JP6146864B2 - Program, server, and method for promoting search interactively with user - Google Patents

Description

  The present invention relates to a technique for a program that allows a search to proceed interactively with a user.

  A dialogue system that realizes natural dialogue with human beings is becoming more popular, especially using smartphones and tablet terminals. Among them, there is a technique for searching for content that a user may need by alternately proceeding with a user according to a dialog scenario.

  Conventionally, there is a technology of an interactive game device that selects the next question sentence based on the psychological state of the respondent and prevents the question pattern from being fixed (see, for example, Patent Document 1). According to this technique, a database is provided in which a large number of question sentences are recorded together with selection criteria for the question sentences. Then, the psychological state of the respondent is analyzed from the voice of the respondent to the question sentence, the psychological state is compared with the selection criteria of the question sentence, and the question sentence is selected from the database. The question text is provided as guidance to the questioner. That is, according to this technique, the order of the question sentences is not fixed and is changed according to the psychological state of the user.

  Further, there is a technique for searching for content that is intended for the user by advancing the dialogue even when the user does not have a clear query (question sentence) (see, for example, Non-Patent Document 1). According to this technique, for example, for the purpose of searching for a recipe, the system presents a question regarding the recipe to the user, and makes the user answer “Yes / No” to narrow down the search candidates. A large number of questions are stored in the database according to the classification of the recipe data. Then, the system selects a question sentence having the largest information gain from the search candidate recipe set. Thereby, it is possible to speed up the narrowing down of the question sentence until the search result to be targeted (reducing the number of questions).

  FIG. 1 is an explanatory diagram showing processing of a dialog search program in the prior art.

According to FIG. 1, it is assumed that the recipe content is searched by interaction with the user.
(S1) It is assumed that the system asks the user “What material do you use?”
(S2) Assume that the user responds that “pork and cabbage are good”. At this time, the user may have desired “healthy” diet food.
(S3) The system searches the content set S of the recipe from the database using “pork” and “cabbage” included in the answer sentence as search keys. From the content set S, an information gain is calculated for each content set Sa related to the category a.
Content collection Sa about Chinese cuisine: Information gain = 0.60
Content set Sa related to party: Information gain = 0.56
Content set Sa regarding microwave oven: Information gain = 0.53
Healthy food content collection Sa: Information gain = 0.50
(S4) “Content set related to Chinese food” with information gain = 0.60 is selected. Therefore, the system will ask the respondent “Chinese food?”.

  According to this technique, the question sentence with the highest “information gain” is selected. The information gain refers to a feature goodness criterion used for feature selection in decision tree algorithm ID3 (Iterative Dichotomiser 3), for example. ID3 is a kind of supervised learning algorithm designed for general purpose.

JP 2002-177647 A

XU Xin, KATO Tsuneo, KDDIRD Lab., `` Interactive Recipe Search Interface using Spoken Dialogue Agent for Tablet Devices '', IEICE Technical Report, 113, 73 (HIP2013 1-31), pp.191-193 ID3, [online], [September 7, 2013 search], Internet <URL: http: //en.wikipedia.org/wiki/ID3>

  However, according to FIG. 1, although the respondent actually wanted a “healthy” diet food, he / she would be asked whether the food is Chinese food. Here, isn't it a problem that the respondent himself does not consider what kind of recipe contents he / she is interested in, and simply determines the next question based on information gain alone? The inventors of the present application thought.

  Accordingly, the present invention provides a program for selecting a question sentence based on an information gain, thereby narrowing down a question sentence as far as possible until reaching a search result to be a target when a search is made interactive with a user. It is an object to provide a server and a method.

According to the present invention, there is provided an interactive search program for causing a computer to function to advance a search interactively with a user,
Content accumulating means for accumulating a large number of contents to be searched and adding one or more categories to each content;
A dialog progress interface means for clearly indicating a question sentence based on a category to the user and obtaining the answer sentence;
Content search means for searching the content set S using the content storage means using the words included in the answer sentence as search keys;
Information gain calculating means for calculating an information gain IG (a) for each category a with respect to the content set S;
Content selection means for allowing the user to arbitrarily select a plurality of contents in the content set S;
Selection probability calculating means for calculating a selection probability P (a) for each category a for a plurality of contents selected by the user;
For each category a, the computer functions as an update information gain calculating means for calculating an update information gain PIG (a) obtained by weighting the selection probability P (a) to the information gain IG (a),
The dialogue progress interface means is characterized in that the computer functions so as to clearly indicate to the user a question sentence based on a category having the highest update information gain.

