JP6220762B2 - Next utterance candidate scoring device, method, and program - Google Patents

Description

  The present invention is suitable as a next utterance to be uttered by the system side in a next utterance candidate set generated by the system after a certain utterance sequence is exchanged with the user in an interactive system that performs a dialogue with the user. The present invention relates to a next utterance candidate scoring apparatus, method, and program for determining a next utterance candidate.

  In the non-task-oriented dialog system, the content of the dialog is chat, and in Non-Patent Document 1, the system is prepared by creating a database of texts such as the web and Twitter (registered trademark) and selecting those similar to user utterances. Generate an utterance.

  There is a method of obtaining a focus (topic) in an utterance sequence exchanged so far, taking a plurality of sentences referring to the focus as a next utterance candidate from a database, and calculating a score of the next utterance candidate. After the score is calculated, the system returns the next utterance candidate with the highest score to the user. Alternatively, the system returns to the user the next utterance candidate randomly selected from the next utterance candidate having a score equal to or higher than a certain threshold or the next utterance candidate having the highest score.

  As such a method, a sequence of N (≧ 1) utterances immediately before is used as a context, and for each next utterance candidate, the closeness between the concept vector of the context utterance sequence and the concept vector of the next utterance candidate is expressed. There is a method based on a concept base for calculating a score.

Bessho, F., Harada, T., and Kuniyoshi, Y. “Dialog System Using Real-Time Crowdsourcing and Twitter Large-Scale Corpus.” In Proc. SIGDIAL, pp. 227-231, 2012.

  In the method based on the concept base described above, if the context utterance sequence and the content of the next utterance candidate are close to each other, the score is high. For this reason, the next utterance candidate that is close to the content but is not suitable as the next utterance is selected, and the content is a little far away. However, there is a problem that the next utterance candidate suitable as the next utterance may not be selected. For example, if the context utterance string is one question sentence and a certain next utterance candidate is the same question sentence, the next utterance candidate is selected, but it cannot be said that the next utterance candidate is suitable as the next utterance. Then, a more appropriate next utterance candidate that answers the question is not selected.

  In order to solve this problem, an object of the present invention is to provide a next utterance candidate scoring apparatus, method, and program capable of selecting a next utterance candidate suitable as the next utterance in the context utterance sequence from the next utterance candidate set. It is to provide.

In order to solve the above-described problem, the next utterance candidate scoring device according to the present invention includes a concept base that is a set of a pair of a word and a concept vector representing the meaning of the word, an utterance string A as a context, and a next utterance. A set of utterance string evaluation data D composed of a combination of the candidate B and a label C indicating whether the next utterance candidate B is appropriate as the next utterance of the utterance string A is input, and for each utterance string evaluation data D, Referencing the concept base, generating a concept vector E of the utterance string A, generating a concept vector F of the next utterance candidate B, and combining the concept vector E and the concept vector F with the concept vector G and An utterance string evaluation data converting means for generating the converted utterance string evaluation data H composed of a combination with the label C, and an arbitrary concept vector having the same dimension as the concept vector G from the set of the converted utterance string evaluation data H. Le, are configured to include a learning means for generating a classification model for calculating a score is classified into one of the values of the label C.
The next utterance candidate scoring apparatus receives the context utterance string I and the next utterance candidate set as input, generates a concept vector J of the utterance string I by referring to the concept base, and generates the next utterance candidate set. For each next utterance candidate K in FIG. 4, a concept vector L of the next utterance candidate K is generated, a concept vector M obtained by combining the concept vector J and the concept vector L is generated, and the classification model generated by the learning means , It is possible to further include an evaluation means for calculating a score for classifying the concept vector M into one value of the label C.

