JP2007102104A - Device and method for answer sentence generation, and program and storage medium thereof - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To eliminate the need for complicated rule setting for an interaction system. <P>SOLUTION: An answer sentence generation device 100 is equipped with an interaction corpus storage part 133, which stores a corpus given utterance pairs of user utterances and their answer sentences and label the values of interaction states of the utterance pairs. Namely, an interaction state management part 140 decides what interaction state an input user utterance is in (what kind of label value the input user utterance corresponds to) based on the user utterances and answer sentences, and label values included in the interaction corpus. An answer sentence generation part 150 retrieves an answer sentence to the user utterance similar to the decided interaction state from the interaction corpus. Namely, the answer sentence which is most suitable as an answer sentence to the input user utterance with high possibility is retrieved from the interaction corpus storage part 133. Then the retrieved answer sentence is corrected and output as the answer sentence with respect to the input user utterance. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a response sentence generation technique in a dialog system.

  Natural language dialogue systems are widely used as a means of realizing natural communication between people and machines. Examples thereof include an automatic answering service by telephone and an interactive service using the Web (World Wide Web) (see Non-Patent Document 1, etc.).

In a conventional dialogue system, a structured back end such as a database or ontology is used (see Non-Patent Document 2 to Non-Patent Document 4).
For example, in a bus information guidance system (system) that has fields such as “departure stop name”, “arrival stop name”, “day of the week”, and “time” in the back-end database, “From Hon-Atsugi to Communication Research Institute” Is input from the user.
In this case, with reference to the back-end database, the system can determine that the “departure stop name” is “Honatsugi” and the “arrival stop name” is “in front of the communication laboratory”. Also, based on the input from the system user and the fields set in the backend database, you can see what information is missing from this question.
That is, the system knows that “day of the week” and “time” are missing from the input content from the user. Therefore, the system can determine that the question regarding “day of the week” and “time” should be made next.
"Speech processing application system / voice interface", [online], [searched on September 1, 2005], Internet <URL: http: //www.icot.or.jp/FTS/REPORTS/H10-reports/AITEC9903Re2_Folder /AITEC9903R2-sec3-2.htm> A. Rundnicky and X. Wu, An Agenda-Based Dialog Management Architecture for Spoken Language Systems, Proceedings of the Workshop on Automatic Speech Recognition and Understanding, 1999 M. Denecke and AhWaibel, Dialogue Strategies Guiding Users to their Commnucative Goals, Proceedings of Eurospeech, Rohdos, Greece, 1997 A. Ferrieux and MDSadek, An Efficient Date-Driven Model for Cooperative SpokenDialogue, Proceedings of International Conference on Spoken Language Proceedings, Yokohama, Japan, 1994

However, it takes time to create the rules used for the back end as described above in the interactive system. In particular, when the user has a wide range of questions, it is necessary to set complicated rules according to the back end. Therefore, there is a problem that it takes much time to create the back end. That is, the conventional technique has a problem that the construction cost becomes high when constructing an interactive system such as an interactive question answering system.
Accordingly, an object of the present invention is to provide a response sentence generation device and the like that solve the above-described problems and reduce the construction cost of a dialogue system.

  In order to solve the above-described problem, the response sentence generation device of the present invention, instead of using a back end having a complicated rule, an utterance pair of a user utterance and its response sentence, and a label value of a conversation state in the utterance pair The dialogue corpus to which That is, the response sentence generation device determines what dialogue state the input user utterance has (based on what label value) based on the user utterance and the label value included in the dialogue corpus. Then, a user utterance similar to this dialog state is searched from the dialog corpus, and a response sentence to this user utterance is searched. That is, the most appropriate response candidate sentence is searched from the dialogue corpus as the response sentence of the input user utterance. And it was set as the structure which corrects this searched response candidate sentence and outputs as a response sentence with respect to the input user utterance.

