JP2015018491A - Information processing device and method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To use user voice of which attributes are unknown, as teacher data for machine learning.SOLUTION: An information processing device comprises: voice acquisition means which acquires voice uttered by a user; attribute estimation means which estimates attributes of the acquired voice by using a discriminator having learned input data and teacher data of which attributes are already known; operation reception means which receives an operation for an advertisement that is associated with attributes and output for the user; and learning means which updates parameters of the discriminator by using combination of the attributes of the operated advertisement and the acquired voice.

Description

  The present invention relates to an online advertisement based on machine learning.

  In smart devices, voice recognition is widely used instead of character input for input of search terms and other purposes. At that time, in addition to the search term representing the utterance content from the spoken voice, the attribute is estimated using a discriminator for identifying the gender, age, etc. of the user, and the result is used for the extraction of the advertisement. It is possible to present an advertisement suitable for. Conventionally, as the teacher data of the classifier for estimating the attribute from the voice, the voice of the user whose attribute is known has been used (see, for example, Patent Documents 1 and 2).

JP 2012-029099 A JP 2012-002856 A

  However, the number of users whose attributes are known based on user registration, access history, and other information is limited. The utterances by many other users cannot be used as teacher data, and there is a limit to improving accuracy.

  An object of the present invention is to utilize voices of users whose attributes are unknown as teacher data for machine learning.

  Based on the above object, an information processing apparatus according to one aspect (1) of the present invention includes a voice acquisition unit that acquires a voice uttered by a user, input data having a known attribute in advance and a teacher from the acquired voice. An attribute estimation unit that estimates an attribute using a discriminator learned using data, an operation reception unit that receives an operation associated with the attribute and an advertisement output to the user, And learning means for updating parameters of the discriminator using a set of the attribute relating to the advertisement and the acquired voice.

  An information processing method according to another aspect (5) of the present invention is obtained by capturing the above aspect in a method category. From an acquired voice, a voice acquisition process for acquiring voice uttered by a user in advance. An attribute estimation process that estimates an attribute using a classifier that is learned using input data with known attributes and teacher data, and an operation that is associated with the attribute and that is output to the user is received. The computer executes an operation reception process, a learning process for updating a parameter of the classifier using a set of the attribute related to the operated advertisement and the acquired voice.

  According to another aspect (2) of the present invention, in any one of the above aspects, the attribute is gender.

  According to another aspect (3) of the present invention, in any one of the above aspects, the speech recognition unit includes a speech recognition unit that recognizes a word from the acquired speech, and the advertisement selection unit is based on the speech-recognized word. , Selecting an advertisement to be output from among the advertisements stored in the advertisement storage means, and the learning means determines a set of the attribute and the voice associated with the operated advertisement as the parameter It is used for updating.

  According to another aspect (4) of the present invention, in any one of the aspects described above, the aspect further includes speech recognition means for recognizing a word from the acquired speech, and the advertisement selection means is configured to recognize the estimated attribute or speech. Based on at least one of the words, an advertisement to be output is selected from the advertisements stored in the advertisement storage unit, and the learning unit includes the attribute associated with the operated advertisement and When the estimated attribute matches, the speech and attribute pair is used for updating the parameter.

  According to the present invention, the voice of a user whose attribute is unknown can be used as teacher data for machine learning.

The functional block diagram which shows a structure about embodiment of this invention. The figure which shows the example of the data in embodiment of this invention. The figure which shows the example of the data in embodiment of this invention. The flowchart which shows the process sequence in embodiment of this invention. The conceptual diagram which shows the whole process in embodiment of this invention. The conceptual diagram which shows an example of the process in embodiment of this invention. The conceptual diagram which shows an example of the process in embodiment of this invention.

  Next, a mode for carrying out the present invention (referred to as “embodiment”) will be illustrated along the drawings. In addition, the assumptions common to the contents already described in the background art and problems are omitted as appropriate.

