JP7407560B2 - Keyword evaluation device, keyword evaluation method, and keyword evaluation program - Google Patents

Description

The present invention relates to a device, method, and program for evaluating keywords.

2. Description of the Related Art Conventionally, technologies related to communication robots that watch videos such as television programs together with people have been proposed.
For example, in Patent Document 1, the robot uses social media comments related to the video, generates utterances according to the robot's internal personality or emotional state, and operates the robot. A technology has been proposed that allows you to perform actions that make you feel as if you are watching a movie.
In addition, in Patent Document 2, the robot responds to commands from humans such as changing channels, and also murmurs spontaneously while facing the direction of the TV, so that the robot behaves as if it were watching TV autonomously. A control technology for robots has been proposed.

Patent No. 6122792 JP 2018-180472 Publication Patent No. 5194198 Patent No. 6486165

Makoto Yamamoto, Hiroaki Tanimoto, Naoko Nitta, Noboru Babaguchi; "Estimating the interest range of a specific person when watching TV to acquire personal preferences", IEICE Transactions D, Vol. J90-D, No. 8, pp. 2202-2211, 2007

By the way, in order for a communication robot that watches TV with people to behave in ways that empathize with people regarding topics related to the program being watched, what kind of things should the people watching with the robot do? It is important to know what your interests are. One type of such information is interest level, which is an index of how much a person is interested in a keyword such as a proper noun.
For example, if it becomes clear that the person watching TV with the robot is a fan of A-san, that is, has a high level of interest in her, he or she will make empathetic utterances such as, "A-san is cute, isn't she?" The robot will be able to utter sentences and informational utterances such as ``A-ko is on TV right now.''

Patent Document 3 proposes a technique for recommending programs that are likely to be of interest to a person based on past viewing history. This is a technology that presents content that the user prefers, and in particular, estimates the user's preferences, including relationships with other people.
However, in the techniques used in such program recommendation, the estimation target is the program, and cannot be used to estimate keywords of interest related to the program being viewed.

Patent Document 4 proposes a technique for extracting interesting candidate keywords from a large number of keywords related to a program. This is a technique that uses a plurality of dictionaries prepared in advance to extract keywords that are presumed to be of general interest from among many keywords related to a program.
However, the extracted keywords are only keywords that are estimated to be of great interest according to multiple dictionaries, and do not reflect the interests of individuals viewing the program.

Non-Patent Document 1 proposes a technique for estimating an interesting section of a television program that is being viewed. This technology uses a camera to take pictures of people's facial expressions while they are watching TV, and then estimates areas of interest based on those facial expressions. Thereby, the keyword tagged in the interesting section within the program can be estimated as the keyword of interest.
However, if there are multiple tagged keywords within the estimated interest interval, it is not possible to estimate which of these keywords the user is interested in. Furthermore, in order to photograph the facial expressions of people watching television programs, it is necessary to install a camera in the home, which is difficult in practice. Even if a camera could be installed, it would be difficult to accurately extract a person's facial expressions because the positions and orientations of people's faces vary in daily life, such as when viewing while lying down.

In this way, if a program is tagged with many related keywords, some kind of reaction from the user while watching the program is required as information in order to estimate the keywords that the viewer is interested in while watching the program. becomes. However, it is often difficult in practice to install sensor devices such as cameras in homes or to observe viewers' reactions by attaching sensors to the body. Additionally, information that can be obtained through viewing history or remote control operations can be used to estimate preferences for each program, but is insufficient to estimate interest in related keywords within a program. was there.

An object of the present invention is to provide a keyword evaluation device, a keyword evaluation method, and a keyword evaluation program that can evaluate a user's degree of interest in each keyword.

The keyword evaluation device according to the present invention includes an utterance generation unit that generates an utterance including an input keyword and outputs it to a user, and a reaction acquisition unit that acquires the type of reaction of the user to the utterance. , an interest degree calculation unit that calculates the user's degree of interest in the keyword based on the reaction type.

The reaction acquisition unit may acquire any one of a plurality of types including positive and negative as the reaction type.

