JP2017027168A - Taste learning method, taste learning program and taste learning device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To extract pieces of data which indicate a taste in an unified manner, from sentences created by plural users.SOLUTION: A taste learning method includes: a decomposition step for reading a sentence associated with a user, and decomposing the sentence into words; a feature word extraction step for extracting prescribed words from the sentence which is decomposed into words; and an interest object list generation step for using a table in which vectors indicating features of the words generated by using plural sentences are clustered, and association of representative words which are words close to gravity of respective clusters with words included in the respective clusters, is stored, and replacing the words extracted in the feature word extraction step with the representative words. A computer executes these steps.SELECTED DRAWING: Figure 3

Description

  The present invention relates to a preference learning method, a preference learning program, and a preference learning device.

  Conventionally, as a method for extracting a user's preference, for example, using a predetermined verb that may be used when a person shows interest, a verb and a noun are set from document data to be analyzed. A technique of extracting has been proposed (Patent Document 1).

JP 2011-180646 A

  Conventionally, a technique for extracting a user's interest from a sentence by extracting a predetermined word has been proposed. However, even users with similar preferences may be expressed in different sentences depending on synonyms and the like.

  Therefore, an object of the present invention is to extract data that represents preferences uniformly from sentences created by a plurality of users.

  The preference learning method according to the present invention includes a decomposition step of reading a sentence associated with a user and decomposing it into words, a feature word extracting step of extracting a predetermined word from the sentence decomposed into words, and a plurality of sentences. Using the table that stores the correspondence between the representative words that are words close to the center of gravity of each cluster and the words included in each cluster The computer executes an interest list generation step of replacing the word extracted in the extraction step with a representative word.

  By substituting with the representative word, it becomes possible to absorb the inconsistency of the script characters or terms freely described by the user and generate the preference information using the representative word that can be used uniformly. That is, data representing preferences in a unified manner can be extracted from sentences created by a plurality of users.

In addition, a step of acquiring a sentence associated with the identification information of the user from information disclosed by SNS (Social Networking Service), and using the parameter weighted for each SNS and each user for the acquired sentence A step of calculating an interest level, and in the interest list generation step, the interest level calculated for the sentence from which the word is extracted is stored in association with the representative word in which the word is replaced. Good. In this way, for example, the degree of interest can be obtained using a parameter weighted for each SNS according to how the user uses the SNS.

  Further, the degree of interest may be calculated using a parameter further weighted by the user's operation on the SNS, the presence / absence of an attached file of posting to the SNS, or the presence / absence of a link to an external site. In this way, it is possible to obtain the degree of interest taking into account the user's operation, the presence or absence of incidental information, and the like based on information published by an external site such as posting to SNS.

  The predetermined word used in the feature word extraction step may be a word related to a predetermined field. In this way, it becomes possible to learn the user's preference regarding a predetermined field.

  The sentence associated with the user is an answer to the question made to the user, the part of speech of the word expected as the answer to the question is defined, and the word expected as the answer in the decomposition step A word corresponding to the part of speech may be extracted as the content of the answer. In this way, the accuracy in extracting the target word from the sentence is improved.

  The contents described in the means for solving the problems can be combined as much as possible without departing from the problems and technical ideas of the present invention. The contents of the means for solving the problems can be provided as a device such as a computer or a system including a plurality of devices, a method executed by the computer, or a program executed by the computer. Note that a computer-readable recording medium may be provided.

Here, the computer-readable recording medium refers to a recording medium that accumulates information such as data and programs by electrical, magnetic, optical, mechanical, or chemical action and can be read from the computer. . Examples of such a recording medium that can be removed from the computer include a flexible disk, a magneto-optical disk, a CD-ROM, a CD-R / W, a DVD, a DAT, an 8 mm tape, and a memory card. Further, there are a hard disk, a ROM (Read Only Memory) and the like as a recording medium fixed to the computer.

  Data representing preferences in a unified manner can be extracted from sentences created by a plurality of users.

