JP2008165718A - Intention determination device, intention determination method, and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To determine a user's intention of an input sentence with high precision. <P>SOLUTION: It is determined whether or not the intention of user's utterance is "question" (106) on the basis of whether or not a sentence of the user's utterance has an interrogative form. When it is determined that the intention of user's utterance is not "question", it is determined whether or not the intention of the user's utterance is "declarative" form (108) on the basis of whether or not the sentence of the user's utterance including a clause. When it is determined that the intention of the user's utterance is "declarative" form, the intention of the user's utterance is determined to be "declaration" or "declarative answer" (110) on the basis of whether or not the sentence of the user's utterance is interrogative form. Meanwhile, when it is determined that the intention of the user's utterance is not "declaration", the intention of the user's utterance is determined to be "Y/N answer" or "OK" on the basis of whether or not the sentence of the user's utterance is interrogative. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to an intention determination device, an intention determination method, and a program, and more particularly, to an intention determination device, an intention determination method, and a program for determining an intention of an input sentence from a user.

  Conventionally, it has an utterance intention extraction means for extracting utterance intention from the utterance analysis result, and when it is determined that the utterance intention is a question, a database search condition is created by referring to the utterance analysis result. An interactive device that searches a database and generates a determination sentence for determining whether or not the answer is correct when the utterance intention is determined to be a response is known (Patent Document 1). In this dialogue apparatus, it is determined whether or not the utterance intention is a question based on the sentence end expression.

Also, a method for determining utterance intention based on statistical means is known (Non-Patent Document 2). In this determination method, N-gram, VSM (Vector Space Model), and SVM (Support Vector Machine) are used for a morpheme string obtained by converting a specific word (proprietary noun, numerical value) in an utterance into a class label expressing a specific meaning. ) Is used for learning for each utterance intention, and is classified into 79 kinds of utterance intentions. Further, at the time of utterance intention determination, the determination accuracy is improved by using not only the utterance information but also the previous utterance intention information. Moreover, in the determination method by SVM, the determination method by one-per-class is used.
JP2003-173335 Masayuki Shiraki, Toshihiko Ito, Mitsuhiko Kai, Hiromasa Nakatani, “Examination of Statistical Intent Understanding Method in Natural Spoken Sentences”, Information Processing Society of Japan Research Report, 2004-SLP-50, 2004

However, in the technique described in Patent Document 1, the intentions are not necessarily the same even if the sentence end expressions are the same as in the following utterance examples 1 and 2, so the intention is erroneously determined. There is a problem that. In addition, since the determination depends on the sentence end expression, there is a problem that it cannot be applied to intention determination other than the question.
Utterance example 1: Well, what did you make (= What did you make?) → Question utterance example 2: He made a gratin → Statement Also, the above non-patent The technique described in Document 2 does not use the previous utterance directly, but uses the utterance intention estimated by the system as the utterance intention of the previous utterance. When the intention of the previous utterance is incorrect, there is a problem that the possibility of the utterance intention of the next utterance is also increased.

  In addition, in the determination method using SVM, a determination method using one-per-class that uses a class number of classifiers such as a certain class or another class is used. When doing this, distinguish between the two features of the intention that the morpheme length of Y / N answer and OK is short and the intention that the statement and the superficial feature are similar, such as a statement-type answer, as a statement feature. For this reason, there is a problem in that it is necessary to set a complicated determination condition and an accurate determination cannot be made.

  The present invention has been made to solve the above problems, and an object of the present invention is to provide an intention determination device, an intention determination method, and a program that can determine the intention of an input sentence from a user with high accuracy. And

  In order to achieve the above object, an intention determination apparatus according to a first invention includes an input unit for inputting an input sentence from a user, an analysis unit for analyzing the structure of the input sentence input by the input unit, Based on an analysis result by the analysis unit, a question determination unit that determines whether or not the intention of the input sentence is a question, and when the question determination unit determines that the intention is not a question, the analysis unit Based on the analysis result, it is determined that the intention of the input sentence is a statement system by a statement system determination unit that determines whether or not the intention of the input sentence is a statement system including a statement and a statement type answer, and the statement system determination unit In this case, based on the analysis result by the analysis means, the statement type answer determination means for determining whether the intention of the input sentence is the statement or the statement type answer, and the statement system determination means An affirmative / negative answer determination means for determining whether the intent of the input sentence is an affirmative / negative answer or an understanding based on the analysis result of the analysis means when it is determined that it is not a clear system; It is comprised including.

  An intention determination method according to a second aspect of the present invention inputs an input sentence from a user, analyzes the structure of the input sentence, and determines whether the intention of the input sentence is a question based on the analysis result. If it is determined that it is not the question, based on the analysis result, it is determined whether the intention of the input sentence is a statement system including a statement and a statement-type answer, When it is determined that it is a statement system, based on the analysis result, it is determined whether the intention of the input sentence is the statement or the statement type answer, and is determined not to be a statement system In this case, it is characterized in that it is determined whether the intention of the input sentence is an affirmative / negative answer or an understanding based on the analysis result.

  According to the first invention and the second invention, the structure of the input sentence from the input user is analyzed, and it is determined whether or not the intention of the input sentence is a question based on the analysis result.

