JP2018028925A - Next utterance candidate ranking device, method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To facilitate interaction between a system and a user by appropriate utterance being returned from the system as a next utterance.SOLUTION: In a next utterance candidate ranking device, concept score calculation means 21 inputs a speech string A to be a context and a next utterance candidate set, refers to a concept base 22, generates a concept vector of the speech string A and a concept vector of respective next utterance candidates B in the next utterance candidate set, calculates a score indicating closeness between the concept vector of the speech string A and the concept vector of the next utterance candidate B for each of the next utterance candidate B, and ranks the next utterance candidate sets based on the score.SELECTED DRAWING: Figure 1

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 ranking 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 technique for 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 ranking the next utterance candidate. After ranking, the system returns the top first utterance candidate in the ranking result or the next utterance candidate randomly selected from the top number to the user.

  As such a method, a sequence of N (≧ 1) utterances immediately before is used as a context, and feature quantities such as words and parts of speech and word semantic categories in the context utterance sequence and the next utterance candidate, and supervised learning in advance. There is a method based on cohesiveness, in which a score representing a degree that the next utterance candidate is appropriate as the next utterance is calculated from the obtained model, and the next utterance candidate set is ranked based on the score.

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

  In the method based on cohesiveness described above, it is possible to consider semantic similarity between words belonging to the same node on the thesaurus using feature quantities such as semantic categories of words. There is a problem that similarity cannot be considered. For this reason, if there is a word that is different in character string between the context utterance sequence and the next utterance candidate but is semantically similar, the similarity between the context utterance sequence and the next utterance candidate is unreasonable. There is a problem that the score becomes unreasonably low even though the next utterance candidate is suitable as the next utterance.

  In order to solve this problem, the first object of the present invention is to provide a next utterance candidate ranking apparatus, method, and program capable of considering semantic similarity between arbitrary words. Specifically, this problem is solved by using a concept base that is a set of pairs of a word and a concept vector representing the meaning of the word.

  The above-mentioned method based on the concept base makes it possible to rank the next utterance candidate in consideration of the semantic similarity between arbitrary words. Generally, on the concept base, only concept words are used for content words such as nouns, verbs, and adjectives. It is shaken. Whether it is appropriate for the next utterance is a context utterance string and an attached word such as a sentence ending expression included in the next utterance candidate, which is not considered in the method using the concept base. Further, when a concept vector is not generated because there is no content word in the context utterance sequence or the next utterance candidate, the score of the next utterance candidate cannot be calculated.

  In order to solve this problem, the second object of the present invention is to utilize the advantages of both the cohesiveness-based method and the concept-based method, and to compensate for the shortcomings on the other side, thereby providing a highly accurate next utterance candidate ranking device and method. And providing a program.

  In order to achieve the first object, the next utterance candidate ranking 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 that is a context, and the next The utterance candidate set is input, the concept base is referred to, a concept vector of the utterance sequence A is generated, a concept vector of each next utterance candidate B in the next utterance candidate set is generated, On the other hand, it includes a score calculation means for calculating a score representing the proximity between the concept vector of the utterance string A and the concept vector of the next utterance candidate B, and ranking the next utterance candidate set based on the score. It is configured.

  Further, the next utterance candidate ranking method according to the present invention provides a next utterance candidate in a next utterance candidate ranking 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, and a concept score calculating means. In the ranking method, the concept score calculating means receives the utterance string A as a context and a next utterance candidate set as input, generates a concept vector of the utterance string A by referring to the concept base, and generates the next utterance. A concept vector of each next utterance candidate B in the candidate set is generated, and a score representing the proximity between the concept vector of the utterance string A and the concept vector of the next utterance candidate B is calculated for each next utterance candidate B; The next utterance candidate set is ranked based on the score.

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

  In order to achieve the second object, the next utterance candidate ranking device includes a model obtained by supervised learning, and a feature amount regarding each utterance candidate B in the utterance string A and the next utterance candidate set. From each of the next utterance candidates B, a unity score calculating means for calculating a score indicating a degree to which the next utterance candidate B is suitable as the next utterance of the utterance string A and ranking the next utterance candidate set based on the score; For each next utterance candidate B in the next utterance candidate set, when the ranking in the ranking result by the concept score calculating means is C and the ranking in the ranking result by the cohesion score calculating means is D, the concept score calculating means In the case where the concept vectors of the utterance string A and the next utterance candidate B are generated, the value obtained by weighting and adding the rank C and the rank D , The combination order of the next utterance candidate B, and if the concept score calculation means does not generate the utterance string A or the concept vector of the next utterance candidate B, the rank D is the combination rank of the next utterance candidate B, and Integration means for ranking the next utterance candidate set in ascending order of the combination order.