FIG. 1 is a diagram illustrating an example of a system configuration. FIG. 2 is a diagram illustrating an example of an agent. FIG. 3 is a functional block diagram illustrating an example of a preference analysis apparatus. FIG. 4 is a diagram illustrating an example of a document table stored in the text storage unit. FIG. 5 is a diagram illustrating an example of an answer table stored in the text storage unit. FIG. 6 is a diagram illustrating an example of a first parameter table stored in the parameter storage unit. FIG. 7 is a diagram illustrating an example of a second parameter table stored in the parameter storage unit. FIG. 8 is a diagram illustrating an example of a third parameter table stored in the parameter storage unit. FIG. 9 is a diagram illustrating an example of a word table stored in the word storage unit. FIG. 10 is a diagram illustrating an example of a feature word table stored in the word storage unit. FIG. 11 is a diagram illustrating an example of data stored in the term storage unit. FIG. 12 is a diagram illustrating an example of a semantic vector. FIG. 13 is a diagram illustrating an example of a list held in the representative word dictionary. FIG. 14 is a diagram illustrating an example of association between representative words and scores. FIG. 15 is an apparatus configuration diagram illustrating an example of a computer. FIG. 16 is a process flowchart showing an example of the document analysis process. FIG. 17 is a process flow diagram illustrating an example of an answer analysis process.

  Hereinafter, embodiments for carrying out the present invention will be described with reference to the drawings. In addition, the structure shown in embodiment is an illustration and this invention is not limited to the following structure.

<System configuration>
FIG. 1 is a diagram illustrating an example of a system configuration according to the present embodiment. The system of FIG. 1 includes a user device 1, a preference analysis device 2, and an SNS (Social Networking Service) providing device 3, which are connected to each other via a network 4 such as the Internet. In the present embodiment, it is assumed that a user installs a software agent (also referred to as an “agent program” or simply “agent”) that extracts a user's preferences and proposes a travel plan, for example, on his / her computer. A plurality of each component may exist. The user device 1 is a computer device carried by the user and executes the agent program according to the present embodiment. The agent program makes a question for learning the user's preference, and transmits, for example, an answer made in a sentence to the preference analysis device 2. In addition, the user may post a sentence to SNS or the like with his / her account. The preference analysis device 2 analyzes the user's answer, SNS posting, and the like, and extracts the user's preference. Note that the preference analysis device 2 or another device may propose a plan such as a trip to the user based on the extracted preferences. The SNS providing device 3 is a server device that provides a place where users communicate with each other like a so-called SNS. In the present embodiment, a service such as a miniblog is also included in the SNS.

FIG. 2 is a diagram illustrating an example of an agent execution screen according to the present embodiment. The agent according to the present embodiment is a mobile device such as a smartphone or a tablet PC (Personal Computer), a laptop PC, or a smart watch, which is an example of the user device 1, a stationary computer such as a desktop PC, or a kiosk. It is an application program (also simply referred to as “application”) that runs on a street computer such as a terminal or digital signage, a computer that provides services in a living space such as a so-called personal robot. It may be resident in the user device 1 like a so-called widget. The agent autonomously inputs and outputs information between the user and software such as an API. Specifically, the agent is information that is measured by a sensor included in the user device 1, a usage status of an application such as a camera installed in the user device 1, a user managed on the user device 1 or a so-called cloud. Information such as a schedule may be acquired to recognize a user's state and output information at a predetermined timing. In addition, the user device 1 includes a voice input unit such as a microphone and a voice output unit such as a speaker, and the agent inputs / outputs information by voice using the existing voice recognition technology and voice synthesis technology. May be performed. As shown in FIG. 2, a character image may be displayed on the display device of the user device 1.

<Functional configuration>
FIG. 3 is a functional block diagram illustrating an example of the preference analysis apparatus 2 according to the present embodiment. 3 includes a document acquisition unit 201, an answer acquisition unit 202, a text storage unit 203, a parameter storage unit 204, an interest level assignment unit 205, a morpheme analysis unit 206, and a word storage unit 207. A feature word extraction unit 208, a feature word storage unit 209, a specific field dictionary 210, a field limitation unit 211, a term storage unit 212, a representative word dictionary 213, a representative word replacement unit 214, and a representative word storage Unit 215, score determination unit 216, and preference information storage unit 217.