The reaction acquisition unit may measure a reaction time of the user to the utterance, and if the reaction time exceeds a predetermined time, the reaction acquisition unit may acquire a non-reaction type as the reaction type.

The keyword evaluation device may include an utterance type selection unit that selects the type of the uttered sentence from a predetermined number of utterance types with a predetermined probability.

The keyword evaluation device includes a preference calculation unit that calculates a preference for an utterance containing the keyword for each utterance type based on the reaction type, and the interest calculation unit calculates the preference statistics for the utterance including the keyword. The degree of interest may be calculated based on information.

The reaction acquisition unit may measure a reaction time of the user to the uttered sentence, and the preference calculation unit may weight the preference level based on the reaction time.

The keyword evaluation device may include a keyword extraction unit that extracts the keywords included in a predetermined database from the broadcast program.

The keyword evaluation method according to the present invention includes an utterance generation step of generating an utterance including an input keyword and outputting it to a user, and a reaction acquisition step of acquiring the reaction type of the user to the utterance. A computer executes the step of calculating the degree of interest of the user in the keyword based on the reaction type.

A keyword evaluation program according to the present invention is for causing a computer to function as the keyword evaluation device.

According to the present invention, it is possible to evaluate a user's degree of interest in each keyword.

It is a block diagram showing the functional composition of the keyword evaluation device in an embodiment. FIG. 3 is a diagram showing a detailed functional configuration of an utterance generation unit in an embodiment. It is a figure showing an example of the disclosure template in an embodiment. It is a figure showing an example of a question template in an embodiment. It is a figure which shows an example of the confirmation template in embodiment. It is a figure showing the example of registration of the keyword dictionary in an embodiment. It is a figure which shows an example of the preference degree recorded in the keyword dictionary in embodiment.

An example of an embodiment of the present invention will be described below.
The keyword evaluation method in this embodiment is a method of evaluating a user's degree of interest in a keyword. In particular, the present embodiment exemplifies a method in which a robot that views a broadcast program such as a television program together with the viewer evaluates the viewer's degree of interest in a keyword related to the program being viewed.
In this embodiment, the index value of interest in a keyword is referred to as the degree of interest, and the larger the value, the higher the interest.

In this embodiment, a keyword evaluation device that evaluates the interest level of a keyword is built into a robot that watches a program together with a person. Evaluate the degree of interest in a keyword by talking to people and obtaining their reactions.

FIG. 1 is a block diagram showing the functional configuration of a keyword evaluation device 1 in this embodiment.
The keyword evaluation device 1 is an information processing device equipped with various interfaces in addition to a control section and a storage section, and the control section executes software (keyword evaluation program) stored in the storage section. Various functions are realized.

The control unit of the keyword evaluation device 1 includes a keyword extraction unit 11 , an utterance type selection unit 12 , an utterance generation unit 13 , a reaction acquisition unit 14 , a preference calculation unit 15 , and an interest calculation unit 16 .
Further, the storage unit of the keyword evaluation device 1 includes various databases such as a keyword dictionary 17 in addition to the keyword evaluation program.

In this embodiment, the keyword evaluation device 1 performs voice input/output as an example of an interface with a user. The utterance output O outputted from the keyword evaluation device 1 is presented to the user as audio output OA by the speech synthesis device 2. The voice input IA, which is the user's response to the voice output OA, is converted into a text sentence by the voice recognition device 3 and input as a response sentence input I to the keyword evaluation device 1.

In this embodiment, a voice method will be described as an interface with the user, but the method is not limited to this. For example, the utterance output O may be displayed on a display, and the user may use a keyboard, etc. It is also possible to input the response sentence input I using

The keyword extraction unit 11 extracts keywords existing in the keyword dictionary 17 from the input sentence T input to the keyword evaluation device 1. In this embodiment, the input sentence T will be explained as subtitle information (closed caption) included in the broadcast program being viewed. For example, if the subtitle sentence "Today I would like to go and eat ramen at this restaurant" is input as input sentence T, and the keyword dictionary 17 contains the keyword "ramen", then The keyword extraction unit 11 extracts the keyword "ramen".

