JP6097791B2 - Topic continuation desire determination device, method, and program - Google Patents

Description

  The present invention relates to a topic continuation desire determination apparatus for determining whether or not a user wants to continue a topic of a system utterance immediately before the user utterance after a user utterance in an interactive system that performs a dialogue with the user, The present invention relates to a method and a program.

  In general, utterances in dialogue have a focal point that represents the object that is the subject of the content. The system utterance and user utterance columns in FIG. 5 indicate the utterance strings of the system and the user that are engaged in the conversation. In each row in order from the first row, the utterance in the order of system utterance / user utterance is shown. Show. The focus of each row represents the focus extracted by the system from the utterance sequence up to the user utterance in the previous row. Focus extraction is performed using a rule-based or machine learning technique. The system selects or generates a sentence referring to the extracted focus from the utterance knowledge data, and utters the sentence as the next utterance. The system utterance of the line is the next utterance.

  The technique of the following nonpatent literature 1 has taken the way as described above.

Ryuichiro Higashinaka, “Efforts for Chat Dialog System,” Artificial Intelligence Laboratories, Vol.SIG-SLUD-70, pp.65-70, Mar. 2014.

  There are cases where the topic of system utterances uttered based on the focus as described above may or may not be continued for the user. Looking at the user utterances in the first to third lines in FIG. 5, the user wants to continue the topic of the immediately preceding system utterance. However, when viewing the user utterances in the 4th to 5th lines, the user does not want to continue the topic of the immediately preceding system utterance. Looking at the user utterance on the sixth line, the user is interested in the topic of the last system utterance and wants to continue.

  If the user wants to continue the topic of the last system utterance, it can be said that the system should continue the same topic or change the topic as appropriate. On the other hand, if the user does not want to continue the topic of the last system utterance, the system must change the topic without fail. Thus, in order for the system to properly speak, information on whether or not the user wants to continue the topic of the immediately preceding system utterance is required.

  However, in the current interactive technology, there is no mechanism for determining whether or not the user wants to continue the topic of the immediately preceding system utterance. For this reason, although the user does not want to continue the topic of the immediately preceding system utterance, the same topic may be continued, and the user's satisfaction with the dialogue is low. To solve this problem, even if you take a procedure to change the topic regularly (for example, every second or third system utterance), even if the user wants to continue the same topic as the previous system utterance, It will not be resolved that the user will be less satisfied with the dialogue. Thus, since an appropriate speech strategy cannot be taken, there exists a subject that the user's satisfaction with respect to a dialogue is low.

  This invention is for solving the said subject, and provides the topic continuation desire determination apparatus, method, and program which can determine whether the user wants to continue the topic of the last system utterance. It is aimed.

  In order to solve the above-described problem, a topic continuation desire determination device according to a first invention includes a sequence of utterances in a dialog between a system and a user, and the user utterances given in advance to each user utterance in the sequence After that, for each flag in each dialog information, the flag is set for each flag in the dialog information including a flag string indicating whether or not the user has a desire to continue the topic of the previous system utterance. The feature value for a predetermined feature is extracted from the utterance sequence up to the utterance corresponding to, and a feature vector that is a set of the feature and the feature value is generated, and a set of each flag and the corresponding feature vector Learning means for generating a classification model for determining a flag corresponding to an arbitrary feature vector from the set, and an utterance string as input, and a feature value for each feature used by the learning means from the utterance string Extracted and generates a feature vector is a set of a set of feature and its feature value, by referring to the classification model, characterized in that it comprises a determining means for determining flag corresponding to the plain of the vector.