Based on the user account information stored in the SNS or the like stored in advance, the document acquisition unit 201 performs SNS on a sentence in which the user has posted to the SNS (including operations such as posting, favorite registration, and sharing (reposting)). Obtained from the providing device 3. In this embodiment, an entry posted by the user to the SNS is called a document. In addition, the answer acquisition unit 202 makes a question to the user via the agent of the user device 1 and receives the answer input by the user from the user device 1. Note that the user may freely input text data, or the text input by the user by voice may be converted into text data by the existing voice recognition in the user device 1. The text storage unit 203 includes a main storage device or an auxiliary storage device, and holds a text acquired as a document or an answer. For example, information as shown in FIG. 4 or 5 is stored in the text storage unit 203.

  FIG. 4 is a diagram illustrating an example of a document table stored in the text storage unit 203. The table of FIG. 4 includes items of user ID, date and time, SNS, operation, text, attachment, and interest level. Identification information for uniquely identifying the user is registered in the user ID field. It is assumed that identification information (not shown) used by each user in each SNS is also stored in association with the user ID. Information indicating the date and time when an operation such as posting is performed is registered in the date and time field. In the SNS field, identification information for uniquely identifying the SNS for which the user has performed an operation such as posting is registered. Information indicating operations performed by the user on the SNS such as posting, favorite registration, sharing (reposting), and the like is registered in the operation field. In the text field, text posted by the user is registered. Information indicating whether or not an attached file is added to the post is registered in the attached field. In the interest level field, an interest level value calculated by a process described later is registered.

  FIG. 5 is a diagram illustrating an example of an answer table stored in the text storage unit 203. The table of FIG. 5 includes items of question ID, question, answer category, user ID, answer, and interest level. Identification information for managing the question is registered in the question ID field. Also, in the question field, contents for questions from the agent to the user are registered. Questions are used to learn user preferences. In addition, information (words, parts of speech, etc.) related to words expected as answers is registered in advance in the answer classification. The questions in this embodiment are mainly questions that can be answered with yes or no (affirmation or denial), questions that can be answered with any of two or more options, so-called factoid questions that can be answered with words such as nouns or adjectives. And so on. Information indicating such a classification is registered in the answer classification of FIG. 5, and the information indicating the classification is used when recognizing the contents of the answer input from the user. In the user ID field, identification information for uniquely identifying the user is registered. The answer from the user is registered in the answer field. In the interest level field, an interest level value set by processing to be described later is registered.

The parameter storage unit 204 is composed of a main storage device, an auxiliary storage device, or the like, and stores in advance a parameter (also referred to as a first parameter) used for calculating the degree of interest for each user and for each SNS type. In other words, each user holds the degree of importance of each of the plurality of SNSs, for example, in numerical form. The value of the parameter can be determined based on, for example, the frequency with which the user posts SNS, the frequency with which the SNS is browsed, or the like. The parameter storage unit 204 stores a parameter (also referred to as a second parameter) used for calculating the degree of interest for each operation such as posting, favorite registration, and sharing (reposting), and the sentence is a URI (Uniform Resource Identifier). ) And whether or not an attached file (image file or the like) is included, and a parameter (also referred to as a third parameter) is stored. Then, the interest level assigning unit 205 calculates the interest level of the text acquired by the document acquisition unit 201 using the first parameter, the second parameter, the third parameter, and a predetermined mathematical formula, and the text storage unit The degree of interest is registered in association with the text stored in 203. On the other hand, for the sentence acquired by the answer acquisition unit 202, a predetermined value is assigned as the degree of interest. In addition, you may give a different value for every question made in order to obtain an answer.

