JP2008293098A - Answer score information generation device and interactive processor - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an interactive processor, capable of making interaction long-lasting by selecting an appropriate answer, and an answer score information generation device provided to the interactive processor. <P>SOLUTION: The answer score information generation device comprises an answer knowledge information storage means which stores answer knowledge information including a plurality of pieces of speech information showing a speech and answer information showing an answer responding to each speech; a conversion means which converts the answer knowledge information stored by the storage means to a speech transition matrix showing a correspondence between speech information and answer information; a calculation means which calculates a maximum unique vector of the speech transition matrix converted by the conversion means; an answer score information generation means which generates answer score information by associating each element of the maximum unique vector calculated by the calculation means with each answer shown in the answer information; and an answer score information storage means which stores answer score information generated by the answer score information generation means. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to a response score information generation device and a dialogue processing device, and more particularly to a response score information generation device and a dialogue processing device for enabling a long-lasting dialogue with a user.

  With recent improvements in speech recognition technology, various interactive processing devices have been proposed. For example, Patent Document 1 proposes an interactive processing device that generates an utterance using response knowledge that can be reduced to a format composed of events and events, events and evaluations, and combinations of evaluations and evaluations.

  In this dialogue processing apparatus, matching is performed with utterances in response knowledge based on the similarity of word sets included in utterances from utterances input by a user, and utterance candidates to which the apparatus should respond are selected.

  When there are a plurality of selected candidates, the dialogue processing device selects one utterance at random from the utterance candidates or performs processing such as masking the utterance that appears in the dialogue history, The utterance is decided.

JP 2006-201870 A

  However, the technique disclosed in Patent Document 1 has a problem in that since a process of selecting one utterance is performed at random, an appropriate response is not selected and the dialogue does not last long.

  In view of the above problems, an object of the present invention is to provide a dialogue processing device that enables a dialogue to last long by selecting an appropriate response, and a response score information generation device provided for the dialogue processing device. To do.

  In order to achieve the above object, the invention of claim 1 includes a response knowledge information storage means storing response knowledge information including a plurality of utterance information indicating utterances and response information indicating responses corresponding to the utterances. Conversion means for converting the response knowledge information stored by the response knowledge information storage means into an utterance transition matrix indicating a correspondence between the utterance information and the response information; and the utterance transition matrix converted by the conversion means Calculation means for calculating the maximum eigenvector of the response, and response score information generation means for generating response score information in which each element of the maximum eigenvector calculated by the calculation means is associated with each response indicated by the response information; Response score information storage means for storing the response score information generated by the response score information generation means.

  According to the first aspect of the present invention, the response knowledge information storage means stores a plurality of utterance information indicating utterances, and response knowledge information including response information indicating a response corresponding to each of the utterances. The response knowledge information stored by the response knowledge information storage means is converted into an utterance transition matrix indicating a correspondence between the utterance information and the response information, and a calculation means converts the utterance transition matrix converted by the conversion means. A response eigenvector is calculated, and response score information generating means generates response score information in which each element of the maximum eigenvector calculated by the calculating means is associated with each response shown in the response information, and response score information The storage means stores the response score information generated by the response score information generation means. Since each element of the maximum eigenvector indicates the degree to which the corresponding response is an appropriate response, it is possible to provide a response score information generating device for making the dialogue last longer.

  In order to solve the above-mentioned problem, the invention of claim 2 has the response score information generation device of claim 1, wherein the input means recognizes speech of a user by speech and recognizes the user's speech as the speech. Response information is input as information, and response information extraction means extracts response information indicating a response corresponding to the utterance indicated by the utterance information input by the input means from the response knowledge information stored by the response knowledge information storage means. The element acquisition unit acquires the element of the maximum eigenvector corresponding to the response indicated by the response information extracted by the response information extraction unit from the response score information stored by the response score information storage unit, and selects response information The means selects response information indicating a response corresponding to the largest element among the elements acquired by the element acquisition means, and the voice output means selects the response information. Response information selected by the-option means, and outputs the converted to speech. Each element of the maximum eigenvector indicates the degree to which the corresponding response is an appropriate response. Therefore, by selecting an appropriate response for the user's utterance and outputting the response, the interaction is performed. It is possible to provide an interactive processing device that can last for a long time.