[1. Constitution〕
FIG. 1 shows a configuration of an information processing apparatus (also referred to as “present apparatus 1”) in the present embodiment. In response to a search request received by speech recognition from the terminal T, the device 1 includes at least a speech-recognized word and an attribute (for example, gender) that is estimated from speech using a result of machine learning. On the other hand, it is a server device that sends a search result with a selected advertisement to the terminal T and performs machine learning according to an operation (for example, a selection operation such as a mouse click or a screen tap) for the sent advertisement. The device 1 may be realized by combining devices or systems for each function or means.

  The terminal T is a smart device (for example, a smart phone, a tablet PC, or the like) that accesses the apparatus 1 via a communication network N (the Internet, a mobile phone network, or the like), and there are many terminals T corresponding to many users.

  The apparatus 1 includes a computer configuration, that is, an arithmetic control unit 6 such as a CPU, a storage device 7 such as a main memory and an auxiliary storage device, a communication device 8 (such as a communication device and a communication adapter) with a communication network N, Is provided. The terminal T also has a computer configuration (not shown), although the specifications are different, and also has a configuration as a mobile terminal (not shown) such as a microphone, a receiving microphone and a speaker.

  In the present apparatus 1, each element shown in FIG. 1 is realized by the arithmetic control unit 6 executing a computer program (not shown) stored in the storage device 7.

  Among the realized elements, the information storage means is not limited to the so-called local storage in the apparatus 1 but may be remote storage by network computing (cloud) or the like. Further, the storage means shown in the present application is a unit that is convenient for explanation and is a main one. The actual storage means may include functions such as input / output and management associated with information storage, the unit of the configuration may be divided or integrated, and other storage means such as a work area may be used as appropriate. Good.

  Of the storage means, the advertisement storage means 45 is means for storing advertisements associated with attributes (for example, FIG. 2). The advertisement illustrated in FIG. 2 is a search-linked advertisement that considers user attributes. For example, the advertisement with the advertisement ID “A01” includes a designated word (for example, “tea”) in response to a search request from the terminal T, or the user attribute matches the advertisement attribute (here, “female”). , Selected as an advertisement to be output.

  In the above example, the output advertisement is an advertisement sentence (for example, “tea free shipping ...” etc.) with a hyperlink to the URL (http: // tea...). As described above, an attribute associated with an advertisement as an output condition is referred to as an “advertisement attribute”. A plurality of advertisement attributes may correspond to one advertisement.

  The teacher data storage means 75 is means for storing teacher data (broad sense teacher data) that is data used for learning in machine learning (for example, FIG. 3). The unit of the teacher data is a set of data having the spoken voice data as input data and the attributes that are preferably output when the voice data is given, that is, teacher data in a narrow sense.

  As illustrated in FIG. 3, the teacher data includes the existing data used in the previous learning (in the example of FIG. 3, “0” indicating existing in the example of FIG. 3), and additional data added after the previous learning. ("Addition flag" is "1" indicating new)).

  Note that the arrows in FIG. 1 and other drawings are examples of main directions of data or control and other flows, and do not mean the denial of other flows or limitation of directions. In addition, each means other than the storage means is a processing means for realizing or executing the following information processing functions or operations, but these functions or operations are units for explanation only, actual hardware and software. The elements of are not limited.

[2. Action)
FIG. 4 is a flowchart showing the operation of the apparatus 1. FIG. 5 is a conceptual diagram showing the entire processing in the present embodiment, and FIGS. 6 and 7 are conceptual diagrams showing examples of processing in the present embodiment. First, an outline of the operation will be described along the flowchart of FIG. 4 with some steps omitted.

[2-1. Overview〕
When the user of the terminal T performs a search by voice recognition (for example, web search, image search, dictionary search, product search, miniblog post search, etc.), the terminal T transmits the spoken voice data to the apparatus 1. .

  In this apparatus 1, the voice acquisition unit 20 acquires the voice uttered by the user from the terminal T (step S11), and the attribute estimation unit 30 determines the classifier having the optimized parameter P from the acquired voice. Is used to estimate the attribute (step S12).

  The algorithm used for optimizing the discriminator and the parameter P may be any method that uses input data, its teacher data (narrow sense teacher data), and (both in a broad sense teacher data) (supervised learning). . For example, when gender or the like is determined from speech data, a GMM (Gaussian Mixture Model) is widely used as a discriminator, and an EM algorithm is used as the machine learning for optimizing the parameters.