  If it is determined that it is not a question, it is determined whether the intention of the input sentence is a statement system including a statement and a statement-type answer based on the analysis result, and the intention of the input sentence is a statement system Is determined, based on the analysis result, it is determined whether the intention of the input sentence is a statement or the statement-type answer. When it is determined that the intention of the input sentence is not a statement system, it is determined whether the intention of the input sentence is an affirmative / negative answer or an understanding based on the analysis result.

  Therefore, whether it is a question, whether it is a statement system, whether it is a statement or a statement-type answer, whether it is an affirmative-negative answer or an understanding By determining the intent of the input sentence by multi-stage determination with the above determination, it is possible to make a combination of simple determinations, so that the intention of the input sentence from the user can be determined with high accuracy.

  An intention determination apparatus according to a third invention is based on an input means for inputting an input sentence from a user, an analysis means for analyzing the structure of the input sentence input by the input means, and an analysis result by the analysis means. The determination of whether or not the intention of the input sentence is a question, and when the question determination means determines that the intention is not a question, the input based on the analysis result by the analysis means An answer system determining means for determining whether the intention of the sentence is an answer system including an affirmative negative answer and a statement type answer, and the analysis when the answer system determining means determines that the answer system is an answer system Based on the analysis result by the means, a statement type answer determination means for determining whether the intention of the input sentence is the positive / negative type answer or the statement type answer; If it is determined that there is no, on the basis of the analysis result by the analyzing means, the intent of the input sentence, is configured to include a determining acknowledgment determining unit whether any of statements and understanding.

  An intention determination method according to a fourth aspect of the present invention is to input an input sentence from a user, analyze the structure of the input sentence, and determine whether the intention of the input sentence is a question based on the analysis result. If it is determined that it is not the question, it is determined whether the intention of the input sentence is an answer system including an affirmative negative answer and a declarative answer based on the analysis result And when it is determined that the answer system is based on the analysis result, it is determined whether the intention of the input sentence is the positive / negative answer or the declarative answer, and the answer When it is determined that it is not a system, it is characterized in that it is determined whether the intention of the input sentence is a statement or an understanding based on the analysis result.

  According to the third and fourth inventions, the structure of the input sentence from the input user is analyzed, and it is determined whether or not the intention of the input sentence is a question based on the analysis result.

  When it is determined that it is not a question, it is determined whether or not the intention of the input sentence is an answer system including an affirmative negative answer and a declarative answer based on the analysis result. In this case, based on the analysis result, it is determined whether the intention of the input sentence is an affirmative / negative answer or the above-mentioned declarative answer.

  On the other hand, when it is determined that it is not an answer system, it is determined whether the intention of the input sentence is a statement or an understanding based on the analysis result.

  Therefore, whether it is a question, whether it is an answer system, whether it is an affirmative negative answer or a declarative answer, and whether it is a statement or comprehension By determining the intent of the input sentence by multi-stage determination with the above determination, it is possible to make a combination of simple determinations, so that the intention of the input sentence from the user can be determined with high accuracy.

  The statement system determination means according to the first invention compares the analysis result of the analysis means with the analysis result prepared in advance based on the analysis by the analysis means, and the intention of the input sentence is a statement type answer or statement, And whether the answer is an affirmative or negative answer or approval. Accordingly, for example, based on the structure of a statement-type sentence having a feature of including a clause, the determination is made by paying attention only to whether it is a statement-type sentence, so that a highly accurate determination can be made.

  The answer system determination means according to the second invention compares the analysis result of the analysis means with the analysis result prepared in advance based on the analysis by the analysis means, and the intention of the input sentence is a declarative answer or affirmative negative A type answer and whether it is a statement or an acknowledgment can be determined. As a result, for example, based on the structure of an answer sentence having a feature that the previous input sentence is in the form of a question sentence, the decision is made with a focus on only whether it is an answer sentence. it can.

  The assertion type answer determination means according to the first invention compares the analysis result of the analysis means with the analysis result prepared in advance based on the analysis by the analysis means, and the intention of the input sentence is Whether it is a statement can be determined. As a result, for example, based on the structure of a sentence-type answer sentence having the characteristic that the previous input sentence is in the form of a question sentence, it is determined by focusing only on whether it is a statement-type answer or statement. An accurate determination can be made.

  The affirmative / negative answer determination means according to the first invention compares the analysis result of the analysis means with an analysis result prepared in advance based on the analysis by the analysis means, and the intention of the input sentence is affirmative / negative It can be determined whether it is an answer or an understanding. Thus, for example, based on the structure of an affirmative / negative answer sentence having the feature that the previous input sentence is in the form of a question sentence, the determination is made by paying attention only to whether it is an affirmative / negative answer or an understanding. It can be determined with high accuracy.

  Further, the assertion type answer determination means according to the second invention compares the analysis result of the analysis means with the analysis result prepared in advance based on the analysis by the analysis means, and the intention of the input sentence is It can be determined whether the answer is an affirmative or negative answer. As a result, for example, based on the structure of a statement-type answer sentence that includes a clause, the decision is made with a focus on only a statement-type answer or an affirmative-negative answer. can do.