  According to the next utterance candidate ranking apparatus, method, and program according to the present invention, the concept base is referred to, the concept vector of the utterance string A is generated, and the concept vector of each next utterance candidate B in the next utterance candidate set is generated. For each next utterance candidate B, a score representing the closeness between the concept vector of the utterance sequence A and the concept vector of the next utterance candidate B is calculated, and the next utterance candidate set is ranked based on the score to obtain the next utterance. Since the system returns a suitable utterance, there is an effect that the interaction between the system and the user becomes smooth.

It is a block diagram which shows the functional structure of the next utterance candidate ranking apparatus which concerns on embodiment of this invention. It is a figure which shows the example of a concept base. It is a flowchart figure which shows the process routine in the next utterance candidate ranking apparatus which concerns on embodiment of this invention. It is a block diagram which shows the functional structure of the next utterance candidate ranking apparatus which concerns on embodiment of this invention. It is a flowchart figure which shows the process routine in the next utterance candidate ranking apparatus which concerns on embodiment of this invention. It is the figure which showed the result of evaluation. It is the figure which showed the result of evaluation for every weight (alpha).

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

<Configuration of next utterance candidate ranking device>
Next, the configuration of the next utterance candidate ranking apparatus according to the embodiment of the present invention will be described. FIG. 1 is a configuration example of a next utterance candidate ranking apparatus according to claim 1 of the present invention. As shown in FIG. 1, a next utterance candidate ranking apparatus 100 according to an embodiment of the present invention includes a CPU, a RAM, and a ROM that stores a program and various data for executing a processing routine described later. Can be configured with a computer. The next utterance candidate ranking apparatus 100 functionally includes an input unit 10, a calculation unit 20, and an output unit 30 as shown in FIG.

  The input unit 10 receives the utterance string A 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 A. As the utterance string A, all utterances (x1, x2,..., Xm) up to immediately before may be input.

  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 20 includes a concept score calculation means 21 and a concept base 22.

  The concept base 22 stores a set of pairs of words (content words such as nouns, verbs, and adjectives) and concept vectors representing the meaning of the words. FIG. 2 is an example of the concept base 22. The concept base 22 is generated by, for example, 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.).

  The content word is registered with the final form of the content word, and when searching the concept base 22, the content word is searched with the final form of the content word. The concept vector of each content word is a k-dimensional vector normalized to a length of 1, and the concept vectors of content words that are semantically close are arranged nearby. The similarity between content words is calculated by, for example, an inner product between corresponding concept vectors.

  Of course, words other than the content words may be registered in the concept base 22, and when the concept vector of the given text is generated, the concept base 22 is searched for words other than the content words in the text. The acquired word concept vector may also be used. In the following, only content words are registered in the concept base 22, and when generating a concept vector of a given text, only the word concept vector obtained by searching the concept base 22 with the content words in the text is used. Describe as using.

  FIG. 3 is an example of the processing flow of the concept score calculation means 21. The processing content of the concept score calculation means 21 is demonstrated along FIG.

  First, in step S1, the concept base 22 is searched for each content word obtained by dividing all the utterances in the utterance string A into words, the acquired concept vectors are added, and the resulting concept vector has a length of 1 Is normalized as a concept vector VA of the utterance string A.

  In step S2, the concept base 22 is searched for each content word obtained by dividing the next utterance candidate B into words for each next utterance candidate B in the next utterance candidate set, and the acquired concept vector is added. The concept vector obtained as a result of normalization to the length 1 is generated as the concept vector VB of the next utterance candidate B.

  In the next step S3, for each next utterance candidate B in the next utterance candidate set, the concept vector VA of the utterance string A generated in step S1 and the concept vector VB of the next utterance candidate B generated in step S2. Based on the above, the score of the next utterance candidate B is calculated according to the following equation 1.

  (VA · VB + 1) / 2 (Formula 1)

  Here, VA · VB is an inner product of VA and VB. The score value is from 0 to 1, and the greater the value, the higher the semantic similarity between the utterance string A and the next utterance candidate B.

  In step S4, the next utterance candidate set is ranked in descending order of the score, and the processing routine is terminated. As a result, the next utterance candidate set is ranked in descending order of the semantic similarity with the utterance string A.

