category determination unit 23 detects from the category database the category associated with the cluster determined based on the feature information of the words included in the search phrase. According to this configuration, even when a search phrase that is not directly associated with a category is input, it is possible to determine a proper category that matches the intention of the user performing the search.

FIG. 3 is a flowchart showing an example of processing for searching for spots based on clusters. A specific example of spot search according to the flow of FIG. 3 will be described below with reference to FIGS. 4 to 6. FIG. FIG. 4 is a diagram showing a specific example of determining clusters from search phrases, FIG. 5 is a diagram showing an example of a data table stored in a category database, and FIG. 6 is a diagram showing an example of a data table stored in a spot database. It is a figure which shows an example of the data table which exists. Note that the category database and the spot database correspond to the accumulation database described above. Also, the acquisition of the search phrase and position information (S201, S202) in FIG. 3 is the same as in FIG. 2, so the description is omitted here.

In S203, the feature word generation unit 24 calculates a feature vector of each word obtained by morphologically analyzing the search phrase. FIG. 4 shows an example in which four types of search phrases were entered: "beautiful scenery" as example 1, "beautiful scenery" as example 2, "beautiful scenery" as example 3, and "fun date" as example 4. ing. For example, when the search phrase of Example 1 is input, in S203, out of the words obtained by morphological analysis of "beautiful scenery," A feature vector is calculated. In the illustrated example, an n-dimensional feature vector is calculated. n may be a natural number, for example n=300.

In S204, the feature word generator 24 determines clusters of words obtained by morphologically analyzing the search phrase. Specifically, the feature word generation unit 24 inputs the feature vector calculated in S203 to a predetermined clustering model, thereby generating a cluster corresponding to the feature vector, that is, each word obtained by morphologically analyzing the search phrase. to get the output of clusters of . For example, in the example of FIG. 4, both "landscape" and "scenery" are determined to belong to the cluster with the cluster ID of "15". Also, both "beautiful" and "beautiful" are determined to belong to the cluster with the cluster ID of "8". Note that the clustering model is a machine-learned model that uses word feature vectors as input data and outputs clusters corresponding to the words. A cluster corresponding to a word is a cluster whose components include feature words whose feature vectors are similar to those of the word. A method of generating a clustering model will be described later.

In S205, the category determination unit 23 determines the category of each word obtained by morphologically analyzing the search phrase based on the cluster determined in S204. For this determination, a category database in which cluster IDs, feature words, weights, and categories are associated with each other can be used as shown in FIG.

Specifically, first, the category determination unit 23 extracts categories associated with the clusters determined in S204 from the category database. For example, in example 1 "beautiful scenery" in FIG. , the category "restaurant" is extracted.

Next, the category determination unit 23 calculates the degree of similarity between each word obtained by morphological analysis of the search phrase and each characteristic word of the category corresponding to the word, and multiplies the calculated degree of similarity by a weight. Note that a fastText inter-word similarity estimation model may be used to calculate the similarity. Then, the calculated values are totaled for each category, and the extracted categories are ranked in descending order of the total value. Hereinafter, the total value summed for each category is also referred to as score.

For example, in the case of example 1 "beautiful scenery" in FIG. is calculated. The same applies to the word "beautiful". Then, the calculated similarity is multiplied by the weight registered in the category database. For example, if the similarity between the word "scenery" and the characteristic word "landscape" is 0.8, and the similarity between the word "beautiful" and the characteristic word "beautiful" is 0.7, then the category "famous place/historic site" ” is calculated as follows:

For the word "scenery": similarity "0.8" x weight "0.88" = 0.704
For the word "beautiful": similarity "0.7" x weight "0.76" = 0.532
Score for the category "famous places/historic sites": 0.704 + 0.532 = 1.236
If a plurality of feature words with a cluster ID of "15" are registered in the category database, similarity calculation and weight multiplication are performed for each feature word in the same manner as described above. The same applies to cluster ID "8". Then, the values obtained by the multiplication are totaled for each category to calculate the score, and the extracted categories are ranked in descending order of the score.

In S206, the search unit 25 searches for spots based on the clusters determined in S204. For this search, a spot database in which cluster IDs, feature words, weights, and spot IDs are associated with each other can be used as shown in FIG.

Specifically, first, the search unit 25 extracts spots associated with the clusters determined in S204 from the spot database. For example, in example 1 “beautiful scenery” in FIG. "Restaurant" is extracted.