  The understanding determination means according to the second invention compares the analysis result of the analysis means with the analysis result prepared in advance based on the analysis by the analysis means, and the intention of the input sentence is either statement or consent Can be determined. Thereby, for example, based on the structure of a statement sentence having a characteristic of including a clause, the determination is made by paying attention only to whether the statement is a statement or an understanding, so a highly accurate determination can be made.

  The analysis unit can analyze the structure of the input sentence input by the input unit and a previous input sentence that is an input sentence input at least one before the input sentence. Thereby, since it determines based on the analysis result of the structure of an input sentence and a previous input sentence, it can be set as a highly accurate determination.

  The question determination means compares the analysis result of the analysis means with the analysis result prepared in advance based on the analysis by the analysis means to determine whether the intention of the input sentence is a question or otherwise. be able to. As a result, since the determination is made by paying attention only to whether or not the question is based on the structure of the question sentence having the feature of being in the question sentence format, the determination can be made with high accuracy.

  Said input means can input the input sentence which shows a user's utterance.

  According to a fifth aspect of the invention, there is provided a program based on input means for inputting an input sentence from a user, analysis means for analyzing a structure of the input sentence input by the input means, and an analysis result by the analysis means. , A question determination means for determining whether the intention of the input sentence is a question, and when the question determination means determines that the intention is not a question, based on the analysis result of the analysis means, A statement system determining means for determining whether the intention is a statement system including a statement and a statement-type answer; when the statement system determination means determines that the intention is a statement system, based on the analysis result by the analysis means Thus, the intent of the input sentence is determined to be either the statement or the statement-type answer by the statement-type answer determination means and the statement-type determination means, and is determined not to be a statement system. A program for causing the input sentence to function as an affirmative / negative answer determination means for determining whether the intention of the input sentence is an affirmative / negative answer or an understanding based on the analysis result of the analysis means. is there.

  According to a sixth aspect of the present invention, there is provided a program for inputting an input sentence from a user, an analysis means for analyzing a structure of the input sentence input by the input means, and an analysis result by the analysis means. Based on the result of analysis by the analysis means when the question determination means determines that the intention of the input sentence is not a question based on the question determination means for determining whether the intention of the input sentence is a question Answer system determination means for determining whether the intention of the sentence is an answer system including an affirmative negative answer and a declaration type answer, and when the answer system determination means determines that the answer system is an answer system, the analysis means Based on the analysis result, the statement type answer determination unit that determines whether the intention of the input sentence is the positive / negative type answer or the statement type answer, and the answer type determination unit Therefore, when it is determined that it is not an answer system, a program for functioning as an understanding determination unit that determines whether the intention of the input sentence is a statement or an understanding based on the analysis result by the analysis unit It is.

  As described above, according to the intention determination device, the intention determination method, and the program of the present invention, the determination of whether or not the question is, the determination of whether or not it is a statement system, and the statement and the declaration type answer By determining the intention of the input sentence, or by determining whether it is a question, by multi-step determination of whether it is any of the above and a determination of whether it is an affirmative / negative answer or approval And whether or not it is an answer system, a determination of whether it is an affirmative negative response or a declarative response, and a determination of whether it is a statement or an understanding. By determining the intention of the sentence, it is possible to obtain a combination of simple determinations, so that it is possible to determine the intention of the input sentence from the user with high accuracy.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. In the present embodiment, a case will be described in which the present invention is applied to an intention estimation apparatus that receives a speech from a user as input and executes predetermined processing to output a voice.

  As shown in FIG. 1, the intention estimation apparatus 10 according to the first embodiment includes a microphone that collects user utterances and generates an audio signal, and recognizes the user utterance based on the audio signal. An input device 12 that generates character string information, an output device 14 that is composed of a speaker and that outputs voice, is connected to the input device 12 and the output device 14, and estimates the intention of the user's utterance, And a computer 16 that executes predetermined processing.

  The computer 16 stores an intention estimation program for executing an intention estimation processing routine, which will be described later, and uses a general morpheme analyzer to analyze morpheme information on character string information indicating a user utterance that has been voice-recognized. A morpheme analysis unit 18 for performing the process, an intention estimation unit 20 for estimating the intention of the user utterance based on the result of the morpheme analysis, and executing a predetermined problem solving process such as a database search according to the estimated intention. And a problem solving processing unit for generating a response sentence for the user utterance.

  As illustrated in FIG. 2, the intention estimation unit 20 includes “declaration”, “question”, “declaration-type answer”, “Y / N answer”, “acceptance” And “other”. The classification of the intention of the user utterance is determined based on a move in the discourse analysis (a structural unit that bears a discourse function such as a question and an answer). When classifying user utterances as one of the above intentions, it is possible to design a classifier that best represents the linguistic features of a class in the classification task of classifying into a plurality of classes according to the linguistic features of the user utterances. It becomes important.

  When the intention of the utterance is “question”, for example, a sentence such as “What is the weather tomorrow?” Is used as a question and takes the form of a question sentence. On the other hand, in cases other than “question”, a plain text is generally used.

  In addition, when the intention of the utterance is “statement”, for example, a sentence having a certain meaning and a clause is included, such as a sentence “Today is sunny”.