Specifically, for posting to the SNS, the degree of interest may be obtained by, for example, the following formula (1).
Interest level = a ₀ + a ₁ × x ₁ + a ₂ × x ₂ + a ₃ × x ₃ (1)
A ₀ is a predetermined coefficient. Further, a _{1 to} a ₃ are the first parameter to the third described above.
Parameter. X ₁ is a weight parameter for weighting according to the type of SNS. For example, it is assumed that a parameter is set in advance such that the above-described degree of interest becomes higher for an SNS that is evaluated to be more important for each user. Further, x ₂ is post to SNS user, browsing, bookmark, a number or a value based thereon operations such as shared (action). For example, it is represented by a vector including a value indicating the number of operations for each type of operation. x ₃ is the number of additional information such as a URI or an attached file, or a value based on this. For example, each type of additional information is represented by a vector including a value indicating the number of postings including additional information.

x ₁ is a parameter corresponding to the type of SNS, and a value weighted according to, for example, the SNS usage status is defined for each user. FIG. 6 is a diagram illustrating an example of a first parameter table stored in the parameter storage unit 204 and holding the _first parameter (x ₁ ). The table in FIG. 6 includes items of user ID, SNS, and parameters. In the user ID field, identification information for uniquely identifying the user is registered. In the SNS field, identification information for uniquely specifying the SNS is registered. In the parameter 1 field, a weight parameter that is weighted in advance for each user and for each SNS is registered. For each user, a vector including weight parameters for a plurality of SNSs in a predetermined order is generated and used as the first parameter (x ₁ ).

x ₂ is the user is weighted in accordance with the operation performed in the SNS parameters.
FIG. 7 shows the second stored in the parameter storage unit 204 and holds the _second parameter (x ₂ ).
It is a figure which shows an example of a parameter table. The table of FIG. 7 includes items of user ID, SNS, operation, and parameter 2. In the user ID field, identification information for uniquely identifying the user is registered. In the SNS field, identification information for uniquely specifying the SNS is registered. In the operation field, operations performed by the user in the SNS, such as posting, favorite, sharing, link destination reference, display, and quotation, are registered. In the parameter 2 field, a second parameter weighted for each operation is registered in advance.

x ₃ is a parameter weighted according to an element accompanying the user's operation. FIG.
These are figures which show an example of the 3rd parameter table memorize | stored in the parameter memory | storage part 204 and hold | maintains the _3rd parameter (x3). The table of FIG. 8 includes items of user ID, SNS, additional information, and parameter 3. In the user ID field, identification information for uniquely identifying the user is registered. In the SNS field, identification information for uniquely specifying the SNS is registered. In the additional information field, a condition associated with the operation, such as a URI or an attached file, is registered. In the parameter 3 field, a third parameter weighted for each additional information is registered in advance.

  In the present embodiment, the degree of interest as described above is calculated as a value representing the degree of interest of the user with respect to a predetermined word included in the sentence based on a post made by the user in a predetermined SNS.

The morpheme analysis unit 206 in FIG. 3 reads the sentence associated with the user stored in the sentence storage unit 203, performs morpheme analysis, and breaks it down into words. The morpheme analysis unit 206 extracts general nouns and proper nouns using a dictionary for morpheme analysis (not shown) or a list of entry words of an online dictionary service that can be edited by an unspecified number of users. It also extracts proper nouns and date expressions that are not registered in the dictionary. As will be described later, a word that is not registered in the dictionary (also referred to as “unknown word”) can be extracted as a specific expression by analyzing the context. The word storage unit 207 includes a main storage device or an auxiliary storage device, and holds the words decomposed by the morpheme analysis unit 206. At this time, the degree of interest given to the sentence before decomposition is stored in association with each word after decomposition.

  FIG. 9 is a diagram illustrating an example of a word table stored in the word storage unit 207. The table of FIG. 9 is substantially the same as FIG. 4, but includes word items instead of text items. In the word field, words included in the above-described sentence, which are decomposed by morphological analysis, are registered.