  That is, the present invention is a response sentence generation device that generates a response sentence for a user utterance, the user utterance input unit that receives input of the user utterance, and the input user utterance, A keyword extraction unit for extracting the included keyword, an utterance state storage unit for storing the extracted keyword and document data including the keyword, and a keyword addition unit for adding the extracted keyword to the utterance state storage unit; A document data search unit for searching for document data including the extracted keyword from a document corpus storage unit that stores a document corpus that is a set of predetermined document data; and the utterance state storage unit for the searched document data Utterance state update unit for updating the document data in the utterance state storage unit, user utterance and its response A dialogue corpus storage unit for storing a dialogue corpus in which a dialogue state label value identified by a combination of the type of the user utterance and the type of the response sentence is stored for each user utterance in the utterance pair consisting of a sentence, and the utterance Based on the keyword stored in the state storage unit and the feature amount of the user utterance calculated based on the retrieved document data, the keyword included in the user utterance of the dialogue corpus, and the label value of the user utterance, the input A dialogue state estimation unit that determines a label value of a user utterance, a response sentence extraction unit that extracts a response sentence to a user utterance having the same label value as the determined label value from the dialogue corpus storage unit, and the extraction A correction location determination unit for determining a correction location of each response sentence, and a position at each of the determined correction locations. The candidate word is selected from the words included in at least one of the dialogue corpus and the document data of the utterance state storage unit, and the correction candidate of the extracted response sentence is selected as the selected word. A response candidate sentence creation unit that creates a response candidate sentence in which the replacement candidate words are replaced, and of the created response candidate sentences, a response candidate sentence that includes the most keywords stored in the utterance state storage unit is selected. And a response sentence output unit that outputs the selected response candidate sentence.

  According to this configuration, since the dialog state estimation unit determines the label value of the dialog state of the user utterance, and selects a response candidate sentence for the user utterance from the dialog corpus using this label value as a key, a complicated rule is used. A response sentence can be generated without it. In addition, the response candidate sentence is replaced with a dialogue corpus or a word included in the retrieved document data, and the response candidate sentence including the most keywords stored in the utterance state storage unit is selected from the replaced response candidate sentences. Therefore, a more appropriate sentence can be generated as a response sentence of the user utterance.

  Further, the response sentence extraction unit of the response sentence generation device of the present invention searches the dialog corpus for a user utterance having the same label value as the determined label value and a response sentence for the user utterance, A similarity between the input user utterance and the searched user utterance is calculated, a predetermined number is selected from the searched user utterances having the highest similarity, and the selected predetermined number of user utterances The response sentence to is extracted from the dialogue corpus.

  According to this configuration, it is possible to narrow down the number of extracted response sentences according to the arithmetic processing capability of the response sentence generation device. Further, since the response sentence generation device extracts various response sentences from the dialogue corpus, it can generate a more appropriate response sentence.

  The feature amount of the user utterance in the response sentence extraction unit of the response sentence generation device of the present invention includes a keyword stored in the utterance state storage unit, an appearance count of the keyword in the utterance state storage unit, and the search It is assumed that it is described by a combination with the number of appearances of the keyword in the document data.

  According to this configuration, it is possible to easily describe the feature amount of the user utterance.

  The correction location determination unit of the response sentence generation device of the present invention determines the position of the word whose appearance frequency in the dialogue corpus is lower than a predetermined threshold among the words included in the extracted response sentence. It was set as the structure determined as.

  According to this configuration, among the response sentences extracted from the dialogue corpus, a set of frequently used words (sentences corresponding to templates) remains, and other words (words that are not used frequently) are replaced and corrected. Therefore, the response sentence generation device can generate a more appropriate response sentence as the response sentence of the user utterance.

  The response sentence generation device of the present invention converts the voice data of the user utterance into text data, outputs the text data to the user utterance input unit, and uses the text data of the response sentence output from the response sentence output unit as voice data The audio processing unit for outputting is further provided.

  According to this configuration, the response sentence generation device can generate a response sentence even when a user utterance is received as voice data. Further, the response sentence generation device can output the generated response sentence as voice data.

  In the response sentence generation device of the present invention, the label value of the dialog state includes whether the user utterance and the response sentence are question sentences, and if they are question sentences, a YES / NO question sentence and a WH question The structure is determined depending on whether the sentence is classified into a sentence, an enumerated question sentence that should list a plurality of answers, or a question sentence other than that.

  According to this configuration, the response sentence generation device can accurately classify the conversation state with a small number of classifications, and thus the processing load of response sentence generation can be reduced.

  According to the present invention, it is not necessary to create complicated rules when constructing an interactive system such as an interactive question answering system. That is, by implementing the present invention in a dialog system, the cost of constructing the dialog system can be reduced.