Further, the next utterance candidate scoring method according to the present invention is a next utterance including a concept base that is a set of a pair of a word and a concept vector representing the meaning of the word, an utterance string evaluation data conversion unit, and a learning unit. The next utterance candidate scoring method in the candidate scoring device, wherein the utterance string evaluation data conversion means includes the utterance string A as a context, the next utterance candidate B, and the next utterance candidate as the next utterance of the utterance string A. A set of utterance string evaluation data D consisting of a combination with a label C indicating whether B is appropriate or not is used as an input, the concept base is referenced for each utterance string evaluation data D, and a concept vector E of the utterance string A is obtained. And generating a concept vector F of the next utterance candidate B, and evaluating a converted utterance string composed of a combination of the concept vector G and the label C obtained by combining the concept vector E and the concept vector F A score for classifying an arbitrary concept vector having the same dimension as the concept vector G into one value of the label C from the set of the converted utterance string evaluation data H. Generate a classification model for calculating.
In the next utterance candidate scoring method, the evaluation means receives the utterance sequence I as a context and the next utterance candidate set as input, generates a concept vector J of the utterance sequence I by referring to the concept base, For each next utterance candidate K in the next utterance candidate set, a concept vector L of the next utterance candidate K is generated, a concept vector M combining the concept vector J and the concept vector L is generated, and generated by the learning means By referring to the classified model, it is possible to further include calculating a score at which the concept vector M is classified into one value of the label C.

  Moreover, the program of this invention is a program for functioning a computer as each means which comprises said next utterance candidate scoring apparatus.

  In the present invention, whether the next utterance candidate is suitable as the next utterance in the context utterance string is solved as a classification problem. Claim 1 constitutes the process of the learning phase, and claim 2 constitutes the process of the classification phase.

  According to the next utterance candidate scoring device, method, and program according to the present invention, the combination of the utterance sequence A as the context, the next utterance candidate B, and the label C indicating whether the utterance sequence A is suitable as the next utterance A set of utterance string evaluation data D consisting of the following is input, and for each utterance string evaluation data D, a concept vector E of the utterance string A is generated, a concept vector F of the next utterance candidate B is generated, and the concept vector E and the concept A post-conversion utterance string evaluation data H composed of a combination of a concept vector G combined with the vector F and a label C is generated, and an arbitrary concept vector having the same dimension as the concept vector G is generated from the set of post-conversion utterance string evaluation data H. Learns a classification model for calculating a score that is classified into one value of label C, and the system returns an utterance suitable for the next utterance. There is an effect that emissions will be smooth.

It is a block diagram which shows the functional structure of the next utterance candidate scoring apparatus which concerns on embodiment of this invention. It is a figure which shows the example of the collection of utterance string evaluation data D. FIG. It is a figure which shows the example of a concept base. It is a figure which shows the example of the set of the utterance string evaluation data H after conversion. It is a flowchart figure which shows the processing routine in the utterance string evaluation data conversion means and learning means of the next utterance candidate scoring apparatus which concerns on embodiment of this invention. It is a block diagram which shows the functional structure of the next utterance candidate scoring apparatus which concerns on embodiment of this invention. It is a flowchart figure which shows the processing routine in the evaluation means of the next utterance candidate scoring apparatus which concerns on embodiment of this invention. It is a figure which shows the set of the group of the concept vector M and the score calculated with respect to it.

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

<Configuration of next utterance candidate scoring device>

  Next, the configuration of the next utterance candidate scoring device according to the embodiment of the present invention will be described. FIG. 1 is a configuration example of a next utterance candidate scoring device according to claim 1 of the present invention. As shown in FIG. 1, the next utterance candidate scoring device 100 according to the embodiment of the present invention includes a CPU, a RAM, and a ROM that stores programs and various data for executing processing routines to be described later. It can be configured with a computer including. Functionally, the next utterance candidate scoring device 100 includes an input unit 10 and a calculation unit 20 as shown in FIG.

  The input means 10 is a set of utterance string evaluation data D comprising a combination of a utterance string A as a context, a next utterance candidate B, and a label C indicating whether the next utterance candidate B is appropriate as the next utterance of the utterance string A. As an input. A set of utterance string evaluation data D is a learning data group in the learning phase. FIG. 2 is a diagram showing an example of a set of utterance string evaluation data D. As shown in FIG. Each row is one utterance column evaluation data D.

  The context utterance string A is a string of utterances exchanged between the system and the user. The number of utterances constituting the utterance string A is arbitrary, but may be a certain number N (for example, 2).

  When the next utterance candidate B is appropriate as the next utterance of the utterance string A, the label C takes the value 1, and when it is not appropriate, the label C takes the value 0. The data composed of the utterance string A and the next utterance candidate B is positive example data if the label C is 1, and is negative example data if the label C is 0.