  The feature word extraction unit 208 shown in FIG. 3 specifies a specific expression using a so-called dependency analyzer, eliminates duplication of words in the sentence, and extracts the specified specific expression as a feature word. Note that the dependency analyzer can determine whether or not it is a unique expression based on a word included in a sentence and a part of speech of the word (that is, based on a context). At the same time, the dependency analyzer can classify the content represented by the word determined to be a specific expression into predetermined items. For example, the extracted unique expressions can be divided into classification items such as a person, a place, and an action. Such a dependency analyzer can employ various existing methods. The feature word storage unit 209 includes a main storage device or an auxiliary storage device, and stores the extracted feature words in association with the above-described degree of interest.

  FIG. 10 is a diagram illustrating an example of a feature word table stored in the feature word storage unit 209. The table of FIG. 10 includes items of feature words, classification, and interest level. In the feature word field, the extracted feature words are registered. In the classification field, the classification items of the content represented by the feature word, such as a person, a place, and an action, classified by the dependency analyzer are registered.

  The specific field dictionary 210 is, for example, a list of words classified into specific fields of an online dictionary service that can be edited by an unspecified number of users. Note that the specific field dictionary 210 may be included in another device connected via the network 4 instead of the preference analysis device 2. In the present embodiment, it is assumed that a dictionary (not shown) of terms categorized in, for example, tourism as a target topic and a field related thereto is prepared. The field limiting unit 211 uses the specific field dictionary 210 to extract terms related to a predetermined field from the feature words stored in the feature word storage unit 209. The term storage unit 212 includes a main storage device or an auxiliary storage device, and stores the terms extracted by the field limiting unit 211. For example, words not registered in the specific field dictionary 210 are deleted from the feature word list shown in FIG. 10, and a list of terms related to the tourism field as shown in FIG. Stored.

  The representative word dictionary 213 includes a main storage device, an auxiliary storage device, or the like, and holds in advance a correspondence relationship between a feature word and a representative word having a similar meaning to the feature word. The representative word is stored in advance in association with the feature word at a predetermined timing.

Specifically, the feature words used by the user in a sentence in a predetermined field are sampled to generate a space vector (also referred to as “semantic vector” or “distributed expression”) representing the co-occurrence relationship. For example, as a sample, sentences referring to a target tourism field are collected without limiting the described users, and a corpus is generated. Further, based on the corpus, a semantic vector reflecting the semantic features of the word is generated using a technique such as Word2Vec. In the present embodiment, it is assumed that a semantic vector as shown in FIG. 12 including the above-described terms is generated in advance. The table of FIG. 12 includes items of words and semantic vectors, and semantic vectors are generated and stored for words obtained by decomposing sentences included in the corpus. Then, the semantic vectors are clustered to collect the feature words estimated to have similar meanings. Clustering can be performed using an existing technique such as the K-means method. Then, for example, the word closest to the centroid of the cluster whose feature vectors included in the same cluster are determined as a representative word, and a list in which the representative words and the feature words are associated is stored in the representative word dictionary 213. Hold. In the present embodiment, it is assumed that a list as shown in FIG. 13 is registered. In the example of FIG. 13, the word “camp” is associated with the representative word “outdoor”.

  Further, the representative word replacement unit 214 replaces the feature words stored in the feature word storage unit 209 with the representative words using information held in the representative word dictionary 213. By replacing with a representative word as described above, even if the word (feature word) used by the user in the sentence is different, the characteristic word with a similar meaning can be replaced with the representative word. It becomes possible to absorb the inconsistency. The representative word storage unit 215 includes a main storage device or an auxiliary storage device, and stores the representative words replaced by the representative word replacement unit 214 in association with the above-described degree of interest.

  The score determination unit 216 generates user preference information based on the representative word stored in the representative word storage unit 215 and the degree of interest, and stores it in the preference information storage unit 217 in association with the representative word. The preference information is represented by a set of combinations of representative words and scores generated based on the degree of interest. For example, when a plurality of feature words are replaced with the same representative word, the score is obtained by substituting the highest interest value associated with the feature word before replacement after replacing the representative word. The representative word score. In the present embodiment, the correspondence between representative words and scores as shown in FIG. 14 is stored. Here, the term “camp” in FIG. 11 is replaced with the representative word “outdoor”, and the records of high interest (with the same interest value in this embodiment) are narrowed down among the overlapping “outdoors”.