Note that the keyword dictionary 17 is a database in which keywords to be evaluated are registered. Keywords are mainly proper nouns and common nouns such as people's names, place names, and facilities. In this embodiment, it is assumed that keywords are registered in the keyword dictionary 17 in advance. Furthermore, for each keyword registered in the keyword dictionary 17, a degree of preference indicating the degree of preference is stored for each type of uttered sentence, which will be described later.

Here, the method of extracting subtitle information from a broadcast program is not limited, but for example, a robot may obtain subtitle information in conjunction with a television receiver, or a service may be distributed using communication equipment such as the Internet as a broadcast-linked service. Methods such as obtaining subtitle information from
Note that in this embodiment, the input sentence T is subtitle information, but it is not limited to this. For example, it may be a sentence extracted by image processing of television images or voice recognition of television audio using a camera or microphone equipped on the robot.

Further, the frequency of extracting keywords from subtitle information or the like may be adjusted as appropriate, such as by randomly sampling with a certain period of time in between. Alternatively, the keyword extraction unit 11 prioritizes keywords with high importance based on the frequency of word appearance, etc., or prioritizes keywords that have been uttered few times and have not received sufficient evaluation, and the number of extracted keywords. You can also narrow down the

The utterance type selection section 12 selects the type of the utterance sentence generated by the utterance generation section 13 from among a predetermined number of utterance types at a predetermined probability.
Here, the utterance type is a classification of utterances in conversations between people while watching TV. For example, in the following document A, utterance types are "question,""instruction,""information,""disclosure,""reflection," and "confirmation." Eight types of "interpretation" and "response" are defined.
Document A: Yuta Hoshi, Yutaka Kaneko, Yuta Hagio, Yasuhiro Murasaki, Michihiro Uehara: "Analysis of dialogue between people while watching TV for robot speech", Institute of Electronics, Information and Communication Engineers, IEICE Technical Report, CNR2019-1 ( 2019-06), pp. 1-6

In this embodiment, there are four types of utterances: "question", "information", "disclosure", and "confirmation", and the utterance type selection unit 12 selects one from these four types with a predetermined probability. .
A "question" is an utterance that asks a question to the other party, such as "Do you like A-san?"
"Information" is an utterance that provides some kind of information to the other party, such as "A-ko-san used to appear in XX drama."
"Disclosure" is an utterance that conveys one's thoughts and feelings to the other party, such as "I love A-san."
"Confirmation" is an utterance that confirms something to the other party, such as "A-san is beautiful."

According to Document A, in conversations between people while watching TV, the most common type of utterance at the beginning of a conversation is "disclosure" at 20-40%, followed by "confirmation" at 14-20%, and "question" at 14-20%. 8% to 16%, and 4% to 9% for "information." Therefore, the utterance type selection unit 12 randomly selects "disclosure" at a ratio of 45%, "confirmation" at 30%, "question" at 15%, and "information" at 10%, for example.

The utterance generation unit 13 generates an utterance based on the keyword extracted by the keyword extraction unit 11 and the utterance type selected by the utterance type selection unit 12, and outputs it as an utterance output O.
Specifically, the utterance generation unit 13 generates an utterance sentence that corresponds to the utterance type selected by the utterance type selection unit 12 and includes the keyword extracted by the keyword extraction unit 11. For this reason, the utterance generation unit 13 may include generation means for generating utterances corresponding to each of the four types of utterances.

For example, in the following document B, subtitles containing emotion words are stored as template sentences using subtitles from past broadcast programs, and by combining them with keywords, utterances expressing emotions containing this keyword are created. A technique has been proposed to automatically generate . In this technology, verb phrases expressing desires such as ``I want to eat,'' ``I want to talk,'' and ``I want to go,'' and adjectives such as ``pretty,''``interesting,'' and ``big'' are called representative words. Representative words with similar feature vectors to the keyword are selected. When a representative word corresponding to the target keyword is selected, an uttered sentence is generated using a template sentence including this representative word. In this embodiment, this technique is applied to generate utterances for each of the four utterance types described above: "disclosure,""question," and "confirmation."
Document B: Unexamined Japanese Patent Publication No. 2018-190077

FIG. 2 is a diagram showing a detailed functional configuration of the utterance generation unit 13 in this embodiment.
The utterance generation section 13 includes a template extraction section 131, a vector distance calculation section 132, a template selection section 133, a keyword search section 134, and an information sentence generation section 135.
Further, the utterance generation unit 13 refers to each database of a template database 18A, a feature vector database 18B, and a keyword dictionary 18C stored in the storage unit.