  The system returns the next utterance candidate in the top rank in the ranking result or the next utterance candidate randomly selected from the top to the top by the output means 30 to the user.

  Although the character string is different between the utterance string A and the next utterance candidate B, even if there are words that are semantically similar, the calculated score is high. The next utterance candidate is easily selected.

  In the above process, the concept vector of the utterance string A was obtained. If a concept vector is obtained for each utterance in the utterance string A, the quality of the concept vector of the utterance may be low when there are not enough content words in a certain utterance. Even if there are not enough content words for each utterance, the entire utterance string A may have sufficient content words, and in such a case, the concept vector of the utterance string A has high quality. By the above process, a higher quality concept vector can be generated.

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

  Another method for calculating the score is as follows. For example, a concept vector is obtained for each utterance in the utterance string A, and a concept vector of the next utterance candidate B is obtained. When the concept vector of at least one utterance in the utterance string A and the concept vector of the next utterance candidate B can be acquired, the average of the squares of the distances between the acquired centroids of all concept vectors and the acquired concept vectors Let the square root be the score of the next utterance candidate B. The score value is 0 or more, and the smaller the value, the higher the semantic similarity between the utterance string A and the next utterance candidate B. The next utterance candidate set is ranked in ascending order of scores.

  In each of the score calculation methods described so far, when a concept vector is not generated from the utterance string A or the next utterance candidate B, the score of the next utterance candidate B is not calculated. In this case, the next utterance candidate whose score is not calculated may be lower in rank than the next utterance candidate whose score is calculated, and the next utterance candidate whose score is not calculated may be randomly ordered.

  FIG. 4 is a configuration example of the next utterance candidate ranking apparatus according to claim 2 of the present invention. In addition, about the structure similar to said structural example, the same code | symbol is attached | subjected and description is abbreviate | omitted.

  As shown in FIG. 4, the next utterance candidate ranking apparatus 200 according to the embodiment of the present invention functionally includes an input means 10, a calculation means 220, and an output means 30 as shown in FIG. Yes.

  The calculation means 220 includes a concept score calculation means 21, a concept base 22, a unity score calculation means 223, and an integration means 224.

  The concept score calculation means 21 performs the same processing as the configuration example of the next utterance candidate ranking apparatus according to claim 1 and outputs the ranking result of the next utterance candidate set.

  In the unity score calculation means 223, a model for determining the degree to which the next utterance candidate is suitable as the next utterance obtained by supervised learning, the word included in each utterance candidate B in the utterance string A and the next utterance candidate set, A score representing the degree of appropriateness of the next utterance candidate B as the next utterance of the utterance string A is calculated for each next utterance candidate B from the feature quantities related to the part of speech, etc., and the next utterance candidate set is ranked based on the score. The ranking result of the utterance candidate set is output. This processing is performed by, for example, the following first method or second method.

  In the first method, as supervised learning, various feature quantities are extracted from an utterance string to which a cohesiveness label is assigned, and when these feature quantities are included in the utterance string under any condition The pattern mining technique is used to discover whether the cohesiveness is high or low. As the feature quantity, dialogue action, question type, predicate term structure, information on specific expression, information on dependency structure are used. These feature quantities in the utterance string are expressed as a tree structure. The pattern (partial tree) discovered by pattern mining is used in a process of calculating a score representing a degree suitable for the next utterance candidate B as the next utterance of the utterance string A. Specifically, the score is calculated on the assumption that the one that includes many subtrees that contribute to the high cohesiveness has high cohesiveness.

  In the second method, as supervised learning, for example, an utterance sequence having high cohesiveness and an utterance sequence having low cohesiveness are prepared, and a classifier that performs binary classification of whether the cohesiveness is high or low is learned from these. . In the second method, a learning example such as a dialogue with various topics is created using a pseudo negative example.

  Specifically, a large amount of dialogue data is prepared first. Any natural dialogue spoken by humans can be used. Then, an utterance string of a certain length is sampled from these dialogue data. Let this utterance string be E. At the same time, an utterance string (utterance string F) in which the last utterance in the utterance string E is randomly replaced with another utterance is also created. The supervised learning method is applied by setting the utterance string E as a positive example and the utterance string F as a negative example (pseudo negative example).

  Note that ranking learning is used as a learning method. Specifically, the ranking rule is learned so that the utterance string E is ranked higher than the utterance string F.