  If a set of combinations of representative words and scores generated in this way (that is, preference information) is used, information corresponding to the preference information can be provided to the user. For example, a user whose preference information shows a predetermined tendency may be extracted by so-called content-based filtering to provide some information. Further, for example, information on which other users have shown interest may be provided to users whose preference information has a similar tendency by so-called collaborative filtering.

<Device configuration>
Note that the user device 1, the preference analysis device 2, and the SNS providing device 3 are computers as shown in FIG. FIG. 15 is an apparatus configuration diagram illustrating an example of a computer. For example, the computer includes a CPU (Central Processing Unit) 1001, a main storage device 1002, an auxiliary storage device 1003, a communication IF (Interface) 1004, an input / output IF (Interface) 1005, a drive device 1006, and a communication bus 1007. The CPU 1001 performs processing described in this embodiment by executing a program. The main storage device 1002 caches programs and data read by the CPU 1001 and develops a work area of the CPU. Specifically, the main storage device is a RAM (Random Access Memory), a ROM (Read Only Memory), or the like. The auxiliary storage device 1003 stores programs executed by the CPU 1001, setting information used in the present embodiment, and the like. Specifically, the auxiliary storage device 1003 is an HDD (Hard-disk Drive), an SSD (Solid State Drive), a flash memory, or the like. The main storage device 1002 and the auxiliary storage device 1003 function as the preference information storage unit 103, the tourism information storage unit 107, the schedule information storage unit 108, and the like. The communication IF 1004 transmits / receives data to / from other computer devices. The communication IF 1004 is specifically a wired or wireless network card or the like. The input / output IF 1005 is connected to the input / output device and accepts input from the user or outputs information to the user. Specifically, the input / output device is an imaging device such as a camera, a keyboard, a mouse, a display, a touch panel, or a sensor such as a GPS receiver, a magnetic sensor, or an acceleration sensor. The drive device 1006 includes a flexible disk, a CD (Compact Disc), and a DVD (Digital Versatile Di).
sc), BD (Blu-ray (registered trademark) Disc), etc., data recorded on a storage medium is read or data is written to the storage medium. The above components are connected by a communication bus 1007. A plurality of these components may be provided, or some of the components (for example, the drive device 1006) may not be provided. Further, the input / output device may be integrated with the computer. The program executed in this embodiment may be provided via a portable storage medium that can be read by the drive device, an auxiliary storage device such as a USB memory, a network IF, or the like. Then, when the CPU 1001 executes the program, the computer as described above works as the preference analysis device 2. In addition, the aspect by which the said function is provided by the some apparatus which bears a part of said structure may be sufficient.

<Document analysis processing>
FIG. 16 is a processing flowchart showing an example of document analysis processing. The document acquisition unit 201 of the preference analysis device 2 monitors an operation performed by the user on the SNS, acquires a posted sentence when a posting is performed by an account used by the user in the SNS, and the sentence is acquired. Assume that the data is stored in the storage unit 203. Then, the interest level assigning unit 205 of the preference analysis device 2 is based on the type of SNS in which the text is posted, the operation performed by the user on the SNS, the presence of a predetermined element associated with the operation, and the like. Is calculated and stored in association with the text (FIG. 16: S1). In this step, for example, the degree of interest is obtained by the above-described equation (1). Then, the interest level value is registered in the data as shown in FIG. The morpheme analysis unit 206 performs so-called morpheme analysis on the sentence stored in the sentence storage unit 203, and decomposes it into words constituting the sentence (S2). In this step, data as shown in FIG. 9 is generated. In addition, the feature word extraction unit 208 extracts feature words that are general nouns and proper nouns from words, and the field limitation unit 211 stores a specific field dictionary that holds in advance the terms of the target field in the present embodiment. The feature words in the predetermined field are extracted and stored in the term storage unit 212 (S3). As described in the explanation of the field limiting unit 211, in this step, for example, only terms that are registered in a dictionary related to the tourism field are narrowed down. Then, a combination of a term and an interest level as shown in FIG. 11 is generated. Note that the process of the feature word extraction unit 208 may be omitted, and the field limitation unit 211 may narrow down the words in the word storage unit 207 to terms in a specific field. Further, the representative word replacement unit 214 replaces the term stored in the term storage unit 212 with the representative word based on the combination of the representative word and the term stored in advance in the representative word dictionary 213 (S4). And the score determination part 216 produces | generates the user preference information represented by the set of the combination of a representative word and a score based on the combination of a representative word and an interest degree, and stores it in preference information storage part 217 (S5). ). In other words, the preference information is data in which a target of the user's interest is listed and a score indicating the degree of interest is set.