Next, the search unit 25 calculates the degree of similarity between each word obtained by morphologically analyzing the search phrase and each feature word of the spot corresponding to the word, and multiplies the calculated degree of similarity by a weight. Note that a fastText inter-word similarity estimation model may be used to calculate the similarity. and,
The calculated values are totaled for each spot, and the extracted spots are ranked in descending order of the total value. Hereinafter, the total value obtained by totaling the values calculated as described above for each spot is also referred to as a score.

For example, in the case of example 1 "beautiful scenery" in FIG. The similarity of "Landscape" is calculated. Note that the word "beautiful" matches the feature word "beautiful" registered in the spot database, so the degree of similarity is 1. Then, the calculated similarity is multiplied by the weight registered in the spot database. For example, if the similarity between the word "scenery" and the characteristic word "scenery" is 0.9, the score of the spot "famous place A" is calculated as follows.

For the word "scene": similarity "0.9" x weight "0.53" = 0.477
For the word "beautiful": similarity "1.0" x weight "0.46" = 0.46
The score of the spot "famous place A": 0.477 + 0.46 = 0.937
If a plurality of feature words with cluster ID "15" are registered in the spot database, similarity calculation and weight multiplication are performed for each feature word in the same manner as described above. The same applies to cluster ID "8". Then, the values obtained by the multiplication are totaled for each spot, and the extracted spots are ranked in descending order of the total score.

In S207, the search unit 25 performs final ranking of the spots detected in S206, taking into account the category determination result of S205. For example, the search unit 25 may rank the spots belonging to the category with the highest score in the category determination in S205 in descending order of the score calculated in S206. It should be noted that which category each spot belongs to is determined in advance.

Then, the search unit 25 ranks all the spots belonging to the category with the highest score, and then ranks all the spots belonging to the category with the next highest score. may be repeated until determined. Note that the search unit 25 may end the above-described repeated processing when a predetermined order is determined. For example, the top 20 ranks are given as the predetermined ranking.

Note that the ranking method based on category and spot scores is not limited to the above example. For example, for each spot detected in S206, the search unit 25 may calculate the total value of the score of the spot and the score of the category to which the spot belongs, and rank the spots in order of the total value. In this case, it can be said that the category score is used to weight the spot score. Moreover, the above-described statistical data may be added to the ranking in S207. In this case, for example, it is possible to make the ranking of spots with high evaluations higher.

Finally, in S209, the search communication device 21 notifies the information processing terminal 1 of the search result. Thereby, the terminal control device 14 of the information processing terminal 1 causes the display device 13 to display the search result. At this time, the search communication device 21 may notify the information processing terminal 1 of only a predetermined number of search results having the highest rank determined in S208. For example, the top 10 items in the search results can be used as the predetermined number of top items.

The search communication device 21 may also notify the information processing terminal 1 of the order determined in S208. In this case, the terminal control device 14 can cause the display device 13 to display the search results in a display mode according to the ranking. For example, when the search results are displayed on a map, a larger marker may be displayed for a spot with a higher rank on the map. This allows the user to recognize which of the detected spots is highly relevant.

As described above, by determining clusters corresponding to search phrases and searching for spots using the determined clusters, it is possible to detect appropriate spots from various search phrases. For example, even if the search phrase is example 1 in FIG. 4 “beautiful scenery”, example 2 “beautiful scenery” in FIG. spots can be detected. Further, when the search phrase is Example 4 "fun date" of FIG. is detected.

A cluster can also be used to specify a category, and by specifying a category, spots within that category can be detected. Furthermore, with respect to the combination of a spot and a feature word or the combination of a category and a feature word, by setting a higher weight for a combination with a higher degree of relevance, it is possible to preferentially detect a highly relevant spot.

(Modified example of category judgment/spot search)
In the category determination in S205, the degree of similarity between words obtained by morphologically analyzing the search phrase and feature words registered in the category database is calculated, but this process can be omitted. In this case, in S205, the category determination unit 23 may calculate the sum of the weight values of the cluster IDs associated with the category, and use the calculated value as the score of the category.

Similarly, the similarity calculation can be omitted in the spot search in S206. In this case, in S206, the search unit 25 may calculate the sum of the weight values of each cluster ID associated with the spot, and use the calculated value as the score of the category.

[Generation of category database and spot database]
(1: Configuration for realization)
The information retrieval system 10 may have a function of generating a category database and a spot database. Components for realizing this function may be provided in the information retrieval device 2 or the information storage device 3, for example. Also, a device having this function may be added to the information retrieval system 10 separately.