  A topic continuation desire determination method according to the second invention is a topic continuation desire determination method in a topic continuation desire determination apparatus including a learning means and a determination means, wherein the learning means is in a dialogue between the system and a user. A sequence of utterances and a column of flags that are given in advance to each user utterance in the sequence and indicate whether or not the user has a desire to continue the topic of the previous system utterance after the user utterance For each flag in each dialogue information, a feature value for a predetermined feature is extracted from each utterance sequence up to the utterance corresponding to the flag, and a set of the feature and its feature value is input. Generate a feature vector that is a set of, and generate a classification model to determine the flag corresponding to an arbitrary feature vector from the set of pairs of feature vectors corresponding to each flag The step and the determination means take the utterance string as input, extract a feature value for each feature used in the learning means from the utterance string, and generate a feature vector that is a set of the feature and the feature value; Referring to a classification model and determining a flag corresponding to the feature vector.

  The feature value used in the learning unit and the determination unit is the frequency of the word or word n-gram being spoken, the frequency of the semantic category of the word being spoken, or the concept vector corresponding to the speech. Each element value, presence / absence of sentence ending during utterance, frequency of positive or negative expression during utterance, number of characters during utterance, number of characters of utterance, or dialogue corresponding to utterance It is characterized by the presence / absence of an action, the similarity between utterances, the number of consecutive appearances of the same topic, or any other feature value extracted from the utterance.

  Moreover, the program of this invention is a program for functioning a computer as each means which comprises said topic continuation desire determination apparatus.

  According to the present invention, the flag corresponding to the utterance string whose tendency is similar to A in the learning data that is input to the learning means becomes the determination result for the utterance string A that is input to the determination means. Can be obtained.

  According to the present invention, it is possible to determine whether or not the user wants to continue the topic of the immediately preceding system utterance, and the system returns an utterance suitable as the next utterance, thereby facilitating smooth interaction between the system and the user. Play.

It is a block diagram which shows the functional structure of the topic continuation desire determination apparatus which concerns on embodiment of this invention. It is a figure which shows an example of the flag corresponding to a feature vector. It is a flowchart figure which shows the learning process routine of the topic continuation desire determination apparatus which concerns on embodiment of this invention. It is a flowchart figure which shows the determination processing routine of the topic continuation desire determination apparatus which concerns on embodiment of this invention. It is a figure which shows an example of the collection of dialogue information.

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

<Configuration of topic continuation desire determination apparatus according to an embodiment of the present invention>

  The configuration of the topic continuation desire determination apparatus according to the first embodiment of the present invention will be described. FIG. 1 is a configuration example of a topic continuation desire determination apparatus according to the present invention. As shown in FIG. 1, the topic continuation desire determination apparatus 100 according to the embodiment of the present invention stores a CPU, a RAM, a program for executing a learning processing routine and a determination processing routine, which will be described later, and various data. It can be composed of a computer including a ROM. Functionally, the topic continuation desire determination apparatus 100 includes an input unit 10, a calculation unit 20, and an output unit 30 as shown in FIG.

  The input means 10 includes a sequence of utterances in the dialogue between the system and the user, and after the user utterance given in advance to each user utterance in the sequence, the user continues the topic of the previous system utterance. A set of dialogue information including a flag string indicating whether or not the applicant has a desire is accepted as an input. The system utterance, user utterance, and flag columns in FIG. 5 are examples of such dialogue information. The flag is set to 1 when the user has a desire to continue the topic of the immediately preceding system utterance, and is set to 0 when the user does not have it. This flag is given by a person looking at the contents of the utterance string. The flag may be given only to some user utterances in the utterance string. Further, the input means 10 accepts an utterance sequence whose last is a user utterance.

  The computing unit 20 includes a learning unit 22, a classification model storage unit 24, and a determination unit 26.

  The learning unit 22 extracts a feature value for a predetermined feature from each utterance string up to the utterance corresponding to the flag for each flag in each piece of dialogue information received by the input unit 10. Then, a feature vector that is a set of features and feature values is generated.

  Hereinafter, examples of features and types of feature values will be described. As for the feature, the same feature character string is treated as different if the type is different.

  For example, a word being uttered as a feature can be taken as a feature value. In addition, as a target utterance, a user utterance corresponding to the flag, a system utterance immediately before the utterance, or the like can be taken, and the same feature in different utterances can be different. The same applies to other features described in the following description.