The keyword Key extracted by the keyword extraction unit 11 and the utterance type C selected by the utterance type selection unit 12 are input to the utterance generation unit 13 .
Here, when the utterance type C is "question", "disclosure", or "confirmation", the keyword Key and the utterance type C are sent to the template extraction unit 131, and when the utterance type C is "information", the keyword Key and the utterance type C is input into the keyword search section 134, respectively.

The template extraction unit 131 extracts a disclosure template when the utterance type C is "disclosure", a question template when it is a "question", and a confirmation template when it is "confirmation" from the template database 18A.
The XX template is data in which a template for generating a "XX" sentence is stored, and includes a representative word and a template sentence corresponding to the representative word.

FIG. 3 is a diagram showing an example of a disclosure template in this embodiment.
The disclosure template is data in which a representative word and a template sentence for generating a "disclosure" sentence for the representative word are registered as a pair.
In this example, six representative words are registered: "I want to talk,""I want to go,""I want to eat,""pretty,""interesting," and "big" and template sentences corresponding to each representative word are registered. . For example, for the representative verb phrase "I want to talk", two template sentences "I want to talk with %key" and "I want to chat with %key" are registered. Furthermore, two template sentences are registered for the representative adjective word "beautiful": "%key is so beautiful" and "%key is so beautiful".
Note that %key in the template sentence indicates the insertion position of the keyword.

FIG. 4 is a diagram showing an example of a question template in this embodiment.
The question template also has the same structure as the disclosure template, and a template sentence for generating a "question" sentence is registered together with a representative word.
In this example, template sentences are registered for representative words of three verb phrases: "I want to talk,""I want to go," and "I want to eat." For example, for the representative verb phrase "I want to talk", two template sentences are registered: "Have you talked to %key?" and "Do you want to chat with %key?".

FIG. 5 is a diagram showing an example of a confirmation template in this embodiment.
The data structure of the confirmation template is the same as that of the disclosure template and the question template, and a template sentence for generating a "confirmation" sentence is registered along with a representative word.
In this example, template sentences are registered for three representative words of the adjective: "beautiful,""interesting," and "delicious." For example, for the representative adjective word ``beautiful'', two template sentences are registered: ``%key is very beautiful, isn't it?'' and ``pretty %key is wonderful, isn't it?''.

The vector distance calculation unit 132 calculates the vector distance between the feature vector of each representative word in the template and the feature vector of the keyword (Key) for the template group extracted by the template extraction unit 131.
In the feature vector database 18B, feature vectors are recorded in advance for at least the keywords recorded in the keyword dictionary 17 and the representative words included in each template.

As a method for calculating a feature vector, there is a method such as word2vec, for example. word2vec uses a three-layer neural network, and it is empirically known that the feature vectors calculated by inputting a large amount of separated sentences are vectors with similar words close in distance. .
Note that the feature vector calculation method is not limited to word2vec, and any calculation method may be used as long as the feature vectors of words that are semantically similar are distributed nearby on the vector space.

Here, when the feature vector of keyword Key is vec_k and the feature vector of representative word t is vec_t, the distance between the vectors is, for example, cosine similarity cos(vec_k, vec_t)=(vec_k・vec_t)/(|vec_k ||vec_t|) can be used.

The template selection unit 133 selects the one closest to the feature vector of the keyword Key, that is, the one with the cosine similarity, from among the results of calculation of the cosine similarity between the feature vector of the keyword Key and the feature vector of each representative word by the vector distance calculation unit 132. Select representative words with feature vectors with large values.