  In the second method, as a feature amount representing an utterance string, a word, a part of speech, and a semantic attribute included in the last utterance, and a word, a part of speech, and a semantic attribute included in each of the previous utterances are included. All combinations are used as feature quantities.

  In the second method, the next utterance candidate B follows the utterance string A from the model obtained by supervised learning as described above and the feature amount of each utterance candidate B in the utterance string A and the next utterance candidate set. A score representing a degree suitable for utterance is calculated.

  The integration unit 224 receives the ranking result of the next utterance candidate set output by the concept score calculation unit 21 and the ranking result of the next utterance candidate set output by the cohesion score calculation unit 223 as inputs.

  The rank in the ranking result by the conceptual score calculation means 21 of each next utterance candidate B in the next utterance candidate set is C, and the rank in the ranking result by the cohesion score calculation means 223 is D.

  The joining rank U of the next utterance candidate B is calculated as follows. The weight α is a constant between 0 and 1.

  When the concept score calculation unit 21 generates the concept vectors of the utterance string A and the next utterance candidate B, that is, when the score of the next utterance candidate B is calculated, the combination rank U is calculated according to the following equation. .

    U: = α · C + (1-α) · D

  When the concept score calculation unit 21 does not generate the concept vector of the utterance string A or the next utterance candidate B, that is, when the score of the next utterance candidate B is not calculated, the combination rank U is calculated according to the following equation.

    U: = D

  The integration unit 224 ranks the next utterance candidate set in ascending order of the combination rank U.

  The system returns the next utterance candidate in the top rank in the ranking result or the next utterance candidate randomly selected from the top to the top by the output means 30 to the user.

  For the next utterance candidate whose score is calculated by the concept score calculating means 21, the ranking considering the semantic similarity between the context utterance sequence and the next utterance candidate by the concept score calculating means 21, and the cohesion score calculating means 223. By combining the context utterance sequence and the rank in consideration of language expressions other than the content words in the next utterance candidate by weighting, a combined rank taking into account the advantages of both can be obtained.

  For the next utterance candidate whose score has not been calculated by the concept score calculation means 21, the rank by the cohesion score calculation means 223 is used as a combination rank, so that a language expression other than the content words in the context utterance string and the next utterance candidate can be obtained. The order considered can be reflected.

  FIG. 5 is an example of a processing flow by the concept score calculation unit 21, the unity score calculation unit 223, and the integration unit 224. In addition, about the process similar to the process flow of FIG. 3, the same code | symbol is attached | subjected and detailed description is abbreviate | omitted.

  First, in step S1, a concept vector VA of the utterance string A is generated. In step S2, a concept vector VB of the next utterance candidate B is generated for each next utterance candidate B in the next utterance candidate set.

  In the next step S3, for each next utterance candidate B in the next utterance candidate set, the concept vector VA of the utterance string A generated in step S1 and the concept vector VB of the next utterance candidate B generated in step S2. Based on the above, the score of the next utterance candidate B is calculated according to Equation 1 above.

  In step S4, the next utterance candidate set is ranked in descending order of the score.

  In the next step S21, feature quantities are extracted for the utterance string A and each utterance candidate B. In step S22, for each next utterance candidate B, the next utterance candidate B is determined based on the model obtained by supervised learning and the feature amount of the utterance string A and the next utterance candidate B extracted in step S21. A score representing a degree suitable for the next utterance of the utterance string A is calculated.

  In step S23, the next utterance candidate set is ranked in descending order of the score calculated in step S22.

  In step S24, each next utterance candidate B is determined based on whether the concept vector is generated in steps S1 and S2, the ranking result obtained in step S4, and the ranking result obtained in step S23. On the other hand, the combination rank U is calculated, the next utterance candidate set is ranked in ascending order of the combination rank U, and the processing routine is terminated.

  The method of the present invention was evaluated as follows.

  In the present invention, the next utterance is one in which the system speaks in accordance with the previous context. Therefore, in a dialogue log between two humans (each utterance is tagged with a speaker and a topic), an utterance boundary E that satisfies the following conditions (1) and (2) is targeted: From the context up to the utterance immediately before, it was taken as a problem of deriving the utterance immediately after E.

Condition (1) [Speaker just before E] ≠ [Speaker just after E]
Condition (2) [Topic of utterance immediately before E] = [Topic of utterance immediately after E]

  The two utterances immediately before E are set as a context utterance string, and the utterance immediately after E is set as a correct answer. In the dialogue log, all utterances from the utterance immediately after E are set as the next utterance candidates, and it is evaluated whether the correct utterance can be ranked higher in the next utterance candidate ranking result.