According to another embodiment of the interactive search program of the present invention,
The update information gain calculation means calculates the update information gain by the following formula: PIG (a) = IG (a) × P (a) + Δr (Δr = 0 to 0.001 random number)
It is also preferable to make the computer function.

According to another embodiment of the interactive search program of the present invention,
The update information gain calculation means calculates the update information gain by the following formula: B = b ^{−P (a)} (b is ^a natural number greater than 1)
PIG (a) = IG (a) ^B
It is also preferable to make the computer function.

According to another embodiment of the interactive search program of the present invention,
As for the information gain calculation means, the information gain IG (a) is based on Iterative Dichotomiser 3 and is calculated by the following formula: IG (a) = H (S) _−Σt∈T p (t) H (t)
H (S) = − _Σx∈T p (x) log ₂ p (x)
H (S): Average information amount of the content set S p (x): It is also preferable to cause the computer to function so that all the content sets S become contents belonging to the category x.

According to another embodiment of the interactive search program of the present invention,
The selection probability P (a) is calculated based on the maximum posterior probability (MAP (Maximum a posteriori) estimation) with the user's preference probability g (a) for the category a calculated from the user's usage history as the prior probability. It is also preferable to make the computer function like this.

According to another embodiment of the interactive search program of the present invention,
The content storage means preferably causes the computer to function so as to use an online learning method (confidence-weighted linear classification) using a machine learning library in order to classify a large number of contents into categories.

According to the present invention, there is provided a dialog search server that interactively advances a scenario with a terminal based on a user operation,
Content accumulating means for accumulating a large number of contents to be searched and adding one or more categories to each content;
A dialogue progress interface means for sending a question sentence based on a category to the user and receiving the answer sentence;
Content search means for searching the content set S using the content storage means using the words included in the answer sentence as search keys;
Information gain calculating means for calculating an information gain IG (a) for each category a with respect to the content set S;
Content selection means for allowing the user to arbitrarily select a plurality of contents in the content set S;
Selection probability calculating means for calculating a selection probability P (a) for each category a for a plurality of contents selected by the user;
Update information gain calculating means for calculating an update information gain obtained by weighting the selection probability P (a) to the information gain IG (a) for each category a;
The dialog progress interface means transmits a question sentence based on the category having the highest update information gain to the terminal.

According to the present invention, there is provided an interactive method for interactively progressing a scenario with a user using an apparatus,
The apparatus has a content storage unit that stores a large number of contents to be searched and assigns one or more categories to each content for storage.
A first step of clearly indicating a question sentence based on the category to the user and obtaining the answer sentence;
A second step of searching the content set S using the content storage unit using a word included in the answer sentence as a search key;
A third step of calculating an information gain IG (A) for each category for the specified content set;
A fourth step for allowing the user to arbitrarily select a plurality of contents in the content set;
A fifth step of calculating a selection probability P (a) for each category for a plurality of contents selected by the user;
A sixth step of calculating an update information gain obtained by weighting the selection probability to the information gain for each category;
And a seventh step of clearly indicating a question sentence based on the category having the highest update information gain to the user.

  According to the program, the server, and the method of the present invention, when selecting a question sentence based on information gain and making a search interactively proceed with a user, narrowing down a question sentence until reaching a search result to be targeted Can be as soon as possible.

It is explanatory drawing showing the process of the dialog search program in a prior art. It is a function block diagram of the dialog search program in this invention. It is explanatory drawing showing the specific process of the dialogue search program in this invention. It is the screen of the user interface displayed on the terminal.

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

  FIG. 2 is a functional configuration diagram of the dialogue search program according to the present invention.

According to FIG. 2, the interactive search program 1 that advances the search interactively with the user is installed and executed on the server side that communicates with the terminal. The user uses a terminal such as a smart phone or tablet, accessed through the network to the server. Of course, the interactive search program 1 may be installed only in a single device such as the terminal itself.

  A terminal 2 as a user interface for the interactive search program 1 has an input device and an output device for the user. When the user utters an answer, the voice signal acquired by the microphone is converted into text by voice recognition, and the text is input to the dialogue search program 1. The question sentence output from the dialogue search program 1 is converted into a voice signal by voice synthesis, and the voice signal is uttered from a speaker. In addition, the terminal 2 is equipped with a touch panel display, and can display a plurality of contents to the user and select the contents.