Next, the template selection unit 133 selects one template from among the template sentences for the selected representative word, and replaces %key in the template sentence with the keyword Key, so that the uttered sentence (“question” sentence, "disclosure" statement or "confirmation" statement).
For example, when "Sanuki udon" is input as the keyword Key and "confirmation" as the utterance type C to the utterance generation unit 13, the template extraction unit 131 extracts the confirmation template shown in FIG. 5. Subsequently, the vector distance calculation unit 132 calculates the vector distances between the feature vector vec_Sanuki udon of "Sanuki udon" and the feature vectors vec_beautiful, vec_interesting, and vec_delicious of the representative words. Then, based on these calculation results, if vec_delicious has the closest vector distance, the template selection unit 133 randomly selects one template sentence for "delicious". As a result, when "%key is delicious, isn't it?" is selected as the template sentence, the template selection unit 133 generates the uttered sentence "Sanuki udon is delicious, isn't it?" as the "confirmation" sentence.

The keyword search unit 134 obtains the explanatory text of the input keyword Key from the keyword dictionary 18C.
Explanatory sentences such as the meaning of keywords are registered in the keyword dictionary 18C.

FIG. 6 is a diagram showing an example of registration of the keyword dictionary 18C in this embodiment.
In this example, the keywords "ammonite", "Sanuki udon", and "Sirius" are registered, and the explanation for the keyword "ammonite" is "creature with a shell like a conch shell that lived in the sea 350 million years ago." , "Kagawa Prefecture's specialty udon" is registered as an explanation for the keyword "Sanuki udon", and "the brightest star in the constellation Canis Major, and the brightest star visible from the earth except the sun" is registered as an explanation for "Sirius". There is.

The information sentence generation unit 135 uses the explanatory text and the keyword Key acquired by the keyword search unit 134 to generate utterances such as “What is (Key)? (Explanatory text)” and “What is (Key) (Explanatory text)?” Generates "informational" sentences such as "Yo".
For example, when "Sirius" is input as a keyword and "information" is input as an utterance type to the utterance generation unit 13, the keyword search unit 134 inputs the explanatory text ``The brightest star in the constellation Canis Major, which is the brightest star on Earth except for the sun. The information sentence generation unit 135 outputs ``Sirius is the brightest star in the constellation Canis Major, and is the brightest star visible from the earth except for the sun.''

The reaction acquisition unit 14 acquires a reaction type Cr indicating the user's preference from the response sentence input I, which is the user's reaction to the utterance output O outputted by the utterance generation unit 13. In addition, a reaction time Tr, which is the time from when the uttered sentence output O is output until the response sentence input I is input, is measured.

The reaction type Cr may be, for example, three types: "yes" indicating a positive preference, "no" indicating a negative preference, and "no reaction" indicating a neutral indifference.
When the reaction time Tr cannot be measured beyond a predetermined time Th (Tr=0), that is, when there is no response sentence input I within the time Th, the reaction acquisition unit 14 sets the reaction type Cr to "None". "reaction". Furthermore, when the reaction time Tr is within the time Th, that is, when there is a response sentence input I within the time Th, the reaction acquisition unit 14 selects either “yes” or “no” for the response sentence input I. Performs two-class classification based on whether it is a sentence or not.

As a method for classifying sentences into two classes, for example, there is a method of creating a classifier using training data in advance, and as a learning method, there are methods using a neural network, a method using a support vector machine, etc. . Note that the classification method is not limited to these.

The preference calculation unit 15 calculates the degree of preference for the keyword Key as numerical data from the reaction type Cr and reaction time Tr acquired by the reaction acquisition unit 14. The calculation result is recorded in the keyword dictionary 17 for each keyword Key and utterance type C as the degree of preference for the utterance sentence of utterance type C that includes the keyword Key.

Here, the preference level for an utterance of utterance type C that includes the keyword Key is assumed to be S_C_Key. Although the method of calculating the preference level is not limited, the preference level calculation unit 15 weights it based on the reaction time Tr, and calculates the S_C_Key using the following formula.
S_C_Key={α_C×n(Cr)-β_C×(1-n(Cr))}/Tr
(When Tr≠0)
=0 (when Tr=0, that is, Cr=“no reaction”)
Note that n(Cr) is a binary function that is 1 when Cr is "yes" and 0 when Cr is "no," and α_C and β_C are constants determined in advance.