  As an evaluation index, MRR (Mean Reciprocal Rank: average of the reciprocal of the rank of correct utterances) was used. The higher this value, the higher the accuracy.

  Since there are 1746 dialogue logs and the average number of utterances in one dialogue log is about 39, in each dialogue log, there is one utterance boundary E that satisfies the above condition that the number of utterance candidates is about half, which is about 20. Determined. As a result, the number of target utterance boundaries is 1746, and the method by the concept score calculation means (the score of the next utterance candidate is calculated by the above equation 1) and the technique by the cohesion score calculation means (the score of the next utterance candidate is the second utterance The MRR calculated by the method is as shown in FIG.

  In the integration unit, the MRR when the weight α is changed from 0 to 1 in increments of 0.1 is as shown in FIG.

  As α increases, the MRR increases and decreases as the maximum value is reached. In the above, when α is 0.7, MRR has a maximum value of 0.424. The integration means improved the MRR from 0.362 to 0.424 as compared to the concept score calculation means alone.

  Among the 1746 target utterance boundaries, the concept score calculation means does not generate a context utterance sequence or a concept vector of correct utterances, and 433 boundaries (total 24.8) have not been calculated. %)Met. For correct utterances for which the score has not been calculated by the concept score calculation means, the order by the unity score calculation means is adopted by the integration means, and it is considered that a higher rank is given.

  As the weight α in the integrating means of the present invention, in some evaluation data, α may be calculated in advance by changing α, and α having the highest accuracy may be adopted.

  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, 220 Calculation means 21 Concept score calculation means 22 Concept base 30 Output means 100, 200 Next utterance candidate ranking apparatus 223 Unity score calculation means 224 Integration means

Claims (7)

A concept base that is a set of pairs of a word and a concept vector representing the meaning of the word;
The context utterance sequence A and the next utterance candidate set are input, the concept base is referred to, a concept vector of the utterance sequence A is generated, and a concept vector of each next utterance candidate B in the next utterance candidate set is generated Then, for each next utterance candidate B, a score representing the closeness between the concept vector of the utterance string A and the concept vector of the next utterance candidate B is calculated, and the next utterance candidate set is ranked based on the score A score calculation means;
Including
The next utterance candidate ranking, wherein the concept vector has a concept for expressing the meaning of a word in each dimension, and the height of the relationship between the word and each concept is the value of each dimension of the concept vector. apparatus.   The next utterance candidate ranking device according to claim 1, wherein the utterance string A includes utterances between a plurality of subjects.   The concept score calculating means generates a concept vector for each utterance of the utterance sequence A when generating the concept vector of the utterance sequence A, and the concept of the utterance sequence A from the generated concept vector for each utterance. The next utterance candidate ranking device according to claim 1 or 2, wherein a vector is generated. The concept score calculating means generates a plurality of concept vectors corresponding to the utterances for a plurality of utterances of the utterance sequence A when generating the concept vectors of the utterance sequence A,
When calculating a score representing the proximity of the concept vector of the utterance string A and the concept vector of the next utterance candidate B, a plurality of concept vectors for a plurality of utterances of the utterance string A, and the next utterance candidate B The next utterance candidate ranking device according to claim 1, wherein a score representing the proximity to the concept vector is calculated. A next utterance candidate ranking method in a next utterance candidate ranking device including a concept base, which is a set of a pair of a word and a concept vector representing the meaning of the word, and a concept score calculation means, wherein the concept score calculation means includes a context Utterance sequence A and next utterance candidate set to be input, refer to the concept base, generate a concept vector of the utterance sequence A, and generate a concept vector of each next utterance candidate B in the next utterance candidate set Calculating a score representing the proximity between the concept vector of the utterance string A and the concept vector of the next utterance candidate B for each next utterance candidate B, and ranking the next utterance candidate set based on the score. Including
The next utterance candidate ranking, wherein the concept vector has a concept for expressing the meaning of a word in each dimension, and the height of the relationship between the word and each concept is the value of each dimension of the concept vector. Method.   The next utterance candidate ranking method according to claim 5, wherein the utterance string A includes utterances between a plurality of subjects.   The program for functioning a computer as each means which comprises the next utterance candidate ranking apparatus of any one of Claims 1-4.