  Further, the invention of claim 2 further comprises speech promotion response information storage means for storing speech promotion response information indicating a predetermined response that prompts the user to prompt the next speech as in the invention of claim 3. The response information selection unit selects the speech promotion response information stored in the speech promotion response information storage unit when the response information is not extracted by the response information extraction unit.

  According to the third aspect of the present invention, when it is not extracted, utterance promotion response information indicating a predetermined response that prompts the user to utter the next utterance is selected, so that the dialogue can be continued for a long time.

  Advantageous Effects of Invention According to the present invention, there is an effect that it is possible to provide a dialog processing device that enables a dialog to last long by selecting an appropriate response, and a response score information generation device provided for the dialog processing device. It is done.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. In the present embodiment, an example in which the response score information generation device and the dialogue processing device are applied to a personal computer will be described.

  First, the configuration of the personal computer 12 will be described with reference to FIG. The personal computer 12 includes a CPU (Central Processing Unit) 60, a ROM (Read Only Memory) 61, a RAM (Random Access Memory) 62, a HDD (Hard Disk Drive) 63, and a display unit 64, each of which is connected by a bus B. An operation input unit 65, a microphone 66, and a speaker 67.

  The CPU 60 controls the overall operation of the personal computer 12, and a program to be described later is executed by the CPU 60. The ROM 61 is a non-volatile storage device that stores a boot program that operates when the personal computer 12 is started up. The RAM 62 is a volatile storage device in which an OS (Operating System), programs, and data are expanded. The HDD 63 includes a later-described response knowledge database (hereinafter referred to as response knowledge DB), a response score database (hereinafter referred to as response score DB), an utterance promotion response database (hereinafter referred to as utterance promotion response DB), an OS, and a program. And the like, and corresponds to response knowledge information storage means, response score information storage means, and utterance promotion response information storage means.

  The display unit 64 displays various information. The operation input unit 65 is used when the user operates the personal computer 12 or inputs information to the personal computer 12. The microphone 66 includes an analog / digital converter, and converts a user's utterance into a digital signal and outputs it to the bus B. The speaker 67 includes a digital / analog converter, and converts a digital signal input from the bus B into an analog signal and outputs a sound.

  Next, the response knowledge DB will be described with reference to FIG. The response knowledge DB is a DB including a plurality of utterance information indicating utterances and response information indicating responses corresponding to the utterances.

  Specifically, as shown in the figure, as the utterance information, a plurality of pieces of information indicating utterances such as “make a dog house”, “the dog has grown”, and the like are included as text information. In addition, the response information indicates a response such as “Dog is pleased” indicating a response corresponding to “Create a doghouse” or “Create a doghouse” indicating a response corresponding to “The dog has grown” Multiple pieces of information are included as text information.

  Note that although only seven utterance information and corresponding response information are included in the response knowledge DB, the number is not limited to seven.

  Next, the response score DB will be described. Prior to the description, the utterance network and the utterance transition matrix will be described.

  FIG. 3 shows an utterance network showing correspondence between utterance information and response information shown in the response knowledge DB. The response corresponding to “Dog has grown”, which is one of the utterances in the response knowledge DB, is only “Create dog house”. Assuming that “the dog has grown” is n1 and “make the doghouse” is n2, it can be expressed as a network connecting the nodes n1 and n2, as shown in FIG.

  Furthermore, there are two responses corresponding to the above-mentioned “create a dog house”: “the dog is pleased” and “there is a hammer”. Similarly to the previous case, if “the dog is pleased” is n3 and “the gold hammer is” is n4, as shown in FIG. 3, it is expressed as a network connecting the nodes n2 and n3 and the nodes n2 and n4. it can.