  By replacing it with a representative word in S4, it becomes possible to absorb the inconsistency of the script characters or terms freely described by the user and generate preference information using representative words that can be used uniformly. In addition, as described for the representative word dictionary 213, for example, the word closest to the centroid of the clustered semantic vector is used as the representative word of the term included in the cluster, so that the term is replaced appropriately in terms of meaning. Can be done. Therefore, according to such a process, it becomes possible to extract data representing preferences uniformly from sentences created by a plurality of users.

<Response analysis process>
FIG. 17 is a process flow diagram illustrating an example of an answer analysis process. The answer acquisition unit 202 of the preference analysis device 2 causes the user device 1 to output a question to the user via the network 4 (FIG. 17: S11). In this step, for example, the question “Where is your favorite sightseeing spot you have been to?” Shown in Q1 of FIG. 5 is output. And the reply acquisition part 202 acquires a user's reply from the user apparatus 1 via the network 4, and memorize | stores it in the text memory | storage part 203 (S12). In this step, it is assumed that the user replied “Oh amusement park”. Questions and answers may be input / output by text or voice, or may be input / output by presenting and selecting options. For example, information is collected via an agent as shown in FIG. Further, the morpheme analysis unit 206 performs so-called morpheme analysis on the sentence stored in the sentence storage unit 203, and decomposes it into words constituting the sentence (S13). This step is the same as the document analysis process. For example, the above-described answer is decomposed into “○ × Amusement park” and “I”. In addition, the feature word extraction unit 208 extracts feature words that are general nouns and proper nouns from words, and the field limitation unit 211 stores a specific field dictionary that holds in advance the terms of the target field in the present embodiment. The feature words in the predetermined field are extracted and stored in the term storage unit 212 (S14). As described in the explanation of the field limiting unit 211, in this step, for example, only terms that are registered in a dictionary related to the tourism field are narrowed down. In this step, for example, “◯ × Amusement park” is extracted from the words generated in S13. In the answer analysis process, a predetermined interest level is associated with the extracted word. In addition, you may make it set the value of the interest level matched for every question. Then, the combination of the term and the degree of interest is stored in the term storage unit 212. Also in the answer analysis process, the processing of the feature word extraction unit 208 may be omitted, and the field limitation unit 211 may narrow down the words in the word storage unit 207 to terms in a specific field. Further, the representative word replacement unit 214 replaces the term stored in the term storage unit 212 with the representative word based on the combination of the representative word and the term stored in advance in the representative word dictionary 213 (S15). In the present embodiment, for example, it is assumed that the representative word “theme park” is registered in the representative word dictionary 213 in association with the term “Ox amusement park” and replaced. And the score determination part 216 produces | generates the user preference information represented by the set of the combination of a representative word and a score based on the combination of a representative word and an interest degree, and stores it in preference information storage part 217 (S16). ). Note that the degree of interest may not be set in S14, and a predetermined score may be set for the first time in S16. Also in the answer analysis process, the preference information is data in which, in other words, the user's interests are listed and a score indicating the degree of interest is set.