  According to FIG. 2, the dialogue search program 1 includes a content accumulation unit 10, a dialogue progress interface unit 11, a content search unit 12, an information gain calculation unit 13, a content selection unit 14, and a selection probability calculation unit 15. And an update information gain calculation unit 16. The interactive search program 1 is executed by a computer mounted on the device, thereby functioning as a server and a single device. The processing flow of these functional components can also be understood as a dialog search method.

  FIG. 3 is an explanatory diagram showing specific processing of the dialog search program in the present invention.

  According to FIG. 3, the question sentence selection method in S3 (S31 to S35) is different from that in FIG. S1 and S2 in FIG. 3 are the same as those in FIG. The process of S3 is executed by the content search unit 12, the information gain calculation unit 13, the content selection unit 14, the selection probability calculation unit 15, and the update information gain calculation unit 16.

  FIG. 4 is a user interface screen displayed on the terminal.

  According to FIG. 4, the touch panel display of the terminal 2 is displayed. “Character Agent” is displayed on the display, and a voice dialogue between the user and the agent may be advanced. Of course, the present invention is not limited to a voice dialogue, but may be a text dialogue.

  As in FIG. 1, the user agent displayed on the display asks the user “What material do you use?” (S 1). It is assumed that the user utters “Terminal pork and cabbage are good” (S2) to the terminal 2.

[Content accumulation unit 10]
The content accumulating unit 10 accumulates a large number of contents to be searched and assigns and stores one or more “categories” to each content. For example, assume that the search target content is a recipe. In this case, a category is assigned to each content as follows.
<Content name><Category>
“Healthy Peppers stuffed with meat”-> (Healthy) (meat)
"Luxury / Chinese cuisine"-> (Chinese) (Party)
"Cabbage salad"-> (salad) (Italian)
"Men's cooking with pork 3 minutes"-> (meat) (microwave)
"Low-calorie side dish * snail and konjac rice"-> (Healthy) (Japanese)
"No sugar! Basically Chikuzenni with microwave oven"-> (Healthy) (Microwave) (Japanese)
"Noodles with cabbage"-> (Noodles) (Japanese)

  The content storage unit 10 preferably uses an online learning method (confidence-weighted linear classification) using a machine learning library in order to classify a large number of contents into categories. A bag-of-words model is used for vector features, and TF-IDF values of nouns and adjectives are used for feature amounts. According to the above-described example, it is possible to classify which category each recipe belongs to.

[Dialogue progress interface unit 11]
The dialogue progress interface unit 11 clearly indicates a question sentence based on the category to the user, and acquires the answer sentence. The dialog progress interface unit 11 uses a dynamic or static dialog scenario to select a question sentence in accordance with the user's answer sentence, and advances the dialog. Here, the dialogue progress interface unit 11 advances the dialogue scenario based on the control logic, but clearly indicates a question sentence based on the category having the highest “information gain” to the user. As the information gain, the update information gain for each category output from the update information gain calculation unit 16 is used. That is, the dialog search program of the present invention is characterized in that the existing dialog progress interface unit 11 uses the update information gain output from the update information gain calculation unit 16.

[Content Search Unit 12]
The content search unit 12 uses the content storage unit 10 to search the content set S using a word included in the answer sentence as a search key. The content set S may be searched for feature quantities of Confidence-weighted linear classification (CW) vector features using a bag-of-words model. As the feature quantity, TF-IDF values of nouns and adjectives are used.

(S31) Similar to S3 in FIG. 1, the content set S of the recipe is searched from the content storage unit 10 using “pork” and “cabbage” included in the answer sentence as search keys. From the content set S, the content set Sa related to the category a can be derived. Among a large number of content sets Sa, according to FIG. 3, for example, the following four content sets Sa are represented.
Chinese food content collection Sa
Content set Sa related to parties
Content set Sa related to microwave oven
Healthy food content collection Sa

[Information Gain Calculation Unit 13]
The information gain calculation unit 13 calculates an information gain IG (a) for each category a for the content set S. Regarding the information gain calculation unit 13, the information gain IG (a) is based on ID3 and is calculated by the following equation.
IG (a) = H (S) _−Σt∈T p (t) H (t)
H (S) = − _Σx∈T p (x) log ₂ p (x)
H (S): Average information amount of content set S p (x): Ratio of all content sets S becoming content belonging to category x