  By doing so, the utterance network shown in FIG. 3 can be obtained from the response knowledge DB. The utterance network represented by a matrix having the start node in the utterance network as a row and the end node as a column is shown in FIG. It is a matrix shown in

  In the matrix shown in FIG. 4A, for example, the first row and second column (n1 row and n2 column) is “1”. This indicates that there is a network in which the start node is n1 and the end node is n2. In this case, the above-mentioned “dog has grown” (n1), “make a doghouse” (n2) It corresponds to the network that connects

  Further, as in the second row and fifth column (n2 row and n5 column), “0” indicates that there is no network in which the start node is n2 and the end node is n5.

  Of the elements of this matrix (FIG. 4 (A)), the utterance transition matrix shown in FIG. 4 (B) is obtained by dividing each element present in the same row by the sum of the elements present in the same row. . For example, in FIG. 4A, since the elements belonging to the second line are two “1” and five “0”, the sum of the elements existing in the same line is 2, so the second line Dividing each element belonging to 2 by 2 is the second line shown in FIG.

  That is, the matrix shown in FIG. 4B shows a probability network indicating the probability of transition from a certain start node to each end node. The limit of the stay probability of a node obtained when the probability network is randomly walked can be obtained as a value obtained by normalizing the maximum eigenvector calculated by eigenvalue decomposition of the utterance transition matrix.

  Therefore, when a certain node is connected to a plurality of nodes (there are a plurality of response candidates corresponding to a certain utterance), the stay probability is the highest among the plurality of nodes (among the elements of the maximum eigenvector). The node that is the largest) is considered the node that gives the appropriate response. Therefore, each element of the maximum eigenvector indicates the degree to which the corresponding response is an appropriate response.

  The normalization of the maximum eigenvector is to obtain a vector obtained by dividing each element by the sum of the elements belonging to the maximum eigenvector.

In addition, in order to guarantee the strong connectivity of the utterance transition matrix shown in FIG. 4B, smoothing may be performed so that each element belonging to the utterance transition matrix does not become zero. This smoothing means that when the number of elements in one row of the utterance transition matrix is n, λ is a positive real number smaller than 1, and the element in the i-th row and j-th column of the utterance transition matrix is aij,
(Aij + λ) / (1 + nλ)
Is to replace

Even if replaced in this way, (aij + λ) / (1 + nλ) is smaller than 1, and Σaij = 1 when Σ is the sum of j (from 1 to n).
Σ (aij + λ) / (1 + nλ) = 1
It becomes.

  FIG. 5 shows the maximum eigenvector calculated by eigenvalue decomposition of the utterance transition matrix described above. N1 to n7 written beside the maximum eigenvector shown in the figure correspond to the words such as “the dog has grown” shown in each node in FIG. The maximum eigenvector shown in the figure is not normalized.

  The response score DB shown in the figure is information in which each element of the maximum eigenvector of the utterance transition matrix is associated with each response shown in the response information. In this response score DB, as shown in the figure, each element of the maximum eigenvector is expressed as a score.

  In the matrix shown in FIG. 4A described above, “1” is stored as an element when a network exists, but there is a method of storing a value other than “1”. This method will be described. First, the contents of the response knowledge DB described above can be automatically acquired from documents published on the network (for example, dryness, automatic causal relationship knowledge from a document set based on “connection indicator“ for ”). Acquired ", IPSJ Proceedings, vol1,45, No. 3, pp. 919-933, 2004).

  At this time, since there are cases where duplicate response knowledge is included in the document set of the acquisition source, this appearance count is stored in the response knowledge DB, and the appearance count is stored in the element.

  Since the appearance count is considered to be a value representative of the strength of utterance connection, a more natural and appropriate response can be selected.

  Next, the speech promotion response DB will be described with reference to FIG. This utterance promotion response DB is information indicating a predetermined response that prompts the user to utter the next utterance when response information is not extracted. Here, “response information was not extracted” means that there was no response corresponding to the user's utterance in the response knowledge DB. In that case, since the dialogue is ended, the speech promotion response DB is used to prompt the user for the next speech.

  As shown in the figure, in the utterance promotion response DB, responses that prompt the user to utter the next utterance such as “So?” And “And then?” Are stored.