α_C and β_C can be set for each utterance type C. For example, an utterance whose utterance type is "Question" is an utterance that directly asks the user about the keyword, and a "yes" or "no" answer from the user is an utterance of another utterance type. This data is more reliable than the reaction type estimated from the response text. Therefore, by setting the values of α_question and β_question to be larger than other utterance types, it is possible to have a greater influence on the interest level than on the preference level of other utterance types.

In addition, in this embodiment, n(Cr) is explained as a binary function of 0 or 1, but it is not limited to this, and n(Cr) may be an integer or a real number. For example, if a value such as reliability can be obtained when classifying reaction types by the reaction obtaining unit 14, this value can be used. Further, when sensor data such as a heart rate monitor is used as input to the reaction acquisition unit 14, this value can also be used.

FIG. 7 is a diagram showing an example of preference degrees recorded in the keyword dictionary 17 in this embodiment.
In this example, "ammonite,""Sanukiudon," and "Sirius" are recorded as keywords. Furthermore, for each keyword, the total of preference degrees calculated so far for each utterance type is recorded.
Every time a reaction to an utterance is obtained, the preference calculation unit 15 adds the calculated preference to the preference recorded in the corresponding keyword and utterance type columns of the keyword dictionary 17, and also calculates the number of utterances. Add +1.

For example, the utterance of "question" regarding the keyword "ammonite" has been uttered once so far, and the preference level is +1.2, and the utterance of "disclosure" has been uttered once so far, and the preference level is +0.2. be. Similarly, the utterance of ``disclosure'' related to the keyword ``Sanuki udon'' is made three times, resulting in a total preference value of +0.3, and the utterance of ``confirmation'' is made once, with a preference value of -0.2. .

Furthermore, if the reaction type to the utterance is "no reaction", the added preference level is 0, but the number of utterances is counted, so the average preference level is neutral, neither positive (like) nor negative (dislike). approaching the state of

When a keyword is specified, the interest level calculation unit 16 calculates the user's interest in the keyword based on statistical information of the preference level calculated based on past reaction types and reaction times for utterances that include this keyword. The degree K is calculated and output.

Specifically, the interest level calculation unit 16 calculates, for example, the average value of the preference levels of the relevant keyword in the keyword dictionary 17 as the interest level. In the example of FIG. 7, the interest level of the keyword "ammonite" is (1.2/1+0.2/1)/2=0.7, and the interest level of the keyword "Sanuki udon" is (0.3/3-0. 2/1)/2=-0.1, and the interest level of the keyword "Sirius" can be calculated as (0.4/2-0.2/1)/2=0.

According to the present embodiment, the keyword evaluation device 1 uses the keyword extraction unit 11 to extract the keywords stored in the keyword dictionary 17 in order to estimate what kind of interest the user has in the keywords. The utterance generation unit 13 generates an utterance output O for the keyword. In the keyword evaluation device 1, the reaction acquisition unit 14 determines the reaction type from the response sentence input I, which is the user's response to the uttered sentence output O, and based on this reaction type, the interest level calculation unit 16 determines the reaction type for the keyword. Calculate interest level.

For example, when the keyword evaluation device 1 is incorporated into a robot, the robot can speak a sentence containing a keyword to a person, and the degree of interest in the keyword can be estimated from the reaction to the sentence. While conventional methods collect passive information from people watching TV to estimate their interest level, this embodiment uses an active method in which the robot speaks to the user to determine the user's interest level for each keyword. Users' interest level for each keyword can be evaluated based on the differences in their reactions.
As a result, the robot can realize natural communication, such as choosing what the user likes or dislikes and speaking, and changing its behavior depending on the user's level of interest in the keyword.

Further, the keyword evaluation device 1 can be incorporated into a robot, for example, to extract keywords included in the keyword dictionary 17 from subtitles etc. while watching a broadcast program on TV or the like together with the user, and Make the robot speak.
Thereby, the keyword evaluation device 1 can evaluate the user's degree of interest in each keyword related to the broadcast program being viewed based on the user's reaction to the robot's utterance.