Hereinafter, the best mode for carrying out the present invention (hereinafter referred to as an embodiment) will be described with reference to the drawings.
FIG. 1 is a block diagram showing a configuration of a response sentence generating apparatus according to the present embodiment.
The response sentence generation apparatus 100 includes a CPU (Central Processing Unit) 10 that performs various arithmetic processes, a main memory 20 that is a storage unit used when the CPU 10 performs arithmetic processes, and an input / output interface that controls input and output of various data. 50 and a storage unit 130 for storing various programs and data. The storage unit 130 is configured by, for example, a hard disk device. The CPU 10 implements each function described later by reading various programs and data stored in the storage unit 130 into the main memory 20 and performing arithmetic processing.

  The storage unit 130 stores a response sentence generation program 101 for the CPU 10 to generate a response sentence based on the input user utterance, and an utterance state storage unit 131 and a document corpus that stores a document corpus in a predetermined area. A storage unit 132 and a dialogue corpus storage unit 133 that stores a dialogue corpus are provided.

  The document corpus is a collection of various document data such as newspaper articles. Further, the dialogue corpus is obtained by adding a dialogue action label (described later) of the utterance pair to a text transcribed of the user utterance and the response sentence (utterance pair). The utterance state storage unit 131 stores a keyword included in the input user utterance and document data including the keyword. The document data including the keyword is document data retrieved from the document corpus storage unit 132.

The response sentence generation program 101 includes a dialog state management module 141 and a response sentence generation module 158.
The dialog state management module 141 is a module for the CPU 10 to infer (determine) the dialog state of the user utterance from the input user utterance.
The response sentence generation module 158 is a module for the CPU 10 to generate a response sentence following the user utterance based on the input user utterance and the estimated dialog state of the user utterance.
Details of the function of each module and the data stored in each storage unit will be described in detail with reference to the block diagram of FIG.

  FIG. 2 is a block diagram showing the expanded function of the response sentence generating apparatus of FIG. The function of the response sentence generating device 100 will be described with reference to FIG.

  The response sentence generation device 100 receives a user utterance input unit 110 that receives an input of a user utterance, and a dialog state management unit 140 that determines a dialog state (dialog action label) of the user utterance from the input user utterance. A response sentence generation unit 150 that generates a response sentence following the user utterance, a response sentence output unit 160 that outputs the generated response sentence, and a storage unit 130.

<Dialogue state management department>
The dialog state management unit 140 is realized by the CPU 10 in FIG. 1 executing the dialog state management module 141. The response sentence generation unit 150 is realized by the CPU 10 executing the response sentence generation module 158.
Further, the user utterance input unit 110 and the response sentence output unit 160 are realized by the execution process of the response sentence generation program 101 by the input / output interface 50 and the CPU 10 of FIG. As described above, the storage unit 130 includes the utterance state storage unit 131, the document corpus storage unit 132, and the dialogue corpus storage unit 133.

  In the present embodiment, the user utterance input to the user utterance input unit 110 is text data, and is, for example, converted from voice data by a voice processing unit (not shown). In addition, the voice processing unit may be provided inside the response sentence generating apparatus 100 or may be connected to the outside. Furthermore, the response sentence generation device 100 is a voice input unit (microphone) that accepts input of a user utterance voice, or a voice output unit (speaker) that outputs the response sentence generated (selected) by the response sentence generation device 100 as a voice. Etc. may be provided.

  The dialogue state management unit 140 includes a keyword extraction unit 142, a keyword addition unit 143, a document search unit 144, an utterance state update unit 145, and a dialogue state estimation unit 146.

  The keyword extraction unit 142 extracts keywords included in the user utterance output from the user utterance input unit 110. The keyword can be extracted at this time by using a known technique.

  The keyword adding unit 143 adds the keyword extracted from the user utterance to the utterance state storage unit 131.

  The document search unit 144 reads a keyword stored in the utterance state storage unit 131 and searches the document corpus storage unit 132 for document data including the keyword.

  The utterance state update unit 145 outputs the retrieved document data to the utterance state storage unit 131 and updates the utterance state of the utterance state storage unit 131. Here, the utterance state (speech state information) stored in the utterance state storage unit 131 indicates a keyword extracted from a user utterance so far and document data including the keyword. This speech state is illustrated in Table 1.

  For example, in the utterance state shown in Table 1, the keywords “Yokohama”, “Sightseeing”, “Facility”, and “Teach” appear once in the user utterance, and the document corpus is stored as document data including these keywords. It is shown that the document data “A guide teaching tourist facilities in Yokohama has been dispatched to ○” is retrieved from the department 132. The document data retrieved from the document corpus storage unit 132 may indicate the number of appearances (score) of each keyword.