  Also, when the intention of the utterance is “statement type answer”, for example, an answer sentence “Tomorrow is sunny” to the question “What is the weather tomorrow?” Becomes Yes, for the 5W1H type question, It is a sentence that answers a question in a format other than No, and has a feature of including a clause. In addition, when the intention of the utterance is “Y / N answer”, for example, the answer is “Yes” to the question “Is it fine tomorrow?” , Yes, No.

  In addition, when the intention of the utterance is “OK”, for example, a sentence that agrees with “Yes” or an OK sentence for the utterance of the other party “Today is sunny”. When the intention of the utterance is “other”, for example, the sentence corresponds to laughter about the utterance of the other party “I am a fine man”.

  As shown in FIG. 3, the intention estimation unit 20 receives a morphological analysis result as an input, and an utterance intention determination unit 24 that determines an intention of a user utterance, and a previous utterance holding that holds a morphological analysis result for the previous user utterance Part 26.

  The result of morpheme analysis is represented by a structural tree as shown in FIG. 4, and the morpheme prototype is expressed as a leaf and the part of speech as a clause. FIG. 4 represents the morphological analysis result of the previous utterance obtained from the previous utterance holding unit 26 and the morphological analysis result of the utterance as one structural tree. Each of the BOU and the EOU represents the utterance of the utterance. It is a symbol representing the beginning and end.

  Also, the utterance intention determination unit 24 uses a question determination unit 28A that determines whether or not the intention of the user utterance is a “question”, and a partial tree as shown in FIG. 5 for use in the determination of the question determination unit 28A. The first dictionary database 30A storing a table in which weights and weights correspond to each other, and whether the intention of the user utterance is “statement” indicating “statement” and “statement-type answer” or not is determined The statement system determination unit 28B, the second dictionary database 30B storing a table in which subtrees and weights used in the determination of the statement system determination unit 28B correspond to each other, and the intention of the user utterance is “statement” and “statement” The third dictionary database storing a statement type answer determination unit 28C for determining whether it is any one of “type answer” and a table in which subtrees and weights for use in the determination of the statement type answer determination unit 28C are associated with each other 30C and Y / N answer determination unit 28D for determining whether the intention of the user's utterance is either “Y / N answer” or “OK”, and a subtree for use in the determination of Y / N answer determination unit 28D And a fourth dictionary database 30D storing a table in which weights correspond to each other. Thus, the utterance intention determination unit 24 includes a multistage determination unit.

  The contents of the tables stored in the first dictionary database 30A to the fourth dictionary database are described in non-patent literature (Kudo et al. “Boosting Algorithm for Semi-Structured Text Classification”, Information Processing Society of Japan Report, Intelligence and Complex Research Group, 2004-ICS-135, pp.163-168). Specifically, with respect to the first dictionary database 30A, among the learning data that is the morphological analysis result of a question case collected in advance or a sentence that is not a question example and a sentence immediately before that sentence, The subtree for maximizing the expression gain (<t, y>) in the document in the case where +1 is set to +1 and the other cases are set to −1, and the weight thereof may be obtained in advance.

  Hereinafter, the intention estimation processing routine executed by the computer 16 will be described with reference to FIG. First, in step 100, it is determined whether or not character string information is input from the input device 12. A voice signal corresponding to the user utterance is generated by the input device 12, and the voice signal is voice-recognized. When the character string information is input to the computer 16, the process proceeds from step 100 to step 102, morphological analysis is performed on the input character string information as a structure analysis. Outputs morpheme information. Then, in step 104, based on the morphological analysis result of the user utterance and the morphological analysis result on the previous user utterance stored in the previous utterance holding unit 26, a structural tree is generated, and the tree represented by the structural tree Output structure data.

  Next, in step 106, it is determined whether or not the intention of the user utterance is “question” by referring to the tree structure data obtained in step 104 and the first dictionary database 30A. Based on whether or not the subtree stored in the first dictionary database 30A is included in the tree structure data indicating the morphological analysis result of the user utterance and the morphological analysis result of the previous user utterance, the user utterance and the previous A score representing the degree of similarity between the morphological analysis result of the sentence indicating the user utterance and the morphological analysis result of the sentence that is the “question” and the previous sentence is calculated. Specifically, the score is calculated by the linear sum of the weights of the subtrees included in the user's utterance, and if the obtained score has a positive sign, it is similar to the morphological analysis result of the sentence “question” The sentence indicating the user utterance is determined to be in the form of a question sentence, and the intention of the user utterance is determined to be “question”. In step 107, “question” is stored in the variable i, and the routine proceeds to step 122.

  On the other hand, if the sign of the obtained score in step 106 is negative, it is not similar to the morphological analysis result of the sentence “question”, and the sentence indicating the user utterance is not in the question sentence format. If it is determined that it is in the form of plain text and it is determined that the intention of the user utterance is not “question”, in step 108, the tree structure data obtained in step 104 and the second dictionary database 30B are stored. Referring to this, it is determined whether or not the intention of the user utterance is “statement”. The user utterance is indicated based on whether the subtree stored in the second dictionary database 30B is included in the tree structure data indicating the morphological analysis result of the user utterance and the morphological analysis result of the previous user utterance. A score representing the degree of similarity between the morphological analysis result of the sentence and the sentence indicating the previous user utterance, and the morphological analysis result of the sentence that is the "statement system" and the previous sentence is calculated, and the obtained score If the sign is positive, it is similar to the morphological analysis result of the sentence that is “statement system”, and it is determined that the sentence indicating the user utterance includes a clause, and the intention of the user utterance is “statement system”. It is determined that there is, and the process proceeds to step 110.