An example in which the information storage device 3 has a function of generating a category database and a spot database will be described below with reference to FIG. FIG. 7 is a block diagram showing an example of the main configuration of the information storage device. As shown, the accumulation control device 32 includes a clustering model generation unit 321, a feature word extraction unit 322, a cluster determination unit 323, a spot database generation unit 324, a category classification unit 325, and a category database generation unit 326. Below, the database may be abbreviated as DB. The functions of these units can also be realized by, for example, a program. Also, it is assumed that word-of-mouth data written on a predetermined site is used to generate the category database and the spot database. It is also assumed that one word-of-mouth includes at least one sentence. Here, the predetermined site may be a plurality of sites or SNS. In addition, word-of-mouth data may be abbreviated as word-of-mouth hereinafter.

The clustering model generation unit 321 generates a clustering model for clustering words. Clustering may be performed, for example, based on the degree of similarity of feature vectors, whereby roughly similar word groups can be treated as one cluster. It should be noted that the above-mentioned "roughly similar word group" can also be said to be a rough synonym group. A fastText vectorization model, for example, may be used to calculate feature vectors of words. The generated clustering model is used for cluster determination in S204 of FIG. By performing clustering, a search can be performed in units of clusters, so that it is possible to improve the efficiency of the search and speed up the search.

The feature word extraction unit 322 extracts feature words in the input sentence. Co-occurrence word analysis, for example, can be used to extract feature words. Specifically, the feature word extraction unit 322 calculates the TF-IDF value for each word included in the word group obtained by morphologically analyzing the input sentence. Then, the feature word extraction unit 322 ranks the words included in the word group according to the calculated TF-IDF value, and extracts the words ranked higher as feature words. Also, since the TF-IDF value is a numerical value indicating the certainty of the feature word, the feature word extraction unit 322 uses the TF-IDF value of the extracted feature word as the weight value of the feature word.

The cluster determination unit 323 determines clusters of feature words extracted by the feature word extraction unit 322 using the clustering model generated by the clustering model generation unit. The determined cluster is composed of a plurality of feature words having similar feature information.

The spot DB generator 324 associates spots, feature words, their clusters, and weight values to generate a spot database as shown in FIG. The feature words and weight values are extracted and calculated by the feature word extraction unit 322 . Also, the cluster is determined by the cluster determination unit 323 .

The category classification section 325 classifies the input sentence into a plurality of predetermined categories. A classification model constructed by supervised machine learning may be used for classification. As the classification model, for example, a fastText text classification model may be used. In the above machine learning, sentences with known categories are used as teacher data. For example, each word-of-mouth written on a predetermined site may be manually associated with a category and used as teacher data.

The category DB generator 326 associates categories, feature words, their clusters, and weight values to generate a category database as shown in FIG. The feature words and weight values are extracted and calculated by the feature word extraction unit 322 . Also, the cluster is determined by the cluster determination unit 323 .

(2: Flow of processing)
FIG. 8 shows an example of processing for generating a spot database. In S301, the clustering model generation unit 321 generates a clustering model. For example, the clustering model generation unit may calculate feature vectors for all words obtained by morphologically analyzing all word-of-mouth written on a predetermined site, and generate a clustering model based on the calculated feature vectors. . In this case, the generated clustering model is a model that classifies the input words into clusters whose feature vectors are similar to those of the words. FastText, for example, can be used to generate such a model.

In S302, the feature word extraction unit 322 extracts feature words and weights from the word-of-mouth of each spot. Feature term and weight extraction can be performed using TF-IDF values, as described above. Note that the correspondence relationship between each word of mouth and the spot is specified in advance according to which spot the word of mouth is written about. For example, if a word-of-mouth written on a page introducing a sight A on a predetermined site is stored in association with the sight A, the characteristic word extracting unit 322 recognizes the word-of-mouth as a word of mouth about the sight A. can be specified.

In S303, the cluster determination unit 323 uses the clustering model generated in S301 to determine clusters of the feature words extracted in S302. Characteristic words, weights, and clusters are determined for each word-of-mouth written for each spot by the processing of S302 and the processing of S303.

In S304, the spot DB generator 324 associates the feature words, weights, and clusters determined as described above with the corresponding spots and stores them in the spot database. As a result, for each spot, a record in which the feature word, cluster, and weight are associated is added to the spot database. Note that if the spot database has not yet been generated, the spot database is generated in S304.

Next, FIG. 9 will be described. FIG. 9 is a flow chart showing an example of processing for generating a category database. In S401, the category classification unit 325 classifies each word-of-mouth written on a predetermined site into categories.

In S402, the feature word extraction unit 322 extracts feature words and weights from word-of-mouth in each category. Feature term and weight extraction can be performed using TF-IDF values, as described above. In S402, the feature words and weights extracted in S302 may be used.

In S403, the cluster determination unit 323 uses the clustering model generated in S301 to determine clusters of the feature words extracted in S402. A feature word, a weight, and a cluster are determined for each word-of-mouth in each category by the processing of S302 and the processing of S403.