  The teacher data is used to optimize the parameters of the GMM that is a discriminator. For example, speech data (input data) in which the narrowly-sensed teacher data is male is used for learning the GMM for identifying the attribute of male. It is done. On the other hand, voice data (input data) whose teacher data in the narrow sense is female is used for optimization of GMM parameters for identifying the attribute of female, as in the case of male. Here, let the optimized parameter be the above-mentioned parameter P.

  When estimating the attributes of the input speech, the input speech is collated (score calculation) with a GMM optimized with male speech data and a GMM optimized with female speech data. The gender that is close as a result of the collation (having a high score) will be referred to as an “estimated attribute”. Here, gender is taken as an example, but it can also be applied to other attributes such as age.

  On the other hand, the voice recognition means 32 recognizes a word from the acquired voice (step S13). The word recognized in this way is called a “recognized word”. Speech recognition is based on a known technique (for example, a statistical method, a dynamic time stretching method, a hidden Markov model, or the like). One of the uses of the recognition word is a search keyword. The search means 35 uses the index data stored in the index storage means 36 to perform a search using the recognition word as a key to obtain a search result (step S14).

  Moreover, the advertisement selection means 42 selects the advertisement output with a search result from the advertisements memorize | stored in the advertisement memory | storage means 45 (step S15). The advertisement output means 50 outputs the selected advertisement to the user by transmitting it to the terminal T in combination with the search result by the search means 35 (step S16).

  The operation accepting means 60 accepts a click on the output advertisement (meaning an operation on the advertisement, including not only a mouse click but also a tap or the like) (step S17).

  When the advertisement is clicked (step S17: “YES”), the learning means 72 learns machine learning, that is, relearning, a set of the attribute related to the clicked advertisement and the acquired voice (steps S19 and S21). (Steps S22 and S23). More specifically, when a large number of teacher data in a broad sense, which is a set of attributes (narrow sense teacher data) related to the clicked advertisement and acquired voice (input data), are stored in the teacher data storage means 75 as additional data, a predetermined number or more. (Step S22: “YES”), relearning is performed (step S23).

  The “attribute related to the clicked advertisement” used for learning may be, for example, an advertisement attribute, an estimated attribute that is a basis for selecting the advertisement, or the like. Two examples are shown below.

[2-2. Attributes used for learning)
The first example is learning based on a match between an estimated attribute and an advertisement attribute (FIG. 6). In this case, the advertisement selection unit 42 selects an advertisement to be output from among the advertisements stored in the advertisement storage unit 45 based on at least one of the estimated attribute or the recognition word (step S15).

  When the advertisement attribute and the estimated attribute of the clicked advertisement match (step S18: “YES”), the learning unit 72 learns the pair of the matching attribute and the voice (step S19). Used (steps S22 and S23). For example, in the example of FIG. 6, the estimated attribute and the advertisement attribute of the advertisement based on the recognition word are all the same for women. Use.

  The second example is learning that prioritizes advertisement attributes (FIG. 7). In the flow of FIG. 4, when the advertisement attribute and the estimated attribute do not match (step S18: “NO”), when the advertisement attribute is set to be prioritized (step S20: “YES”), in the second example, Learning (FIG. 7) giving priority to a certain advertisement attribute is performed (steps S21 to S23).

  Learning that prioritizes advertisement attributes is effective when the advertisement selection means 42 selects an advertisement to be output from advertisements stored in the advertisement storage means 45 based on the recognition word (step S14). ). In this case, the advertisement attribute of the advertisement selected based on the recognition word does not necessarily match the estimated attribute based on the voice.

  And the learning means 72 uses the group of the advertisement attribute and audio | voice of the clicked advertisement (step S21) for learning (step S22 and S23). For example, in the example of FIG. 7, the estimated attribute (here, “male”) is ignored, and the advertisement attribute (“female”) based on the recognition word (for example, “tea”) and the voice are used for learning. .

[3. effect〕
(1) As described above, in the present embodiment, the voice uttered by the user whose attribute is unknown is combined with the attribute related to the operated advertisement (FIGS. 5 to 7), and is suitable for the purpose of selecting the advertisement. Can be used as machine learning teacher data.