  When the user utterance on the first line in FIG. 5 is targeted, (Ramen: Feature value) can be taken as (Ramen: 1), (Gas: 1), (Like: 1), (I: 1). The number of features is the total number of different words that appear in the user utterance in the data that is input to the learning means 22, and the word “think” that appears in the user utterance on the third row is one feature, but one row For the user's utterance of eyes, (feature: feature value) for this feature “I think” is (I think: 0). In the following description, a feature value of 0 (feature: feature value) is not specified unless otherwise specified. Moreover, only content words can be taken as features, such as (Ramen: 1) and (Like: 1).

  The word n-gram being uttered as a feature can be taken as a feature value. When n = 2, from the user utterance on the first line in FIG. 5, as (feature: feature value), (ramen / ga: 1 :), (ga / like: 1), (like / like: 1) Can be taken.

  The semantic category of the word being uttered as a feature can be taken as the feature value. Here, the semantic category means a group of similar words as one category. For example, when the meaning category of “ramen” is “noodles” and the meaning category of “like” is “+ evaluation”, from the user utterance on the first line in FIG. 5 (feature: feature value), (Noodles: 1), (+ Evaluation: 1) can be taken.

  Each element of the concept vector corresponding to the utterance can be taken as a feature value as a feature value. Here, the concept vector is obtained by, for example, synthesizing the word concept vector of the word in the target utterance using the word concept vector generated by the method of Non-Patent Document 2.

[Non-patent literature 2] Katsuto Bessho, Toshiro Uchiyama, Kei Uchiyama, Ryoji Kataoka, Masahiro Oku, “Corpus concept-based generation method based on co-occurrence between words and semantic attributes,” Information Processing Society of Japan, Dec. 2008, Vol.49, No.12, pp.3997-4006.

  The end-of-sentence expression being uttered as a feature and the presence / absence (if present: 1; absent: 0) can be taken as a feature value. Here, the sentence end expression is a character string obtained by concatenating the attached words at the end of the sentence. As (feature: feature value), (I: 1) can be obtained from the user utterance on the first line in FIG. 5, and (Dayo: 1) can be obtained from the user utterance on the second line.

  A character string of a positive expression or a negative expression being uttered as a feature can be used as a feature value. From the user utterance on the first line in FIG. 5 above, since “like” is a character string of positive expression, (like: feature value) can be taken as (like: 1). Also, the features may be “positive expression” and “negative expression”, and the feature value may be the sum of the frequencies of utterance character strings belonging to each expression. In this case, (positive expression: 1) can be taken as (feature: feature value) from the user utterance on the first line in FIG.

  Moreover, the number of characters in speech can be taken as the feature value. From the user utterance on the first line in FIG. 5, (character number: 9) can be taken as (feature: feature value).

  It is possible to take the character type being uttered as a feature and the number of characters of that character type as a feature value. From the user utterance in the first line in FIG. 5, (Hiragana: 4), (Katakana: 4), and (Kanji: 1) can be taken as (Feature: Feature Value).

  Among the types of features described so far, words, word n-grams, word semantic categories, positive / negative expressions (and their character strings), and the type of character being uttered, for each type of feature, As the feature value, the frequency corresponding to the feature may be a ratio of the total frequency corresponding to all the features. For example, for the character type being uttered, (Hiragana: 4/9), (Katakana: 4/9), (Kanji: 1) as (feature: feature value) from the user utterance on the first line in FIG. / 9).

  As a feature, a dialogue action corresponding to an utterance can be taken as a feature value (1 if present, 0 if not present). Here, the dialogue action of the utterance is the utterance intention of the speaker who made the utterance, and a predetermined number of conversation actions are determined in advance. As the target utterance, a certain number of utterances up to the user utterance corresponding to the flag are taken. For example, the target utterances are two utterances up to the user utterance corresponding to the flag (that is, the user utterance and the system utterance immediately before the utterance). For each utterance, a corresponding dialogue action is derived using a method such as machine learning. As (feature: feature value), (question_evaluation: 1) can be obtained from the system utterance on the first line in FIG. 5, and (self-disclosure_evaluation +: 1) can be obtained from the user utterance on the first line.