  Note that, for example, the user utterance input unit 110 receives an input of an utterance “a place that can be enjoyed by the family” as a user utterance next to the current user utterance, and the keyword extraction unit 142 uses “family” and “enjoy” as keywords. , The keyword adding unit 143 additionally registers the keywords “family” and “enjoyable” in the utterance state described above. Further, the document search unit 144 searches the document corpus storage unit 132 for document data including the keywords “Yokohama”, “sightseeing”, “facility”, “family”, and “enjoyable”. Then, the utterance state updating unit 145 updates the utterance state by replacing the document data registered in the utterance state with the retrieved document data.

The dialogue state estimation unit 146 estimates the dialogue state of the user utterance based on the utterance state stored in the utterance state storage unit 131, and assigns a label value (interaction action label) indicating the estimation result to the user utterance. For example, a support vector machine (VNVapnic. The Nature of Statistical Learning Theory, Springer 1995), which is a statistical classifier, is used for the estimation of the dialog state.
This support vector machine is based on a keyword included in a user utterance of a dialogue corpus and a dialogue action label attached to the user utterance, and a user utterance calculated from a keyword stored in the utterance state storage unit 131 and searched document data. It is assumed that the user has learned to output the dialogue action label of the user utterance when receiving the feature amount (denoted by vector).

Here, a method of creating a vector input to the support vector machine will be described. Here, as features of user utterances, (1) keywords included in user utterances, (2) the number of occurrences of each keyword in the utterance state (see Table 1), and (3) each keyword in the retrieved document data It is assumed that the number of appearances is described by a vector.
The description will be made assuming that the keywords registered as the utterance state in the utterance state storage unit 131 and the number of appearances of each keyword are as shown in Table 2 below.

  In addition, a number assigned to each keyword is determined. For example, assignment numbers are determined as shown in Table 3 below. This allocation number indicates the position where the numerical value of the number of appearances of each keyword is written in the vector. This information is set in advance by the administrator of the response sentence generating apparatus 100 and stored in the storage unit 130. Note that when creating a vector, a field for writing the number of appearances of each keyword in the utterance state and a field for writing the number of appearances in the retrieved document data are different. FIG. 3 is a diagram illustrating vectors input to the support vector machine of the present embodiment. For example, the number of appearances in the utterance state is written in the field indicated by reference numeral 701, and the number of appearances in the retrieved document data is written in the field indicated by reference numeral 702.

  The dialog state estimation unit 146 creates a vector to be input to the support vector machine based on these pieces of information (see FIG. 3). That is, for example, since the assigned number of the keyword “Yokohama” is “2” and the number of appearances is “3”, “3” is written in the second position from the left. Similarly, the number of appearances is written for other keywords. In addition, since keywords other than allocation numbers 2, 3, 5, and 8 do not appear, 0 is set. Similarly, the dialogue state estimation unit 146 also writes the number of times each keyword appears in the retrieved document data (see reference numeral 702). Note that the assignment number of each keyword in the reference numeral 702 is prepared separately from the assignment number (Table 3) of the number of appearances in the user utterance.

  The dialog state estimation unit 146 inputs the vector created in this way to the support vector machine, and estimates the dialog state of the user utterance. Then, the dialogue action label as the estimation result and the input user utterance are output to the response sentence generation unit 150.

  The dialogue action label provided by the dialogue state estimation unit 146 will be described. The dialogue action label is a label given to a pair of a user utterance and a response sentence (wizard utterance). In the present embodiment, the types of user utterances and response sentences are (1) whether the question is not a question, and (2) if the question is a question, depending on what kind of question, eight types shown in Table 4 below. And an interactive action label is assigned.

  In Table 4, UT indicates a user utterance, and WT indicates a wizard utterance (response sentence). The WH type question is a question regarding Who (who), Whoch (what), What (what), When (when), Where (where), and How (how). For example, if the user utterance is a YES / NO question and the wizard utterance is not a question, the dialogue action label is “1”.

Next, the dialogue corpus stored in the dialogue corpus storage unit 133 will be described. This dialogue corpus is obtained by adding the dialogue action label described above to text data in which an utterance pair consisting of a user utterance and a response sentence (wizard utterance) is transcribed. The dialog action label at this time is determined manually.
An example of an utterance pair included in the dialogue corpus is shown below.

UT: 5: I would like to stay at a hotel in Hakone and enjoy a holiday while relaxing in a hot spring, but the budget is 20000 yen per person, and I would like to stay if there is a good hotel within that range. WT :: Yes, do you have any hope for the hot springs?