  In particular, for the purpose of selecting an advertisement to be output to a spoken user, the attribute estimated from speech need not exactly match the user's true attribute (eg, biological gender). For example, the voice of a person who tends to manipulate advertisements for men even for men is suitable for the purpose of advertisement selection even if it is used for learning as teacher data to be estimated as women.

(2) Further, in this embodiment, by enhancing machine learning for the sex that is the most basic user attribute, the advantage of improving the accuracy of attribute estimation can be exerted in many application fields.

(3) Also, in this embodiment, by using a set of attributes (advertising attributes) and speech corresponding to the operated advertisement for machine learning (FIG. 6), learning suitable for the purpose of advertisement selection can be realized. Improvement of CTR can be expected.

(4) In this embodiment, even if the basis for selecting an advertisement is a speech-recognized word or an estimated attribute, the attribute (advertisement attribute) corresponding to the operated advertisement is estimated. If the attribute matches the attribute, the pair of the attribute and speech is used for learning (FIG. 7), so that highly reliable teacher data can be increased.

[4. Other embodiments]
In addition, the content of the said embodiment and figure is only an illustration, and the presence or absence and arrangement | positioning of each element, the order and content of a process, etc. can be changed suitably. For this reason, this invention also includes the modification illustrated below and other embodiment other than that. For example, the attribute is not limited to gender, but may be age, residence, home region, or the like.

  Moreover, each aspect of the present invention can be understood as other categories (methods, programs, systems including terminals, etc.) that are not specified. In the category of method or program, “means” shown in the category of apparatus is appropriately read as “process” or “step”. In addition, all or any part of “means” can be read as “part” (unit, section, module, etc.).

  Also, the processes and steps shown in the embodiment can be changed by changing the order, executing some of them collectively, or executing them part by part. In addition, hardware elements that implement and execute individual means, processes, and steps may be common, or may differ for each means, process, step, or timing.

  The individual means shown in the present application may be realized by calling a function provided by an external server by an API (Application Program Interface) or network computing (so-called cloud or the like). Furthermore, elements such as means are not limited to computers, and may be realized by other information processing mechanisms that appear now or in the future.

1. Information processing device (this device)
6 arithmetic control unit 7 storage device 8 communication device 20 voice acquisition means 30 attribute estimation means 32 voice recognition means 35 search means 36 index storage means 42 advertisement selection means 45 advertisement storage means 50 advertisement output means 60 operation acceptance means 72 learning means 75 teacher Data storage means N Communication network T Terminal

Claims (5)

Voice acquisition means for acquiring voice spoken by the user;
Attribute estimation means for estimating an attribute from the acquired speech using a discriminator previously learned using input data and teacher data with known attributes;
Operation accepting means for accepting an operation associated with an attribute and an advertisement output to the user;
Learning means for updating the parameters of the classifier using a set of the attribute related to the operated advertisement and the acquired voice;
An information processing apparatus comprising:   The information processing apparatus according to claim 1, wherein the attribute is gender. Comprising speech recognition means for recognizing words from the acquired speech;
The advertisement selection means selects an advertisement to be output from the advertisements stored in the advertisement storage means based on the speech-recognized word,
The information processing apparatus according to claim 1, wherein the learning unit uses a set of the attribute and the voice associated with the operated advertisement for updating the parameter. Comprising speech recognition means for recognizing words from the acquired speech;
The advertisement selection means selects an advertisement to be output from the advertisements stored in the advertisement storage means based on at least one of the estimated attribute or the speech-recognized word.
The learning means uses the voice and the attribute set for updating the parameter when the attribute associated with the operated advertisement matches the estimated attribute. The information processing apparatus according to claim 1 or 2. Voice acquisition processing for acquiring voice spoken by the user;
Attribute estimation processing for estimating an attribute from the acquired speech using a discriminator previously learned using input data and teacher data with known attributes;
An operation reception process for receiving an operation associated with an attribute and an advertisement output to the user;
A learning process for updating the parameters of the classifier using a set of the attribute relating to the operated advertisement and the acquired voice;
An information processing method characterized in that a computer executes.

Images