  In step 110, referring to the tree structure data obtained in step 104 and the third dictionary database 30C, it is determined whether the intention of the user utterance is one of “statement” and “statement type answer”. To do. The user utterance is indicated based on whether the subtree stored in the third dictionary database 30C is included in the tree structure data indicating the morphological analysis result of the user utterance and the morphological analysis result of the previous user utterance. The score obtained by calculating the score indicating the degree of similarity between the morphological analysis result of the sentence indicating the sentence and the previous user utterance, and the morphological analysis result of the sentence that is the “statement type answer” and the previous sentence If the sign of is positive, it is similar to the result of morphological analysis of a sentence that is a “statement type answer”, and it is determined that the sentence indicating the previous user utterance is a question type, and the intention of the user utterance is “statement type” In step 112, “statement type answer” is stored in the variable i, and the process proceeds to step 122. On the other hand, if the sign of the score calculated in step 110 is negative, it is not similar to the morphological analysis result of the sentence that is “statement type answer”, and it is determined that the sentence indicating the previous user utterance is not questionable. Then, it is determined that the intention of the user utterance is “statement”, and in step 114, “statement” is stored in the variable i, and the process proceeds to step 122.

  For example, as shown in FIG. 7A, when the previous user utterance is “how is the weather” and the user utterance is “this is rain”, the score is calculated based on the table shown in FIG. When calculated, a positive score is obtained (0.39 + 0.15 = 0.54).

  On the other hand, as shown in FIG. 7B, when the previous user utterance is “this is sunny” and the user utterance is “this is rain”, based on the table shown in FIG. When the score is calculated, a negative score is obtained (0.39-0.40 = -0.01).

  If it is determined in step 108 that the intention of the user utterance is not “statement”, in step 116, the tree structure data obtained in step 104 and the fourth dictionary database 30D are referred to. It is determined whether the intention of the user utterance is either “Y / N answer” or “OK”. The user utterance is shown based on whether or not the subtree stored in the fourth dictionary database 30D is included in the tree structure data indicating the morphological analysis result of the user utterance and the morphological analysis result of the previous user utterance. The score indicating the degree of similarity between the morphological analysis result of the sentence indicating the sentence and the previous user utterance, and the morphological analysis result of the sentence that is “Y / N answer” and the previous sentence was obtained. If the sign of the score is positive, it is similar to the morphological analysis result of the sentence “Y / N answer”, and the sentence indicating the previous user utterance is determined to be questionable, and the intention of the user utterance is “ It is determined that the answer is “Y / N answer”. In step 118, “Y / N answer” is stored in the variable i, and the process proceeds to step 122. On the other hand, if the sign of the score calculated in step 116 is negative, it is determined that the sentence indicating “Y / N answer” is not similar to the morphological analysis result and the sentence indicating the previous user utterance is not questionable. Then, it is determined that the intention of the user utterance is “OK”, and “OK” is stored in the variable i in step 120, and the process proceeds to step 122.

  In step 122, the morphological analysis result of the utterance generated in step 102 is stored in the previous utterance holding unit 26 as a previous utterance. In step 124, a variable i indicating the estimated intention is output. The estimation processing routine is terminated.

  As described above, in the intention estimation processing routine, the determination condition of each determination process is based on the surface features of the utterance intention to be determined, and the identification boundary that characterizes the difference in intention is simple. In the processing, the intention of the speech can be determined with high accuracy, and the user's speech can be estimated with high accuracy by executing these determination processing in multiple stages.

  Then, according to the user utterance intention output in the intention estimation processing routine, a predetermined process is executed in the problem solving processing unit 22, and a response sentence indicating the execution result is generated. Further, the output device 14 synthesizes the generated response sentence with speech and outputs the speech.

  For example, as shown in FIG. 9, when the intention estimation device 10 is an airline reservation system, the voice output from the intention estimation device 10 is “the time is full. Do you want to change flights?” If the user utterance is “yes”, the intention estimation device 10 estimates that the intention of the user utterance is “Y / N answer”, executes a search process for vacant flights on the database, and searches for vacant flights. Then, a response sentence representing the searched vacant flight is generated, and the intention estimation device 10 outputs a voice message “Changed to 24 flights”.

  Further, if the user utterance for the voice output from the intention estimation device 10 is “I want to change the departure time”, the intention estimation device 10 estimates that the intention of the user utterance is a “statement type answer”, and requests for another flight The processing is executed, a response sentence for requesting another flight is generated, and a voice message “Please tell me the departure time” is output from the intention estimation device 10.

  As described above, according to the intention estimation device according to the first embodiment, the determination as to whether or not the intention of the user utterance is a “question”, and the determination as to whether or not it is a “declaration system” , The intention of the user's utterance is determined by multi-stage determination of whether it is “statement” or “statement type answer” and whether it is “Y / N answer” or “acceptance”. Since the estimation can be a combination of simple determinations, the intention of the user utterance can be estimated with high accuracy.