  In the above example, since the wizard utterance is a question, the interactive action label of “1 to 4” is excluded from the candidates. Since the wizard utterance is a YES / NO question, a dialogue action label of “5” is given.

  This dialogue corpus divides the utterances in the dialogue corpus into phrase and word units using existing morphological analysis technology, and gives part of speech information to each word. For words whose part of speech is any of a noun, a verb, and an adjective, the number of appearances in the document corpus of the document corpus storage unit 132 is given. In addition, the extraction of specific expressions and the types of specific expressions using specific expression extraction technology (for example, Hideki Isozaki, Hideto Kazawa: Efficient Support Vector Classifiers for Named Entity Recognition, Proceedings of COLING-2002, pp.390-396, 2002) (For example, person name, place name, etc.). In addition, it is assumed that there are a sufficiently large number of utterance pairs in this dialogue corpus.

<Response sentence generator>
The response sentence generation unit 150 adds a response sentence extraction unit 151 that extracts a response candidate sentence from the dialogue corpus of the conversation corpus storage unit 133, modifies the extracted response candidate sentence, and selects one of the corrected response candidate sentences. And a response sentence correction unit 152 that selects one sentence.

  When the response sentence extraction unit 151 receives the user utterance and the dialogue action label of the user utterance from the dialogue state estimation unit 146, the response sentence extraction unit 151 selects a user utterance whose dialogue state is similar to the user utterance based on the information. Extracted from the dialogue corpus storage unit 133. Then, a response sentence (wizard utterance) for this user utterance is extracted as a response candidate sentence and transferred to the response sentence correction unit 152. That is, a response sentence suitable as a response sentence (wizard utterance) following the input user utterance is extracted from the dialogue corpus as a response candidate sentence. Details of the response sentence extraction unit 151 will be described later using a flowchart.

Here, the configuration of the response sentence correction unit 152 will be described with reference to FIG. 4 (see FIG. 2 as appropriate). FIG. 4 is a block diagram illustrating a configuration of the response sentence correction unit in FIG.
As shown in FIG. 4, the response sentence correction unit 152 determines a correction part determination unit 153 that determines a correction part in each response candidate sentence output from the response sentence extraction part 151, and determines a replacement candidate word in each correction part. A replacement candidate determination unit 154, a response candidate sentence creation unit 155 that creates a response candidate sentence in which the word of the corrected portion of each response candidate sentence is replaced with a replacement candidate word, and a score that gives a score to the created response candidate sentence The adding unit 156 and a response sentence selection unit 157 that selects a response candidate sentence with the highest score among the created response candidate sentences are configured.
Details of these components will be described later with reference to the flowchart of FIG.

<Processing procedure>
Next, the outline of the processing procedure of the response sentence generation device 100 will be described with reference to FIGS. FIG. 5 is a flowchart showing an outline of a processing procedure of the response sentence generation device of FIG.

First, the response sentence generator 100 (see FIG. 2) initializes the system (S401). Then, the user utterance input unit 110 receives an input of the user utterance (S402), and the dialogue state management unit 140 performs dialogue state management of the input user utterance (S403). That is, the dialogue state management unit 140 extracts a keyword included in the user utterance and searches for document data including the keyword. Next, the dialogue action label of the user utterance is given from the keyword and the document data including the keyword. Then, the response sentence generation unit 150 generates a response sentence (S404). That is, the response sentence generation unit 150 extracts a response candidate sentence from the dialogue corpus storage unit 133 based on the user utterance and the dialogue action label of the user utterance, and corrects the extracted response candidate sentence. Then, a score is given to the corrected response candidate sentence, and the response candidate sentence having the highest score is selected as the response sentence. Next, the response text output unit 160 outputs the selected response text (S405), and returns to S402. That is, the input of the next user utterance is received.
In this way, the response sentence generating apparatus 100 generates and outputs a response sentence for the input user utterance.

  Next, the dialog state management in S403 of FIG. 5 will be described in detail with reference to FIGS. FIG. 6 is a flowchart for explaining the details of the dialog state management in S403 of FIG.

  The keyword extraction unit 142 performs morphological analysis on the text data of the user utterance input from the user utterance input unit 110, and extracts keywords included in the text data (S501). For example, the keyword extraction unit 142 performs morphological analysis on the user utterance “Please tell me about tourist facilities in Yokohama” and extracts the keywords “Yokohama”, “sightseeing”, “facility”, and “teach”.