As a result, when a keyword is input, the robot can perform different actions depending on the user's level of interest in the keyword. In other words, the robot can determine whether or not to utter a certain keyword, select the content of the action, and present information such as related programs of interest to the user in a timely manner.
Note that the content uttered regarding a keyword with a high degree of interest and the content uttered for evaluating the degree of interest may be the same. In this case, the various templates described above are shared. In either case, the preference level and interest level may be updated according to the user's response.

By acquiring any one of a plurality of reaction types including positive and negative reaction types, the keyword evaluation device 1 can evaluate the user's interest level in a keyword by quantifying it using a certain scale.

The keyword evaluation device 1 measures the user's reaction type to the uttered sentence, and when the reaction time exceeds a predetermined time, acquires a non-reaction type as the reaction type.
Thereby, the keyword evaluation device 1 can determine the type of preference in which the user is indifferent to the keyword, and can appropriately evaluate the degree of interest.

In the keyword evaluation device 1, the utterance type selection unit 12 selects the type of the uttered sentence from among a predetermined number of utterance types with a predetermined probability.
For example, it is possible to evaluate the level of interest by using only "questions" as the type of utterance, but if the user only uses "questions", even if the user initially responds to the questions from the robot, he or she will eventually respond due to reasons such as being annoying. Therefore, variations in utterance types are necessary. Document A shows that conversations between people often begin with a "disclosure" sentence, and that "questions,""information," and "confirmation" are uttered more often than "disclosure" sentences. However, the results show that the response rate of the other party is high.
Therefore, by selecting the type of speech from a robot to a person with a probability close to that of a conversation between people, the keyword evaluation device 1 can prevent the user from getting bored by, for example, using a robot that only asks questions or a robot that only reads out information. It is possible to have a robot speak at the same rate as a human, without causing confusion, and to achieve natural dialogue.

The keyword evaluation device 1 uses the preference calculation unit 15 to calculate the preference for the utterance including the keyword for each utterance type based on the reaction type, and updates the keyword dictionary 17.
For example, when the utterance type "question" is selected for the keyword "A-san," the keyword evaluation device 1 asks the user, "Do you like A-san?" Based on responses such as "I like it,""Idon't really like it," and "I don't like it," the degree of preference for the "Question" sentence regarding "Miss A" is estimated.
As a result, the keyword evaluation device 1 calculates the interest level based on the statistical information of the preference level estimated for each utterance type. The reliability of a user's response to a sentence can be reflected in the level of interest.

The keyword evaluation device 1 can appropriately evaluate the degree of interest by weighting the degree of preference based on the reaction time to the uttered sentence, thereby reflecting the degree of the user's likes and dislikes for the keyword in the degree of preference.

Although the embodiments of the present invention have been described above, the present invention is not limited to the embodiments described above. Further, the effects described in the above-described embodiments are only a list of the most preferable effects resulting from the present invention, and the effects according to the present invention are not limited to those described in the embodiments.

In the embodiment described above, a template of a representative word selected for a keyword is used when generating an utterance, but the present invention is not limited to this as long as the method generates an utterance related to a keyword. Although vector distances based on feature vectors are used to select representative words, for example, representative words may be determined in advance for each keyword and selected from among these representative words. Furthermore, instead of using representative words, an individual utterance may be prepared in advance for each keyword registered in the keyword dictionary 17, and this may be used directly.

Furthermore, although a method has been described in which the keyword dictionary 18C is used when generating the "information" sentence, the method is not limited to this. For example, a method of connecting to the Internet and referring to an online encyclopedia such as Wikipedia to obtain explanations of keywords is also applicable.

In the above-described embodiment, when determining the user's reaction type, the response sentence input I obtained by converting the voice input IA into text data by the voice recognition device 3 is used, but the present invention is not limited to this. Any method may be used as long as it is possible to determine "no", that is, to determine whether the user is positive or negative regarding the utterance output O. For example, an image of the user's face captured by a camera may be input, or a device for acquiring physical data such as an electrocardiograph, heart rate monitor, or electroencephalogram may be attached to the user.
Furthermore, when a plurality of people are using the service at the same time, the keyword evaluation device 1 may acquire the reaction type based on, for example, a reaction such as whether or not the conversation between the users was lively, depending on the utterance.