  In addition, for example, based on the structure of a sentence that is a “question” having the characteristic of being in the form of a question sentence, the determination is made by focusing only on whether or not the intention of the user utterance is a “question”. It can be determined. In addition, for example, based on the structure of a sentence that is a “statement” that has a feature of including a clause, it is determined only by whether or not it is a “statement”, so that a highly accurate determination can be made. .

  Also, for example, based on the structure of a sentence that is a “statement-type answer” that has the characteristic that the previous user utterance is in the form of a question sentence, focus only on whether it is a “statement-type answer” or “statement” Therefore, the determination can be made with high accuracy. Also, for example, based on the structure of a sentence that is a “Y / N answer” having the feature that the previous user utterance is in the form of a question sentence, it is only one of “Y / N answer” and “OK”. Therefore, the determination can be made with high accuracy.

Also, by determining the intention of the user utterance using the information of the previous user utterance held by the previous utterance holding unit, even in the same utterance as in the following user utterance examples 1 to 4, It can be determined that the intention is different depending on the user utterance.
(User utterance example 1)
A: How is the weather?
B: This is rain. → "Statement type answer"
(User utterance example 2)
A: Tokyo is sunny.
B: This is rain. → "Statement"
(User utterance example 3)
A: Fukui is delicious, B: Yeah. → "OK"
(User utterance example 4)
A: Going to eat salmon?
B: Yeah. → "Y / N answer"
In addition, the intention of the previous user utterance estimated last time is not used for estimation of the intention of the user utterance, but the intention of the user utterance is directly determined from the morphological analysis result of the previous user utterance and the morphological analysis result of the user utterance. Since the estimation is performed, it is possible to prevent a chain of errors from occurring.

  In the above-described embodiment, the case where voice output by a speaker is performed has been described as an example. However, the present invention is not limited to this, and a response sentence may be displayed on a display.

  Moreover, although the case where the voice is input from the user to the microphone has been described as an example, the user may input text as an input sentence using a keyboard or the like.

  Moreover, although the case where the structural tree is generated based on the result of the morphological analysis has been described as an example, by introducing the dependency analysis processing after the morphological analysis, the structural tree is changed to a dependency tree as shown in FIG. You may make it express. In this case, the dependency (modification relationship) of each clause is analyzed based on the result of the morphological analysis, and the structural tree is generated based on the result of the dependency analysis. For example, as shown in FIG. 10, a dependency tree indicating that “weather is” is related to “how is” and “here is” and “violent” is related to “falling” is generated.

  Further, in all the determination processes, the case where the intention of the user utterance is determined using the morphological analysis result of the previous user utterance and the morphological analysis result of the user utterance has been described as an example. Depending on the intention of the user utterance to be made, the determination may be made using a structural tree indicating only the morphological analysis result of the user utterance. For example, in the determination of whether or not it is a “question”, since the contribution of what the previous user utterance is is not so great, only the determination process of whether or not it is a “question” for the sake of processing efficiency, The determination is made without using the information of the previous user utterance. In this case, tree structure data that does not include the previous user utterance is generated at the time of generating the tree structure, and the tree structure data of the previous user utterance may be added in the subsequent determination process when it is determined not to be a “question”. .

  In addition, although the case where the determination process is performed using the morphological analysis result of the previous user utterance and the morphological analysis result of the user utterance has been described as an example, it is not limited thereto, and the context information is more abundant. In order to use, the determination process may be performed using the previous user utterance in addition to the previous user utterance. In this case, a subtree representing the previous user utterance may be added to the root node of the structural tree.

  Moreover, although the case where the intention of the user utterance is classified into any of “question”, “statement answer”, “statement”, “Y / N answer”, and “acceptance” has been described as an example, a filler whose intention cannot be recognized Since tweets and tweets may be input as user utterances, a classification of “others” may be added. In this case, as a dictionary database for determining “others”, similar to other intentions, a sentence other than “other” cases or “other” cases collected in advance and a sentence immediately before that sentence are used. In the learning data that is the result of morphological analysis, a subtree that maximizes the gain (<t, y>) and its weight when the “other” case is +1 and the other cases are −1 is obtained and obtained. The subtree and its weight may be stored in the dictionary database. Then, a determination is made using this dictionary database. In the intention estimation processing routine, “other” is determined between the generation of the tree structure data (step 104) and the question determination processing (step 106). If the user utterance is “other”, the variable i is set to “ After substituting “others”, the process from step 122 is performed. If not “others”, the process from step 106 is executed.

  Next, a second embodiment will be described. In addition, about the part which has the structure similar to 1st Embodiment, the same code | symbol is attached | subjected and description is abbreviate | omitted.

  In the second embodiment, the content of the subsequent determination process when it is determined that the intention of the user utterance is not a question is different from the first embodiment.