  The positive example data is a sequence of N + 1 consecutive utterances from a person-to-person dialogue log, the utterance sequence A is the first N utterance sequence, and the next utterance candidate B is the N + 1th utterance. Can also be configured. The negative example data is a sequence of N + 1 consecutive utterances from the person-to-person dialogue log, the utterance string A is the first N utterance strings, and the next utterance candidate B is other than the N + 1th utterance. It can also be configured by making an arbitrary utterance not suitable as the next utterance.

  The calculation means 20 includes an utterance string evaluation data conversion means 21, a concept base 22, a learning means 23, and a classification model 24.

  The concept base 22 stores a set of pairs of a word and a concept vector representing the meaning of the word. Only the content words such as nouns, verbs, and adjectives may be registered in the concept base 22. FIG. 3 is an example of the concept base 22. The concept base 22 is generated, for example, by the method of Non-Patent Document 2 (Non-Patent Document 2: Katsuto Bessho, Toshiro Uchiyama, Jun Uchiyama, Ryoji Kataoka, Masahiro Oku, “Corpus Concept Based on Co-occurrence between Words and Meaning Attributes” Generating method of the base, “Journal of Information Processing Society of Japan, Dec. 2008, Vol.49, No.12, pp.3997-4006.).

  In the concept base 22, the word is registered with the word end form. When the concept base 22 is searched, the word is searched with the word end form. The concept vector of each word is a d-dimensional vector normalized to a length of 1, and the concept vectors of semantically close words are arranged nearby.

  In the processing of the present invention, a concept vector of a given text is searched for the concept base 22 with each word obtained by dividing the text into words, the acquired concept vectors are added, and the resulting concept vector is obtained. Is generated by normalizing to length 1. Here, the concept vector of the text may be generated using only the content words in the words in the text.

  When the text is an utterance sequence composed of a plurality of utterances, when a concept vector is obtained for each utterance, the quality of the utterance concept vector is low when there are not enough content words in a certain utterance. there is a possibility. Even when there are not enough content words for each utterance, the content words may be sufficiently arranged in the entire utterance string. In such a case, the concept vector of the utterance string has high quality.

  Needless to say, a concept vector may be obtained for each utterance in the utterance sequence, and the sum of the acquired concept vectors normalized to length 1 may be used as the concept vector of the utterance sequence.

  Hereinafter, processing of the speech string evaluation data conversion unit 21 will be described.

  The following processing is performed on each utterance string evaluation data D.

  Referring to the concept base 22, a concept vector E of the utterance string A is generated, and a concept vector F of the next utterance candidate B is generated.

  A post-conversion utterance string evaluation data H composed of a combination of a 2d dimensional concept vector G obtained by combining the d dimensional concept vector E and the d dimensional concept vector F and a label C is generated. The post-conversion utterance string evaluation data H is learning data composed of a 2d-dimensional feature vector and a classification label C. If the label C is 1, this 2d-order elemental vector is positive data, and the label C is 0. If so, it is negative example data. The post-conversion utterance string evaluation data H represents whether the concept vector of the context utterance string and the concept vector of the next utterance candidate are a suitable combination as the context utterance string and the next utterance regardless of the similarity. Yes.

  The utterance string evaluation data conversion means 21 outputs a set of converted utterance string evaluation data H generated for each utterance string evaluation data D. FIG. 4 is a diagram showing an example of a set of post-conversion utterance string evaluation data H. Each row is the post-conversion utterance string evaluation data H.

  The learning means 23 classifies an arbitrary 2d-dimensional concept vector as a value 1 of the label C using an algorithm such as a support vector machine (abbreviation: SVM) using the set of converted utterance string evaluation data H as a learning data group. The classification model 24 for calculating the score to be generated is generated.

  Regardless of the similarity of a set of concept vectors of an arbitrary context utterance sequence and a concept vector of an optional next utterance candidate, the combined concept vectors are generally high scores if they are close to the positive example data group. The model is generated so that the score is low if it is close to the negative example data group.

  FIG. 5 is an example of the processing flow of the utterance string evaluation data conversion means 21 and the learning means 23.