Although the above embodiment has been described assuming that there is one user, the present invention is not limited to this. By identifying each of a plurality of users and managing the degree of preference in the keyword dictionary 17 for each user, it is possible to obtain the degree of interest for each user.

In the above-described embodiment, the keyword evaluation device 1 was described as being built into the robot; however, the present invention is not limited to this, and the keyword evaluation device 1 is placed outside the robot and is connected to the robot for communication by wire or wirelessly or via a network. may be done.
Further, although the keyword evaluation device 1 is configured to include various databases, the present invention is not limited to this, and the databases may be placed in an external server such as a cloud.

In the present embodiment, the configuration and operation of the keyword evaluation device 1 have been mainly described, but the present invention is not limited to this, and may be configured as a method or program for evaluating keywords, including each component. .

Furthermore, the functions of the keyword evaluation device 1 may be realized by recording a program on a computer-readable recording medium, causing the computer system to read and execute the program recorded on the recording medium. good.

The "computer system" here includes hardware such as an OS and peripheral devices. Furthermore, the term "computer-readable recording medium" refers to portable media such as flexible disks, magneto-optical disks, ROMs, and CD-ROMs, and storage devices such as hard disks built into computer systems.

Furthermore, a ``computer-readable recording medium'' refers to a computer-readable storage medium that dynamically retains a program for a short period of time, such as a communication line when transmitting a program via a network such as the Internet or a communication line such as a telephone line. It may also include a device that stores a program at a fixed time, such as a volatile memory inside a computer system that is a server or client in that case. Further, the above program may be for realizing a part of the above-mentioned functions, or may be one that can realize the above-mentioned functions in combination with a program already recorded in the computer system. .

1 Keyword evaluation device 2 Speech synthesis device 3 Speech recognition device 11 Keyword extraction unit 12 Utterance type selection unit 13 Utterance generation unit 14 Reaction acquisition unit 15 Preference level calculation unit 16 Interest level calculation unit 17 Keyword dictionary 18A Template database 18B Feature vector database 18C Keyword Dictionary 131 Template Extraction Unit 132 Vector Distance Calculation Unit 133 Template Selection Unit 134 Keyword Search Unit 135 Information Sentence Generation Unit

Claims (9)

an utterance generation unit that takes as input a keyword extracted from at least one of subtitles, audio, and images included in the broadcast program being viewed , generates an utterance including the keyword, and outputs the generated utterance to the user;
a reaction acquisition unit that acquires the user's reaction type to the utterance;
A keyword evaluation device comprising: an interest level calculation unit that calculates the user's interest level for the keyword based on the reaction type. The keyword evaluation device according to claim 1, wherein the reaction acquisition unit acquires one of a plurality of types including positive and negative as the reaction type. The keyword evaluation device according to claim 2, wherein the reaction acquisition unit measures reaction time of the user with respect to the utterance, and when the reaction time exceeds a predetermined time, acquires a non-reaction type as the reaction type. . The keyword evaluation device according to any one of claims 1 to 3, further comprising an utterance type selection unit that selects the type of the uttered sentence from a predetermined number of utterance types with a predetermined probability. comprising a preference calculation unit that calculates a preference for an utterance containing the keyword for each utterance type based on the reaction type;
The keyword evaluation device according to claim 4, wherein the interest level calculation unit calculates the interest level based on statistical information of the preference level. The reaction acquisition unit measures reaction time of the user to the utterance,
The keyword evaluation device according to claim 5, wherein the preference calculation unit weights the preference based on the reaction time. The keyword evaluation device according to any one of claims 1 to 6, further comprising a keyword extraction unit that extracts the keywords included in a predetermined database from a broadcast program. an utterance generation step of inputting a keyword extracted from at least one of subtitles, audio, and images included in the broadcast program being viewed , generating an utterance including the keyword, and outputting the utterance to the user;
a reaction acquisition step of acquiring the user's reaction type to the utterance;
A keyword evaluation method in which a computer executes the step of calculating the degree of interest of the user in the keyword based on the reaction type. A keyword evaluation program for causing a computer to function as the keyword evaluation device according to any one of claims 1 to 7.

Images