  As illustrated in FIG. 11, the utterance intention determination unit 224 of the intention estimation unit 220 of the intention estimation apparatus according to the second embodiment is configured to determine whether the intention of the user utterance is a “question”. 28A, the first dictionary database 30A for use in the determination by the question determination unit 28A, and the intention of the user utterance is “answer system” indicating “Y / N answer” and “statement type answer”, An answer system determination unit 228B that determines whether or not the second dictionary database 230B that stores a table in which a partial tree and a weight for use in the determination of the answer system determination unit 228B are stored, and the intention of the user utterance is A Y / N answer determination unit 228C that determines whether the answer is “Y / N answer” or “statement type answer”, and a subtree and a weight to be used in the determination of the Y / N answer determination unit 228C. Store the corresponding table A third dictionary database 230C, an approval determination unit 228D for determining whether the intention of the user's utterance is “statement” or “acceptance”, and a subtree for use in the determination of the approval determination unit 228D The fourth dictionary database 230D stores a table corresponding to the weights.

  Next, an intention estimation processing routine according to the second embodiment will be described with reference to FIG. In addition, about the process similar to 1st Embodiment, the same code | symbol is attached | subjected and detailed description is abbreviate | omitted.

  First, in step 100, it is determined whether or not character string information is input from the input device 12, and when character string information is input from the input device 12 to the computer 16 as a result of speech recognition for a user utterance, step 102 is performed. Then, morphological analysis is performed on the input character string information. In step 104, a structure tree is generated based on the analysis result, and tree structure data represented by the structure tree is output.

  Next, in step 106, it is determined whether or not the intention of the user utterance is “question”. When it is determined that the intention of the user utterance is “question”, in step 107, the variable i is set to “question”. Is stored and the process proceeds to step 122.

  On the other hand, if it is determined in step 106 that the intention of the user utterance is not “question”, in step 208, the tree structure data obtained in step 104 and the second dictionary database 230B are referred to. Then, it is determined whether or not the intention of the user utterance is “answer system”. Based on whether the subtree stored in the second dictionary database 230B is included in the tree structure data indicating the morphological analysis result of the user utterance and the morphological analysis result of the previous user utterance, the user utterance and the previous A score indicating the degree of similarity between the morphological analysis result of the sentence indicating the user's utterance and the morphological analysis result of the sentence that is the “answer system” and the previous sentence is calculated, and the sign of the obtained score is correct. If so, it is similar to the morphological analysis result of the sentence that is “answer system”, and it is determined that the sentence indicating the previous user utterance is a question form and the intention of the user utterance is “answer system” Then, the process proceeds to step 210.

  In step 210, referring to the tree structure data obtained in step 104 and the third dictionary database 230C, the intention of the user utterance is either “Y / N answer” or “statement type answer”. Determine whether. Based on whether or not the subtree stored in the third dictionary database 30C is included in the tree structure data indicating the morphological analysis result of the user utterance and the morphological analysis result of the previous user utterance, the user utterance and the previous A score indicating the degree of similarity between the morphological analysis result of the sentence indicating the user's utterance, the sentence that is the "statement type answer" and the morphological analysis result of the previous sentence is calculated, and the sign of the obtained score is If it is negative, it is determined that the sentence indicating “user's utterance” does not resemble the morphological analysis result of the sentence, and the sentence indicating the user utterance does not contain a clause, and the intention of the user utterance is “Y / N In step 212, “Y / N answer” is stored in the variable i, and the process proceeds to step 122. On the other hand, if the sign of the score calculated in step 210 is positive, it is similar to the morphological analysis result of the sentence that is a “statement type answer”, and it is determined that the sentence indicating the user utterance includes a clause. Then, it is determined that the intention of the user utterance is “statement type answer”, and in step 214, “statement type answer” is stored in the variable i, and the process proceeds to step 122.

  In the above step 208, if the sign of the obtained score is negative, it is determined that the sentence that is “answer system” is not similar to the morphological analysis result, and the sentence indicating the previous user utterance is not questionable, If it is determined that the intention of the user utterance is not “answer system”, the intention of the user utterance is “step 216” with reference to the tree structure data obtained in step 104 and the fourth dictionary database 230D. It is determined whether it is one of “statement” and “acceptance”. Based on whether the subtree stored in the fourth dictionary database 230D is included in the tree structure data indicating the morphological analysis result of the user utterance and the morphological analysis result of the previous user utterance, the user utterance and the previous A score indicating the degree of similarity between the morphological analysis result of the sentence indicating the user's utterance, the sentence that is “statement”, and the morphological analysis result of the previous sentence is calculated, and the sign of the obtained score is positive. If so, it is similar to the morphological analysis result of the sentence “statement”, determines that the sentence indicating the user utterance includes a clause, determines that the intention of the user utterance is “statement”, and In step 218, “statement” is stored in the variable i, and the routine proceeds to step 122. On the other hand, if the sign of the score calculated in step 216 is negative, it is determined that the sentence that is “statement” is not similar to the morphological analysis result and the sentence indicating the user utterance does not include a clause, It is determined that the intention of the user utterance is “OK”, “OK” is stored in the variable i in Step 219, and the process proceeds to Step 122.

  In step 122, the morphological analysis result of the utterance generated in step 102 is stored in the previous utterance holding unit 26 as a previous utterance. In step 124, a variable i indicating the intention of the user utterance is output. The estimation processing routine is terminated.

  As described above, according to the intention estimation device according to the second embodiment, the determination as to whether the intention of the user utterance is a “question” and the determination as to whether it is a “answer system” , The intention of the user utterance is determined by multi-stage determination of whether the answer is “Y / N answer” or “statement type answer” and whether it is “statement” or “acceptance”. Since the determination can be a simple combination of determinations, the intention of the user utterance can be determined with high accuracy.