  First, in step S1, the concept base 22 is referred to each utterance string evaluation data D, a concept vector E of the utterance string A is generated, a concept vector F of the next utterance candidate B is generated, and a d-dimensional concept vector E And post-conversion utterance string evaluation data H consisting of a combination of a 2d-dimensional concept vector G and a label C, which is a combination of the d-dimensional concept vector F.

  Then, in step S2, a classification for calculating a score by which an arbitrary 2d-dimensional concept vector is classified as a value 1 of label C based on the set of post-conversion utterance string evaluation data H generated in step S1. A model 24 is generated.

  FIG. 6 is a configuration example of the next utterance candidate scoring device according to claim 2 of the present invention. In addition, about the part which becomes the structure similar to the structural example of the next utterance candidate scoring apparatus of Claim 1, the same code | symbol is attached | subjected and description is abbreviate | omitted.

  As shown in FIG. 6, the next utterance candidate scoring device 200 according to the embodiment of the present invention includes a CPU, a RAM, and a ROM that stores programs and various data for executing processing routines to be described later. It can be configured with a computer including. Functionally, the next utterance candidate scoring device 200 includes an input unit 10, a calculation unit 220, and an output unit 230 as shown in FIG.

  The input means 10 is a set of utterance string evaluation data D comprising a combination of a utterance string A as a context, a next utterance candidate B, and a label C indicating whether the next utterance candidate B is appropriate as the next utterance of the utterance string A. As an input.

  Moreover, the input means 10 receives the utterance string I that is the input context and the next utterance candidate set.

  The utterances exchanged so far between the system and the user are arranged in chronological order, and are set as x1, x2,..., Xm. N (≧ 1) is determined, and the sequence of the last N utterances (x (m−N + 1), x (m−N + 2),..., Xm) As an utterance string I. All utterances (x1, x2,..., Xm) up to immediately before may be input as the utterance string I as the context.

  Separately, a plurality of sentences referring to the focus (topic) extracted from the utterance sequence so far are taken as the next utterance candidates from the database, and are input as the next utterance candidate set.

  The calculation means 220 includes an utterance string evaluation data conversion means 21, a concept base 22, a learning means 23, a classification model 24, and an evaluation means 221.

  FIG. 7 is an example of a processing flow of the evaluation unit 221. The processing content of the evaluation means 221 is demonstrated along FIG.

  First, in step S3, the concept base 22 is referred to generate a concept vector J of the utterance sequence I.

  Next, the following steps S4 to S6 are performed on each next utterance candidate K in the next utterance candidate set.

  In step S4, the concept base 22 is referred to generate a concept vector L of the next utterance candidate K.

  In step S5, a 2d-dimensional concept vector M is generated by combining the d-dimensional concept vector J and the d-dimensional concept vector L.

  In step S <b> 6, by referring to the classification model 24, a score for classifying the 2d-dimensional concept vector M into the value 1 of the label C is calculated. This score becomes larger as the degree of classification into value 1 is larger.

  For a set of concept vector J and concept vector L, regardless of their similarity, the combined concept vector M generally has a high score if it is close to the positive example data group, and a low score if it is close to the negative example data group. It becomes. That is, regardless of the closeness of the contents of the context utterance sequence I and the next utterance candidate K, a high score is obtained if the correspondence between the contents is close to the correspondence between the context utterance sequence A and the next utterance candidate B in the positive example data. If the correspondence between the context utterance string A and the next utterance candidate B in the negative example data is close, the score is low. Therefore, the next utterance candidate suitable as the next utterance of the context utterance string I is selected regardless of the closeness of the content with the context utterance string I.

  FIG. 8 is a diagram showing a set of sets of the concept vector M output from the evaluation unit 221 and the score calculated for the concept vector M.

  In step S7, the next utterance candidate that gave the highest score is selected, and the system returns it to the user. Alternatively, the system returns to the user the next utterance candidate randomly selected from the next utterance candidate having a score equal to or higher than a certain threshold or the next utterance candidate having the highest score.