  Next, the keyword adding unit 143 adds a keyword included in the user utterance to the utterance state storage unit 131 (S502). For example, in the above-described example, the keywords “Yokohama”, “sightseeing”, “facility”, and “teaching” are additionally registered in the utterance state storage unit 131.

  Then, the document search unit 144 searches the document corpus storage unit 132 for document data including all keywords in the utterance state storage unit 131 (S503). For example, if the keywords “Yokohama”, “Sightseeing”, “Facility”, and “Teach” are registered in the utterance state storage unit 131, the document search unit 144 searches for document data including this keyword.

  The utterance state update unit 145 outputs the retrieved document data to the utterance state storage unit 131 (S504), and updates the utterance state.

  The dialogue state estimation unit 146 estimates the dialogue state of the user utterance based on the utterance state stored in the utterance state storage unit 131, and assigns a dialogue action label indicating the estimation result to the user utterance (S505). For example, “3” is assigned as a dialogue action label of the user utterance “Please tell me about sightseeing spots in Yokohama”. Then, the user utterance and the dialogue action label are transferred to the response sentence extraction unit 151.

  Next, the response sentence generation in S404 of FIG. 5 will be described in detail with reference to FIGS. FIG. 7 is a flowchart for explaining the details of the response sentence generation in S404 of FIG.

  First, when the response sentence extracting unit 151 receives the user utterance and the dialog action label of the user utterance from the dialog state estimating unit 146, the response sentence extracting unit 151 receives the most recent user utterance from the dialog corpus (dialog corpus storage unit 133) based on the information. A near user utterance is searched (S601).

  That is, first, the response sentence extraction unit 151 searches for a user utterance having the same label as the dialogue action label received from the dialogue state estimation unit 146. Then, the degree of similarity between the user utterance in the searched utterance pair and the user utterance received from the dialogue state estimation unit 146 is obtained, and N (predetermined number) user utterances are selected from those having a high degree of similarity, for example. In addition, you may make it select the user utterance here that the similarity exceeds a predetermined threshold value.

The similarity here may be a distance scale, for example. As this distance measure, a measure (such as a Vapnic95 kernel) that can receive two character strings and return a similarity can be used.
For example, the utterance “I am looking for a hot spring inn in Nara” and the utterance “I am looking for a hot spring inn in Kyoto” and the utterance “I am looking for a hot spring inn in Nara” The distance measure is smaller in the former pair than in the latter pair. Therefore, when the response sentence extraction unit 151 uses the distance scale described above as the similarity in S601, N user utterances are selected from those having the smallest dialog distance scale.
Then, the response sentence extraction unit 151 reads the response sentence for the user utterance from the dialogue corpus storage unit 133.

Next, the correction part determination unit 153 determines the correction part (P _i1 ,..., P _imi ) of the response sentence of the user utterance selected in S601 (S602). Then, the determined correction part of the response sentence is stored in the main memory 20 (see FIG. 1) or the like.

Here, in FIG.
m _i : Number of corrections in the i-th response sentence
P _ij : j-th corrected portion in the i-th response sentence i: utterance index j: index of the corrected portion.
First, the correction location determination unit 153 determines i by substituting i = 1.

Here, the correction location is determined by first obtaining the position of a word whose part of speech of the response sentence is a noun, verb, or adjective, and the number of occurrences (appearance frequency) of the word in the dialogue corpus is predetermined. The position of the word is lower than the threshold. If this word does not appear in the dialogue corpus, the number of appearances (appearance frequency) in the document corpus (document corpus storage unit 132) is used. That is, the correction location determination unit 153 leaves words (sentences) that are frequently used as response sentences, and sets words that are not frequently used as correction targets (replacement targets).
For example, when a response sentence “what kind of sightseeing facility are you looking for” is extracted from the dialogue corpus, if the number of appearances (frequency of appearance) of the word “tourism facility” is lower than a predetermined threshold, this word Is determined as a correction point. In other words, a set of frequently used words (sentences) such as “What are you looking for?” Should be left.

Next, the replacement candidate determination unit 154 selects replacement candidates (S _ij ¹ ,..., S _ij ^l ) of the correction portion (P _ij ) determined in S602 (S603). That is, the part of speech of the corrected part of the response sentence determined in S602, whether or not the word is a proper expression (a proper noun or a noun representing a number), and in the case of a proper expression, information on the type is used. From the keywords of the utterance information in the utterance state storage unit 131, the word having the most similar information is selected as a replacement candidate. Then, the replacement candidate word is stored in the main memory 20. First, the correction location determination unit 153 substitutes j = 1 and selects a replacement candidate.