  Next, functional blocks of the response score information generation device and the dialogue processing device will be described with reference to FIG. In the figure, a transition matrix conversion unit 70, a maximum eigenvector calculation unit 71, an utterance input unit 72, a response selection unit 73, a response output unit 74, a response knowledge DB 80, an utterance transition matrix 81, a response score DB 82, and an utterance promotion response DB 83 It is shown.

  Among these, the transition matrix conversion unit 70 converts the response knowledge DB into an utterance transition matrix indicating correspondence between utterance information and response information. The maximum eigenvector calculation unit 71 calculates the maximum eigenvector of the utterance transition matrix converted by the transition matrix conversion unit 70.

  The utterance input unit 72 recognizes utterances by the user and inputs the utterances of the user as utterance information. Note that the utterance input unit 72 also inputs the utterance information as “create kennel” even when the user utters “I made a kennel”, for example. That is, the utterance input unit 72 is adapted to match a noun and a verb.

  The response selection unit 73 extracts response information indicating a response corresponding to the utterance indicated by the utterance information input by the utterance input unit 72 from the response knowledge DB 80, and converts the response information indicated by the extracted response information from the response score DB 82. A corresponding maximum eigenvector element is acquired, and response information indicating a response corresponding to the largest element among the acquired elements is selected.

  Moreover, the response selection part 73 is also what selects the speech promotion response information memorize | stored by speech promotion response DB83, when response information is not extracted.

  The response output unit 74 converts the response information selected by the response selection unit 73 into sound and outputs it.

  The response knowledge DB 80, the utterance transition matrix 81, the response score DB 82, and the utterance promotion response DB 83 are as described above.

  In the functional blocks described above, the functional blocks as the response score information generation device include the transition matrix conversion unit 70, the maximum eigenvector calculation unit 71, the response knowledge DB 80, the utterance transition matrix 81, and the response score DB 82.

  In addition to the functional block as the response score information generation device, the functional block as the dialogue processing device includes an utterance input unit 72, a response selection unit 73, a response output unit 74, and an utterance promotion response DB 83.

  Hereinafter, processing executed by the personal computer 12 will be described using a flowchart. This process is executed by the CPU 60.

  First, a response score information generation process executed by the personal computer 12 as a response score information generation device will be described using the flowchart of FIG. In this flowchart, the utterance in the response knowledge DB 80 shown in FIG. 2 and the response corresponding to the utterance are expressed as an utterance pair.

  First, in step 101, the matrix element corresponding to the utterance pair stored in the response knowledge DB 80 is set to 1 (see FIG. 4A). In step 102, it is determined whether or not the processing in step 101 has been completed for all utterance pairs included in the response knowledge DB 80. If not, the processing in step 101 is performed. The processing is shifted to 103.

  In step 103, each element present in one same row is divided by the sum of the elements present in the same row (see FIG. 4B). The smoothing described above is performed in step 104 on the elements in the row. In the next step 105, it is determined whether or not the processing of steps 103 and 104 for all rows has been completed. If not, the processing of step 103 is performed. If the processing has been completed, the processing of step 106 is performed. Transition. Through the above processing, the response knowledge DB 80 is converted into an utterance transition matrix indicating correspondence between utterance information and response information.

  In the next step 106, the maximum eigenvector is calculated by eigenvalue decomposition of the utterance transition matrix. In the next step 107, a response score DB is generated using the maximum eigenvector (see FIG. 5), and the process ends.

  Since each element of the maximum eigenvector in the response score information generation process described above indicates the degree to which the corresponding response is an appropriate response, a response score information generation device for continuing the dialogue is provided. be able to.

  Next, dialogue processing executed by the personal computer 12 as the dialogue processing device will be described using the flowchart of FIG. First, in step 201, the user's speech is recognized and speech information is input. In the next step 202, the utterance information stored in the response knowledge DB 80 is compared with the inputted utterance information to determine whether there is a match. If a negative determination is made in step 202, the process proceeds to step 204.

  On the other hand, if an affirmative determination is made in step 202, a response corresponding to the utterance stored in the response knowledge DB 80 that matches the utterance information by the user is set as a response candidate in step 203, and the process proceeds to step 204.