  Also in the answer analysis process, by replacing the term with a representative word in S15, the user can absorb the inconsistency of the answer script or term made by the sentence and generate preference information using a representative word that can be used uniformly. become able to. Also, in the answer analysis process, by asking questions that can be assumed in advance as part of speech that is assumed as answers, it is possible to improve the accuracy of extracting words related to user preferences from sentences.

<Modification>
In the embodiment described above, even if the same word appears in one post or answer, the feature word extraction unit 208 eliminates duplication. Also, even if the same word appears in different posts and answers, the score determination unit adopts the highest degree of interest. However, the present invention is not limited to this mode, and when a plurality of applications for the same word are applied, for example, a mathematical expression that increases the value of the degree of interest may be adopted.

  As described above, the question in the embodiment may be a question that can be answered in an affirmative or negative manner or a question that can be answered in any of two or more options. In this case, the user's preference may be expressed by a combination of a question and an answer, for example. Specifically, for example, user preference information can be generated in the form of a feature vector whose elements are answers to questions.

In the embodiment, the information stored in the table is exemplified by a general table including rows and columns, but the table design and format are not particularly limited. For example, the table design can be properly normalized in consideration of software performance and the like. The database format is not limited to RDBMS (Relational Database Management System), but a key-value management system such as so-called NoSQL may be adopted.

DESCRIPTION OF SYMBOLS 1 User apparatus 2 Preference analysis apparatus 201 Document acquisition part 202 Answer acquisition part 203 Text storage part 204 Parameter storage part 205 Interest degree provision part 206 Morphological analysis part 207 Word storage part 208 Feature word extraction part 209 Feature word storage part 210 Specific field dictionary 211 Field Limiting Unit 212 Term Storage Unit 213 Representative Word Dictionary 214 Representative Word Replacement Unit 215 Representative Word Storage Unit 216 Score Determination Unit 217 Preference Information Storage Unit 3 SNS Providing Device 4 Network

Claims (7)

Reading a sentence associated with the user and breaking it down into words;
A feature word extraction step of extracting a predetermined word from the sentence decomposed into words;
Clustering vectors representing the characteristics of words generated using multiple sentences, using a table that stores correspondence between representative words that are words near the center of gravity of each cluster and words included in each cluster An interest list generation step of replacing the word extracted in the feature word extraction step with the representative word;
A preference learning method that the computer executes. Obtaining a sentence associated with the user from information published by SNS (Social Networking Service);
For the acquired sentence, calculating a predetermined degree of interest using a parameter weighted for each SNS and for each user;
Further including
The preference learning method according to claim 1, wherein in the interest list generation step, the degree of interest calculated for the sentence from which the word is extracted is stored in association with the representative word in which the word is replaced. The preference according to claim 2, wherein the degree of interest is calculated using a parameter further weighted according to a user operation on the SNS, presence / absence of a file attached to the SNS, or presence / absence of a link to an external site. Learning method. The preference learning method according to any one of claims 1 to 3, wherein the predetermined word used in the feature word extraction step is a word related to a predetermined field. The sentence associated with the user is an answer to a question made to the user,
The preference learning according to claim 1, wherein a part of speech of a word expected as an answer to the question is defined, and a word corresponding to the part of speech of the word expected as the answer is extracted as the content of the answer in the decomposition step. Method. Reading a sentence associated with the user and breaking it down into words;
A feature word extraction step of extracting a predetermined word from the sentence decomposed into words;
Clustering vectors representing the characteristics of words generated using multiple sentences, using a table that stores correspondence between representative words that are words near the center of gravity of each cluster and words included in each cluster An interest list generation step of replacing the word extracted in the feature word extraction step with the representative word;
A preference learning program that causes a computer to execute. Reading a sentence associated with the user and breaking it into words,
A feature word extraction unit that extracts a predetermined word from a sentence decomposed into words;
Clustering vectors representing the characteristics of words generated using multiple sentences, using a table that stores correspondence between representative words that are words near the center of gravity of each cluster and words included in each cluster An interest list generator that replaces the word extracted in the feature word extraction step with the representative word;
Preference learning device.

Images