According to S31 of FIG. 3, it is assumed that the information gain IG (a) is calculated for each content set Sa as follows, similarly to S3 of FIG. As shown below, the top four sets of contents Sa are shown in descending order of information gain IG (a).
Content collection Sa about Chinese cuisine: Information gain = 0.60
Content set Sa related to party: Information gain = 0.56
Content set Sa regarding microwave oven: Information gain = 0.53
Healthy food content collection Sa: Information gain = 0.50

[Content selection unit 14]
The content selection unit 14 allows the user to arbitrarily select a plurality of contents in the content set Sa. According to the present invention, the user is allowed to select content once without using the information gain for each content set Sa obtained by the information gain calculation unit 13 as it is.

(S32) According to FIGS. 3 and 4, some of the content names included in the content set S selected in S31 are displayed in a list form on the terminal 2 as follows. Here, it is preferable that an average is selected from all content sets without biasing to a specific content set S.
"Healthy bell pepper stuffing"
"Luxury / Chinese cuisine"
"Cabbage salad"
"3 minutes cooking of a man using pork"
"Low-calorie side dish * and konjac rice"
“No sugar! Boiled in Chikuzen with a microwave oven”
"Noodle dishes with cabbage"

Here, according to FIG. 4, the user selects a plurality of desired contents. For example, assume that the following three contents are selected.
"Healthy bell pepper stuffing"
"Low-calorie side dish * and konjac rice"
“No sugar! Boiled in Chikuzen with a microwave oven”

[Selection probability calculation unit 15]
The selection probability calculation unit 15 calculates a selection probability P (a) for each category a for a plurality of contents selected by the user.

(S33) “Category” is extracted from each selected content. According to FIG. 4, the following categories are extracted.
“Healthy Peppers stuffed with meat”-> (Healthy) (meat)
"Low-calorie side dish * snail and konjac rice"-> (Healthy) (Japanese)
"No sugar! Basically Chikuzenni with microwave oven"-> (Healthy) (Microwave) (Japanese)

(S34) Next, the selection probability P (a) for each category a is calculated. The selection probability P (a) is the number of appearances divided by the total number of selected contents for each category a.
p (Chinese) = 0/3 = 0
p (party) = 0/3 = 0
p (microwave oven) = 1/3 = 0.333
p (healthy) = 3/3 = 1

[Update Information Gain Calculation Unit 16]
For each category a, the update information gain calculation unit 16 calculates an update information gain PIG (a) obtained by weighting the information gain IG (a) with the selection probability P (a). As the weighting, for example, there are the following two calculation methods.
(Update information gain calculation method 1)
PIG (a) = IG (a) × P (a) + Δr (Δr = 0 to 0.001 random number)
(Update information gain calculation method 2)
B = b- ^{P (a)} (b is ^a natural number greater than 1)
PIG (a) = IG (a) ^B

(S35) According to FIG. 3, for example, the update information gain is calculated as follows by the update information gain calculation method 1 (Δr = 0.001) described above.
"Is it Chinese food?": Information gain 0.60 x Weight 0 + 0.001 = 0.001
"Is it a party food?": Information gain 0.56 x Weight 0 + 0.001 = 0.001
"Do you use a microwave oven?": Information gain 0.53 x Weight 0.333 + 0.001 = 0.177
"Is it a healthy dish?": Information gain 0.50 x Weight 1 + 0.001 = 0.5
As a result, the question sentence with the highest update information gain for each category is "Healthy dish?"

The selection probability P (a) is a maximum probability of the posterior probability (the user's preference probability g (a) for the category a calculated from the user's usage history (retrieved or stored recipe content). Calculation is based on MAP (Maximum a posteriori) estimation. According to the MAP estimation, if a conditional distribution model P (X | a) (X: training sample set) and a parameter prior distribution g (a) are given, the posterior distribution of the parameter is obtained by Bayes' theorem. Is calculated as follows.
MAP = arg max P (a) = arg max P (X / a) · g (a)
X: Number of contents of each category selected by the user P (X / a): Calculation method similar to the selection probability P (a) described above g (a): Number of searched category a / Retrieved recipe Times

  As described above in detail, according to the program, server, and method of the present invention, a search to be a target when a search is advanced interactively with a user by selecting a question sentence based on information gain. It is possible to narrow down the question sentences up to the result as quickly as possible.