  In the next step 204, it is determined whether or not all the utterance information stored in the response knowledge DB 80 and the utterance information by the user have been compared. If so, the process proceeds to step 205.

  Through the processing from step 202 to step 204, response information (response candidate) indicating the response corresponding to the utterance indicated by the input utterance information is extracted.

  In the next step 205, it is determined whether or not the response candidate is 0. If the response candidate is 0, that is, if response information is not extracted, in step 206, the speech promotion response is selected from the speech promotion response DB 83 as response information, and the process proceeds to step 209.

  As described above, when no response candidate is extracted, utterance promotion response information indicating a predetermined response that prompts the user to utter the next utterance is selected, so that the conversation can be continued for a long time.

  On the other hand, if a negative determination is made in step 205, the element (score) corresponding to the response candidate is acquired from the response score DB 82 in step 207, and the response candidate corresponding to the largest element is selected as the response information in step 208. Choose as.

  In the next step 209, the response information is converted into voice and output, and the process of step 201 is performed again.

  According to this dialogue processing, for example, when the user speaks “make a doghouse”, “the dog is pleased” or “there is a hammer” is a response candidate. The element corresponding to “the dog is pleased” is 0.569, and the element corresponding to “has a hammer” is 0.012. Therefore, “the dog is pleased” is selected as the response. Such a statement may be processed by adding “ne” to the end of the sentence and outputting the sound.

  Since each element of the maximum eigenvector in the dialog processing described above indicates the degree to which the corresponding response is an appropriate response, an appropriate response is selected for the user's utterance and the response is output. Thus, it is possible to provide a dialogue processing apparatus that makes it possible to continue the dialogue.

It is a figure which shows the structure of a personal computer. It is a figure which shows a response knowledge database. It is a figure which shows an utterance network. It is a figure which shows an utterance transition matrix. It is a figure which shows a maximum eigenvector and a response score database. It is a figure which shows an utterance promotion response database. It is a functional block diagram of a response score information generation device and a dialogue processing device. It is a flowchart which shows a response score information generation process. It is a flowchart which shows a dialogue process.

Explanation of symbols

12 PC 60 CPU
63 HDD
66 microphone 67 speaker 70 transition matrix conversion unit 71 maximum eigenvector calculation unit 72 utterance input unit 73 response selection unit 74 response output unit 80 response knowledge DB
81 Utterance transition matrix 82 Response score DB
83 Utterance promotion response DB

Claims (3)

Response knowledge information storage means storing response knowledge information including a plurality of utterance information indicating utterances, and response information indicating responses corresponding to each of the utterances;
Conversion means for converting the response knowledge information stored by the response knowledge information storage means into an utterance transition matrix indicating correspondence between the utterance information and the response information;
Calculating means for calculating a maximum eigenvector of the utterance transition matrix converted by the converting means;
Response score information generation means for generating response score information in which each element of the maximum eigenvector calculated by the calculation means is associated with each response indicated in the response information;
Response score information storage means for storing the response score information generated by the response score information generation means;
A response score information generation device having A response score information generating device according to claim 1;
Input means for recognizing speech by a user and inputting the user's speech as the speech information;
Response information extraction means for extracting response information indicating a response corresponding to the utterance indicated by the utterance information input by the input means from the response knowledge information stored by the response knowledge information storage means;
Element acquisition means for acquiring the element of the maximum eigenvector corresponding to the response indicated by the response information extracted by the response information extraction means from the response score information stored by the response score information storage means;
Response information selection means for selecting response information indicating a response corresponding to the largest element among the elements acquired by the element acquisition means;
Voice output means for converting the response information selected by the response information selection means into voice and outputting the voice;
A dialogue processing apparatus. Utterance promotion response information storage means for storing utterance promotion response information indicating a predetermined response that prompts the user for the next utterance;
3. The dialogue according to claim 2, wherein the response information selection unit selects the utterance promotion response information stored in the utterance promotion response information storage unit when the response information is not extracted by the response information extraction unit. Processing equipment.

Images