Then, j <when a m _i (S604 of Yes), that is, when the i-th (m _i-th from the first) All corrected portion of the response sentence does not select a replacement candidate for substitution candidate determination section 154 Increments the value of j (S605) and executes the process of S603. When j = m _i (No in S604), that is, when the replacement candidate determination unit 154 selects replacement candidates for all the corrected portions of the i-th response sentence, the process proceeds to S611. In S611, when i <N (Yes in S611), that is, when there is a response sentence in which the correction part determination unit 153 has not yet determined a correction part among the N user utterance response sentences selected in S601. The value of i is incremented (S612), and the process of S602 is executed. When i = N (No in S611), that is, when all the corrected parts of the response sentences of the N user utterances selected in S601 are determined, the response candidate sentence creation unit 155 is stored in the main memory 20. The correction part of each response sentence and the replacement candidate word to be replaced with the correction part are read, and each response sentence is corrected according to the read contents (S621). In this way, the response candidate sentence creation unit 155 creates a list of response candidate sentences that are replaced with the words of the respective replacement candidates at the correction portions of the respective response sentences.

  Next, the score giving unit 156 gives a score to each response candidate sentence created by the response candidate sentence creating unit 155 (S622). The score at this time is determined by how many keywords (keywords registered in the utterance state storage unit 131) are included in each response candidate sentence. Then, the response candidate sentence with the highest score is output as a response sentence (S623). That is, the response sentence selection unit 157 selects the response candidate sentence having the highest score as a response sentence, and the response sentence output unit 160 outputs the selected response sentence. That is, the response sentence output unit 160 outputs a response sentence that includes more keywords included in the user utterances so far.

The response sentence generation device 100 generates and outputs a response sentence for the input user utterance by the procedure as described above.
As described above, the response sentence generation apparatus 100 according to the present embodiment can create and output a response sentence for a user utterance without using a back end having a complicated rule as in the prior art.

  The response sentence generation device 100 according to the present embodiment can be realized by the response sentence generation program 101 that executes the processing as described above, and this program can be stored in a computer-readable storage medium (CD-ROM or the like). It can be stored and provided. It is also possible to provide the program through a network such as the Internet.

It is a block diagram which shows the structure of the response sentence production | generation apparatus of this Embodiment. It is the block diagram which expanded and showed the function of the response sentence production | generation apparatus of FIG. It is the figure which illustrated the vector inputted into the support vector machine of this embodiment. It is a block diagram which shows the structure of the response sentence correction part of FIG. It is a flowchart which shows the outline | summary of the process sequence of the response sentence production | generation apparatus of FIG. 6 is a flowchart for explaining details of dialog state management in S403 of FIG. It is a flowchart explaining the detail of the response sentence production | generation of S404 of FIG.

Explanation of symbols

10 CPU (Central Processing Unit)
DESCRIPTION OF SYMBOLS 20 Main memory 50 Input / output interface 100 Response sentence production | generation apparatus 101 Response sentence production | generation program 110 User utterance input part 130 Storage part 131 Speech state storage part 132 Document corpus storage part 133 Dialogue corpus storage part 140 Dialogue state management part 141 Dialogue state management module 142 Keyword Extraction Unit 143 Keyword Addition Unit 144 Document Search Unit 145 Speech State Update Unit 146 Dialogue State Estimation Unit 150 Response Sentence Generation Unit 151 Response Sentence Extraction Unit 152 Response Sentence Correction Unit 153 Correction Location Determination Unit 154 Replacement Candidate Determination Unit 155 Response Candidate Sentence creation unit 156 Score assignment unit 157 Response statement selection unit 158 Response statement generation module 160 Response statement output unit

Claims (9)