In S404, the category DB generation unit 326 associates the feature words, weights, and clusters determined as described above with their corresponding categories and stores them in the category database. As a result, for each category, a record in which feature words, clusters, and weights are associated with each other is added to the category database. If the category database has not yet been generated, the category database is generated in S404.

(3: Summary)
As described above, the clustering model generation unit 321 calculates feature vectors for each word obtained by morphologically analyzing word-of-mouth, and generates a clustering model based on the calculated feature vectors. According to the clustering model generated in this way, it is possible to accurately cluster search phrases containing expressions similar to word-of-mouth. In addition, this clustering model can also be generated only from word of mouth.

Then, by using the constructed clustering model and word of mouth, a spot database and a category database can be generated. This spot database and category database can also be generated only from word of mouth. In addition, these databases can be expanded by acquiring the latest word-of-mouth information from time to time. Word of mouth reflects the latest terms used to evaluate or describe spots and categories. It is also possible to respond to searches using

In addition, the cluster determination unit 323 determines clusters of feature words using the clustering model, and the spot DB generation unit 324 associates and stores the clusters determined by the cluster determination unit 323 and the spots. As a result, even when a search phrase made up of words that are not directly associated with spots is input, the search unit 25 can detect appropriate spots by using clusters corresponding to the words. be possible.

Similarly, the category DB generation unit 326 stores the clusters determined by the cluster determination unit 323 and the categories classified by the category classification unit 325 in association with each other. As a result, even when a search phrase composed of words not directly associated with a category is input, the category determination unit 23 can determine an appropriate category by using the cluster corresponding to the word. becomes possible.

Furthermore, the feature word extraction unit 322 extracts feature words and calculates their weights, and the spot DB generation unit 324 stores spots, clusters, feature words, and weights in association with each other. As a result, for each feature word included in one cluster, the degree of relevance to the spot can be specified by the weight value. Therefore, the search unit 25 can rank the search results in consideration of the degree of relevance between the feature word corresponding to the search phrase and the spot.

Similarly, the category DB generator 326 associates and stores categories, clusters, feature words, and weights. As a result, for each feature word associated with one category, the degree of relevance to the category can be specified by the weight value. Therefore, the category determination unit 23 can rank the category determination results in consideration of the degree of relevance between the characteristic word corresponding to the search phrase and the category.

(Regarding re-learning)
When a model generated by supervised machine learning is used in a search, teacher data may be generated according to the search results of the search, and machine learning may be performed again using the teacher data to update the model. For example, suppose that a certain search phrase is input, category determination (S204) and spot search (S206) are performed, and one of the spots notified to information processing terminal 1 is selected by the user. In this case, it is considered that the selected spot was the spot that the user intended. Therefore, the search phrase or the words forming the search phrase, their feature vectors, and the spot or its category can be associated with each other and used as teacher data. For example, by associating the search phrase with the category of the selected spot, the teacher data may be used to update the classification model used by the category classification unit. By updating the classification model by re-learning using such teacher data, the classification accuracy in S401 can be improved.

1 Information Processing Terminal 2 Information Retrieval Device 3 Information Storage Device 10 Information Retrieval System 11 Input Device 12 Terminal Communication Device 13 Display Device 15 Terminal Storage Device 21 Search Communication Device 22 Position Information Acquisition Part 23 Category Discrimination Part 24 Characteristic Word Generation Part 25 Search Part 28 Search storage device 31 Accumulation storage device

Claims (1)

A program for searching a spot belonging to a category associated with position information indicating a geographical position, using a computer as an information search device,
an obtaining step of obtaining a search phrase;
a location information acquisition step of acquiring the location information;
a calculating step of calculating a feature vector of each word obtained by morphologically analyzing the search phrase;
a determination step of determining each cluster containing a feature word having a high degree of similarity with each of the feature vectors calculated for each of the words;
a determination step of determining each of the categories corresponding to each of the determined clusters from a category database in which each of the clusters, the feature words, the first weighting factors, and the categories are associated;
Corresponding to each cluster determined from the spot database in which each cluster, the feature word, the second weighting factor, and the spot are associated, and included in a predetermined range from the position indicated by the acquired position information a searching step of searching for each of the said spots;
Calculate the similarity between each word and the feature word of each category, and calculate the first score based on a first score obtained by multiplying the calculated similarity by the first weighting factor Rank each category in descending order, calculate the similarity between each word and the feature word of each spot, and multiply the calculated similarity by the second weighting factor. and a ranking step of ranking each of the spots within each of the categories in descending order of the second score based on the second score.

Images