  Various changes, modifications, and omissions of the above-described various embodiments of the present invention can be easily made by those skilled in the art. The above description is merely an example, and is not intended to be restrictive. The invention is limited only as defined in the following claims and the equivalents thereto.

DESCRIPTION OF SYMBOLS 1 Dialog search program 10 Content storage part 11 Dialog progress interface part 12 Content search part 13 Information gain calculation part 14 Content selection part 15 Selection probability calculation part 16 Update information gain calculation part 2 Terminal

Claims (8)

An interactive search program that allows a computer to function interactively with a user.
Content accumulating means for accumulating a large number of contents to be searched and adding one or more categories to each content;
A dialogue progress interface means for clearly indicating a question sentence based on the category to the user and acquiring the answer sentence;
Using the content storage means, a content search means for searching the content set S using a word contained in the answer sentence as a search key;
Information gain calculating means for calculating an information gain IG (a) for each category a for the content set S;
Content selection means for allowing a user to arbitrarily select a plurality of contents in the content set S;
Selection probability calculating means for calculating a selection probability P (a) for each category a for a plurality of contents selected by a user;
For each category a, let the computer function as update information gain calculation means for calculating an update information gain PIG (a) obtained by weighting the selection probability P (a) to the information gain IG (a),
The dialog search program characterized in that the dialog progress interface means causes a computer to function so as to clearly show a question sentence based on a category having the highest update information gain to a user. The update information gain calculating means calculates the update information gain by the following formula: PIG (a) = IG (a) × P (a) + Δr (Δr = 0 to 0.001 random number)
The interactive search program according to claim 1, wherein the computer functions as described above. The update information gain calculating means calculates the update information gain by the following equation: B = b ^{−P (a)} (b is ^a natural number greater than 1)
PIG (a) = IG (a) ^B
The interactive search program according to claim 1, wherein the computer functions as described above. Regarding the information gain calculating means, the information gain IG (a) is based on Iterative Dichotomiser 3 and is calculated by the following formula: IG (a) = H (S) _−Σt∈T p (t ) H (t)
H (S) = − _Σx∈T p (x) log ₂ p (x)
H (S): Average information amount of the content set S p (x): The computer is caused to function so that all the content sets S become contents belonging to the category x. The interactive search program according to item 1. The selection probability P (a) is based on the maximum posterior probability estimation (MAP (Maximum a posteriori) estimation) with the user's preference probability g (a) for the category a calculated from the user's usage history as the prior probability. The interactive search program according to any one of claims 1 to 4, wherein the computer is operated so as to calculate. The content storage means causes a computer to function so as to use an online learning method (confidence-weighted linear classification) using a machine learning library in order to classify a large number of contents into categories. 6. The interactive search program according to any one of 5 above. A dialogue search server that interactively advances a scenario with a terminal based on a user operation,
Content accumulating means for accumulating a large number of contents to be searched and adding one or more categories to each content;
Dialogue progress interface means for sending a question sentence based on the category to the user and receiving the answer sentence;
Using the content storage means, a content search means for searching the content set S using a word contained in the answer sentence as a search key;
Information gain calculating means for calculating an information gain IG (a) for each category a for the content set S;
Content selection means for allowing a user to arbitrarily select a plurality of contents in the content set S;
Selection probability calculating means for calculating a selection probability P (a) for each category a for a plurality of contents selected by a user;
Update information gain calculation means for calculating an update information gain obtained by weighting the selection probability P (a) to the information gain IG (a) for each category a,
The dialog search server, wherein the dialog progress interface means transmits a question sentence based on a category having the highest update information gain to the terminal. A dialogue method in which a scenario is progressed interactively with a user using a device,
The apparatus has a content storage unit that stores a large number of contents to be searched, and stores one or more categories for each content.
A first step of clearly indicating a question sentence based on the category to the user and obtaining the answer sentence;
Using the content storage unit, a second step of searching the content set S using a word included in the answer sentence as a search key;
A third step of calculating an information gain IG (A) for each category with respect to the specified content set;
A fourth step for allowing the user to arbitrarily select a plurality of contents in the content set;
A fifth step of calculating a selection probability P (a) for each category for a plurality of contents selected by a user;
A sixth step of calculating an update information gain obtained by weighting the selection probability to the information gain for each category;
And a seventh step of clearly indicating a question sentence based on the category having the highest update information gain to the user.

Images