  As the feature value, the similarity between two utterances in the utterance string up to the user utterance corresponding to the flag (for example, the user utterance and the immediately preceding system utterance) can be taken. The similarity between utterances can be a cosine measure between vectors corresponding to each utterance. Here, as a vector corresponding to an utterance, a word frequency vector, a word n-gram frequency vector, a word meaning category frequency vector, or a concept, which is a set of the features and feature values described above. Can take vectors etc. Alternatively, the similarity between utterances may be set to 1 when there is an intersection between the focus sets extracted from each utterance, and may be set to 0 when there is no intersection.

  As the feature value, in the system utterance or user utterance up to the user utterance corresponding to the flag, the number of times that the same topic appears continuously until the latest utterance can be taken. As a topic, the focus extracted from the utterance can be taken. For system utterances, the focus (value in the focus column in FIG. 5) that is the basis for selecting the system utterances can also be taken. In this case, for the flag in the sixth line in FIG. 5, the number of times that the same focus (“ramen”) appears continuously until the most recent system utterance is five. As another method of certifying the same topic, if the similarity between the topic vector such as the concept vector of each utterance and the topic vector of the latest utterance is equal to or greater than a certain threshold, the topic of the utterance group is the same A way to do it is also possible.

  In addition, an arbitrary feature extracted from the utterance string up to the user utterance corresponding to the flag and its feature value can be used for subsequent processing.

  The learning means 22 generates, for each flag, a feature vector that is a set of several types of features and their feature values among the types of features described above. FIG. 2 shows an example of a set of sets of flag and corresponding feature vector.

  Next, the learning means 22 regards each flag as a classification category, and uses a classification algorithm such as a support vector machine to set a flag corresponding to an arbitrary feature vector from a set of pairs of feature vectors corresponding to each flag. A classification model for determination is generated. Then, the learning unit 22 stores the generated classification model in the classification model storage unit 24.

  The classification model storage unit 24 stores the classification model generated by the learning unit 22.

  The determination unit 26 receives an utterance sequence that is accepted by the input unit 10 and ends with a user utterance, extracts a feature value for each feature used by the learning unit 22 from the utterance sequence, and extracts the feature and its feature value. Generate a feature vector that is a set of sets of Then, the flag corresponding to the feature vector is determined with reference to the classification model stored in the classification model storage unit 24. If the determination result flag is 1, it is estimated that the user wants to continue the topic of the immediately preceding system utterance, and if the determination result flag is 0, the user estimates that the user does not want to continue the topic of the immediately preceding system utterance. To do.

  The output unit 30 outputs the determination result obtained by the determination unit 26.

<Operation of topic continuation desire determination device>
FIG. 3 is an example of a learning process flow of the topic continuation desire determination apparatus 100. The content of the learning process of the topic continuation desire determination apparatus 100 will be described with reference to FIG.

  First, in step S100, the input means 10 receives a set of dialogue information.

  In step S102, the learning means 22 obtains a feature value for a predetermined feature from the utterance string up to the utterance corresponding to the flag for each flag in the dialogue information of the dialogue information set received in step S100. Extraction is performed to generate a feature vector which is a set of features and feature values.

  In step S104, the learning unit 22 regards each flag as a classification category, and uses a classification algorithm such as a support vector machine from a set of pairs of feature flags corresponding to the flags generated in step S102. A classification model for determining a flag corresponding to an arbitrary feature vector is generated.

  In step S106, the learning unit 22 stores the classification model generated in step S104 in the classification model storage unit 24, and ends the learning processing routine.

  FIG. 4 is an example of a determination process flow of the topic continuation desire determination apparatus 100. The processing content of the determination process of the topic continuation desire determination apparatus 100 will be described with reference to FIG.