  If no concept vector is generated from the utterance string I or the next utterance candidate K, the concept vector M is not generated, and therefore the score of the next utterance candidate K is not calculated. The next utterance candidate whose score is not calculated is assumed to have a lower rank than the next utterance candidate whose score is calculated, and the next utterance candidate whose score is not calculated is randomly ordered.

  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.

  In both claim 1 (learning phase) and claim 2 (classification phase), the number of utterances in the context utterance sequence is a fixed number N, and a concept vector is generated for each utterance in the context utterance sequence. It is possible to extend such that the d (N + 1) -dimensional concept vector obtained by combining the concept vector and the concept vector generated from the next utterance candidate is processed as a concept vector G or a concept vector M equivalent. This corresponds to interpretation of the concept vectors E and J of the context utterance sequence as a combination of the concept vectors of the N utterances in the context utterance sequence.

  When the k-nearest neighbor method is adopted as the classification method, the learning unit 23 does not perform anything in particular, and the set of the converted utterance string evaluation data H output from the utterance string evaluation data conversion unit 21 is used as the classification model 24. The evaluation means 221 takes k concept vectors G from the classification model 24 in the order closer to the concept vector M, and uses the value of the concept vector G with the larger number as the score of the concept vector M. And labels.

  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, 220 Calculation means 21 Utterance string evaluation data conversion means 22 Concept base 23 Learning means 24 Classification model 100, 200 Next utterance candidate scoring device 221 Evaluation means 230 Output means

Claims (5)

A concept base that is a set of pairs of a word and a concept vector representing the meaning of the word;
An input is a set of utterance string evaluation data D composed of a combination of a utterance string A as a context, a next utterance candidate B, and a label C indicating whether the next utterance candidate B is appropriate as the next utterance of the utterance string A. , Referring to the concept base for each utterance string evaluation data D, generating a concept vector E of the utterance string A, generating a concept vector F of the next utterance candidate B, and the concept vector E and the concept vector Utterance string evaluation data conversion means for generating post-conversion utterance string evaluation data H comprising a combination of the concept vector G combined with F and the label C;
Learning means for generating a classification model for calculating a score in which an arbitrary concept vector having the same dimension as the concept vector G is classified into one value of the label C from the set of the converted utterance string evaluation data H When,
The next utterance candidate scoring device, comprising: With the utterance string I as a context and the next utterance candidate set as inputs, refer to the concept base, and generate a concept vector J of the utterance string I,
For each next utterance candidate K in the next utterance candidate set, a concept vector L of the next utterance candidate K is generated, a concept vector M obtained by combining the concept vector J and the concept vector L is generated, and the learning means The next utterance candidate score according to claim 1, further comprising: an evaluation unit that calculates a score by which the concept vector M is classified into one value of the label C by referring to the generated classification model. Ring device. A next utterance candidate scoring method in a next utterance candidate scoring device including a concept base that is a set of a pair of a word and a concept vector representing the meaning of the word, an utterance string evaluation data conversion unit, and a learning unit ,
The utterance string evaluation data conversion means includes a combination of a utterance string A as a context, a next utterance candidate B, and a label C indicating whether the next utterance candidate B is appropriate as the next utterance of the utterance string A. A set of sequence evaluation data D is input, and for each utterance sequence evaluation data D, the concept base is referred to, a concept vector E of the utterance sequence A is generated, and a concept vector F of the next utterance candidate B is generated. Generating post-conversion speech string evaluation data H composed of a combination of a concept vector G obtained by combining the concept vector E and the concept vector F and the label C;
A classification model for the learning means to calculate a score in which an arbitrary concept vector having the same dimension as the concept vector G is classified into one value of the label C from the set of converted utterance string evaluation data H The next utterance candidate scoring method. The evaluation means receives the utterance sequence I as a context and the next utterance candidate set as input, refers to the concept base, and generates a concept vector J of the utterance sequence I,
For each next utterance candidate K in the next utterance candidate set, a concept vector L of the next utterance candidate K is generated, a concept vector M obtained by combining the concept vector J and the concept vector L is generated, and the learning means The next utterance candidate scoring method according to claim 3, further comprising: calculating a score by which the concept vector M is classified into one value of the label C by referring to the generated classification model.   The program for functioning a computer as each means which comprises the next speech candidate scoring apparatus of Claim 1 or Claim 2.

Images