A response sentence generation device for generating a response sentence for a user utterance,
A user utterance input unit that accepts input of the user utterance;
A keyword extraction unit that analyzes the input user utterance and extracts a keyword included in the user utterance;
An utterance state storage unit for storing the extracted keyword and document data including the keyword;
A keyword adding unit for adding the extracted keyword to the utterance state storage unit;
A document data retrieval unit that retrieves document data including the extracted keyword from a document corpus storage unit that stores a document corpus that is a set of predetermined document data;
Outputting the searched document data to the utterance state storage unit, and updating the document data in the utterance state storage unit;
Dialog corpus storage for storing a dialogue corpus in which a dialogue state label value identified by a combination of the type of the user utterance and the type of the response sentence is assigned to each user utterance in the utterance pair composed of the user utterance and the response sentence. And
(1) a keyword stored in the utterance state storage unit and a feature amount of the user utterance calculated based on the retrieved document data, and (2) a keyword included in the user utterance of the dialog corpus and a label of the user utterance A dialogue state estimation unit that determines a label value of the input user utterance based on the value and
A response sentence extraction unit that extracts a response sentence to a user utterance having the same label value as the determined label value from the dialogue corpus storage unit;
A correction location determination unit for determining a correction location of each extracted response sentence;
A replacement candidate determination unit that selects a replacement candidate word at each of the determined correction locations from words included in at least one of the document data of the utterance state storage unit and the dialogue corpus;
A response candidate sentence creating unit that creates a response candidate sentence by replacing the selected replacement candidate word at the correction location of each extracted response sentence;
Among the created response candidate sentences, a response sentence selection unit that selects a response candidate sentence that includes the most keywords stored in the utterance state storage unit;
A response sentence output unit for outputting the selected response candidate sentence;
A response sentence generation device comprising: The response sentence extraction unit
A user utterance having the same label value as the determined label value and a response to the user utterance are searched from the dialogue corpus, and the similarity between the input user utterance and the searched user utterance is determined. Calculating, selecting a predetermined number from the searched user utterances having the highest similarity, and extracting response sentences for the selected predetermined number of user utterances from the dialogue corpus. Item 4. The response sentence generation device according to Item 1. The feature amount of the user utterance is:
It is described by a combination of a keyword stored in the utterance state storage unit, the number of appearances of the keyword in the utterance state storage unit, and the number of appearances of the keyword in the retrieved document data. The response sentence generation device according to claim 1 or 2. The correction location determination unit
4. The position of a word whose appearance frequency in the dialogue corpus is lower than a predetermined threshold among the words included in the extracted response sentence is determined as the correction portion. The response sentence generation device according to any one of the above. A voice processing unit that converts voice data of the user utterance into text data, outputs the text data to the user utterance input unit, and outputs text data of the response sentence output from the response sentence output unit as voice data; The response sentence generation device according to claim 1, wherein: The label value of the conversation state is
Whether or not the user utterance and the response sentence are question sentences, and if they are question sentences, a YES / NO question sentence, a WH question sentence, an enumerated question sentence that should list a plurality of answers, and others The response sentence generation device according to any one of claims 1 to 5, wherein the answer sentence generation device is determined depending on which of the question sentences is classified. A response sentence generation method for generating a response sentence for a user utterance,
Dialog corpus storage for storing a dialogue corpus in which a dialogue state label value identified by a combination of the type of the user utterance and the type of the response sentence is assigned to each user utterance in the utterance pair composed of the user utterance and the response sentence. A response sentence generation device comprising a unit,
Receiving an input of the user utterance;
Analyzing the input user utterance and extracting a keyword included in the user utterance;
Adding the extracted keyword to an utterance state storage unit that stores the extracted keyword and document data including the keyword;
Retrieving document data including the extracted keyword from a document corpus storage unit that stores a document corpus that is a set of predetermined document data;
Outputting the retrieved document data to the utterance state storage unit and updating the document data in the utterance state storage unit;
(1) a feature amount of a user utterance calculated based on a keyword stored in the utterance state storage unit and the retrieved document data;
(2) determining a label value of the input user utterance based on a keyword included in the user utterance of the dialogue corpus and a label value of the user utterance;
Extracting a response sentence to a user utterance having the same label value as the determined label value from the dialogue corpus storage unit;
Determining a correction location of each extracted response sentence;
Selecting a replacement candidate word at each determined correction location from words included in at least one of the document data of the utterance state storage unit and the dialogue corpus;
Creating a response candidate sentence in which the selected replacement candidate word is replaced at the correction location of each extracted response sentence;
Selecting a response candidate sentence including the most keywords stored in the utterance state storage unit from among the created response candidate sentences;
Outputting the selected response candidate sentence;
The response sentence generation method characterized by performing.   A response sentence generation program for causing a computer which is the response sentence generation apparatus to execute the response sentence generation method according to claim 7.   A computer-readable storage medium storing the response sentence generating program according to claim 8.

Images