  In step S <b> 200, the input unit 10 receives an utterance sequence whose last is a user utterance.

  In step S202, the determination unit 26 extracts a feature value for each feature used by the learning unit 22 from the utterance sequence received in step S200, and generates a feature vector that is a set of the feature and the feature value. .

  In step S204, the determination unit 26 refers to the classification model stored in the classification model storage unit 24, and determines a flag corresponding to the feature vector generated in step S202.

  In step S206, the output unit 30 outputs the determination result obtained in step S204 and ends the determination processing routine.

  It is possible to construct the processing described so far as a program, install the program from a communication line or a recording medium, and implement it by means such as a CPU.

  The present invention is not limited to the above-described embodiments, and various modifications and applications are possible within the scope of the claims.

  The present invention can be applied to a dialogue processing technique that realizes a smooth interaction between a system and a user.

DESCRIPTION OF SYMBOLS 10 Input means 20 Calculation means 22 Learning means 24 Classification model memory | storage part 26 Determination means 30 Output means 100 Topic continuation desire determination apparatus

Claims (5)

The sequence of utterances in the dialogue between the system and the user, and whether the user has a desire to continue the topic of the immediately preceding system utterance after the user utterances given in advance to each user utterance in the sequence A set of dialogue information including a flag sequence indicating whether or not a flag is input, and for each flag in each dialogue information, a feature value for a predetermined feature is extracted from the utterance sequence up to the utterance corresponding to the flag. Then, a feature vector that is a set of features and feature values is generated, and a classification model for determining a flag corresponding to an arbitrary feature vector is generated from the set of feature vectors corresponding to each flag. Learning means to
Using an utterance string as an input, extracting a feature value for each feature used by the learning means from the utterance string, generating a feature vector that is a set of the feature and the feature value, and referring to the classification model, A topic continuation desire determination device comprising: determination means for determining a flag corresponding to the feature vector. The feature value used in the learning means and the determination means is the frequency of the word or word n-gram being uttered, the frequency of the semantic category of the word being uttered, or each element value of the concept vector corresponding to the utterance, Or, the presence or absence of sentence ending expressions during utterance, the frequency of positive or negative expressions during utterance, the number of characters being uttered, the number of characters in the utterance, or the presence or absence of interactive actions corresponding to the utterance, The topic continuation desire determination apparatus according to claim 1, wherein the topic continuation desire determination apparatus is a similarity between utterances, a continuous appearance count of the same topic, or any other feature value extracted from the utterance. A topic continuation desire determination method in a topic continuation desire determination apparatus including a learning means and a determination means,
The learning means includes a sequence of utterances in the dialogue between the system and the user, and after the user utterance given in advance to each user utterance in the sequence, the user continues the topic of the immediately preceding system utterance. A set of dialogue information including whether or not the applicant has a desire is input, and for each flag in each dialogue information, an utterance sequence up to the utterance corresponding to the flag is used for a predetermined feature. To extract a feature value, generate a feature vector that is a set of a feature and a set of feature values, and determine a flag corresponding to an arbitrary feature vector from a set of feature vectors corresponding to each flag Generating a classification model of
The determination unit receives an utterance string, extracts a feature value for each feature used by the learning unit from the utterance string, generates a feature vector that is a set of the feature and the feature value, and the classification model And a step of determining a flag corresponding to the feature vector. The feature value used in the learning means and the determination means is the frequency of the word or word n-gram being uttered, the frequency of the semantic category of the word being uttered, or each element value of the concept vector corresponding to the utterance, Or, the presence or absence of sentence ending expressions during utterance, the frequency of positive or negative expressions during utterance, the number of characters being uttered, the number of characters in the utterance, or the presence or absence of interactive actions corresponding to the utterance, The topic continuation desire determination method according to claim 3, wherein the method is a similarity between utterances, a continuous appearance count of the same topic, or another arbitrary feature value extracted from the utterance.   The program for functioning a computer as each means which comprises the topic continuation desire determination apparatus of Claim 1 or Claim 2.

Images