  Also, for example, based on the structure of a sentence that is an “answer system” with the characteristic that the previous user utterance is in the form of a question sentence, paying attention only to whether or not the intention of the user utterance is “answer system” Since it determines, it can be set as highly accurate determination. For example, based on the structure of a sentence that is a “declaration-type answer” having a feature that the user utterance includes a clause, paying attention to only one of the “declaration-type answer” and the “Y / N answer” Since it determines, it can be set as highly accurate determination.

  In addition, for example, based on the structure of a sentence that is a “statement” having a feature that a user utterance includes a clause, the determination is made by focusing only on whether it is “statement” or “acceptance”. It can be determined.

  Next, a third embodiment will be described. In addition, since the structure of the intention estimation apparatus which concerns on 3rd Embodiment is the structure similar to 1st Embodiment, it attaches | subjects the same code | symbol and abbreviate | omits description.

  The third embodiment is different from the first embodiment in that the tree structure data indicating the morphological analysis result indicates not only morphological information but also information on other linguistic features.

  As shown in FIG. 13, tree structure data in which not only morphological information but also other linguistic features are added to the root node is used for the structure tree indicating the morphological analysis result. For example, “OK” or “Y / N response” has a feature that the morpheme length is short, and “OK” has a rough feature that it does not include a predicate, and therefore a part indicating whether the morpheme length is equal to or less than a threshold value. A subtree indicating whether or not a tree or predicate is included is added to the root node.

  In the intention estimation processing routine, when determining whether or not the intention of the user utterance is “declaration system”, the score is calculated with reference to the second dictionary database 30B, and the morpheme length is referred to. If the score is negative and the morpheme length is short, it is determined that the statement is not “statement”.

  Further, when determining whether the intention of the user utterance is any one of “Y / N answer” and “OK”, the score is calculated with reference to the fourth dictionary database 30D, With reference to the presence or absence of the predicate, if the score is negative and the predicate is not included, it is determined as “OK”.

  In this way, by using information indicating characteristics such as the morpheme length and the presence / absence of a predicate together with the score calculated by referring to the dictionary database, the user utterance intention determination process is performed, so that the user utterance can be more accurately recorded. Intent can be estimated.

It is the schematic which shows the structure of the intention estimation apparatus which concerns on the 1st Embodiment of this invention. It is a table | surface which shows the classification | category of the intention of a user utterance, a role, and the example of utterance. It is a block diagram which shows the structure of the intention estimation part of the intention estimation apparatus which concerns on the 1st Embodiment of this invention. It is an image figure which shows the content of tree structure data. It is an image figure which shows the content of the dictionary database used when determining whether the intention of a user utterance is a "question". It is a flowchart which shows the content of the intention estimation process routine of the intention estimation apparatus which concerns on the 1st Embodiment of this invention. (A) Image diagram showing contents of tree structure data when intention of user utterance is “statement type answer”, and (B) Image diagram showing contents of tree structure data when intention of user utterance is “statement” It is. It is an image figure which shows the content of the dictionary database used when determining whether the intention of a user utterance is a "statement type answer". It is an image figure which shows a process in case an intention estimation apparatus is an airline reservation system. It is an image figure which shows the content of the tree structure data produced | generated based on the result of dependency analysis. It is a block diagram which shows the structure of the intention estimation part of the intention estimation apparatus which concerns on the 2nd Embodiment of this invention. It is a flowchart which shows the content of the intention estimation process routine of the intention estimation apparatus which concerns on the 2nd Embodiment of this invention. It is an image figure which shows the content of the tree structure data which added the partial tree which shows the presence or absence of a morpheme length and a predicate.

Explanation of symbols

10 intention estimation device 12 input device 14 output device 16 computer 18 morpheme analysis unit 20 intention estimation unit 22 problem solving processing unit 24 utterance intention determination unit 26 previous utterance holding unit 28A question determination unit 28B statement system determination unit 28C statement type answer determination unit 28D Y / N answer determination unit 30A, 30B, 30C, 30D, 230B, 230C, 230D Dictionary database 220 Intention estimation unit 224 Utterance intention determination unit 228B Answer system determination unit 228C Y / N answer determination unit 228D Comprehension determination unit

Claims (15)

An input means for inputting an input sentence from the user;
Analyzing means for analyzing the structure of the input sentence input by the input means;
Question determination means for determining whether the intention of the input sentence is a question based on the analysis result by the analysis means;
A statement for determining whether the intention of the input sentence is a statement system including a statement and a statement type answer based on the analysis result by the analysis unit when the question determination unit determines that the question is not a question System determination means;
When the statement system determination unit determines that it is a statement system, it determines whether the intention of the input sentence is either the statement or the statement type answer based on the analysis result by the analysis unit A statement-type answer determination means;
Affirmation to determine whether the intention of the input sentence is an affirmative negative answer or an understanding based on the analysis result by the analysis unit when it is determined that the statement system determination unit is not a statement system A negative answer determination means;
An intention determination device including: An input means for inputting an input sentence from the user;
Analyzing means for analyzing the structure of the input sentence input by the input means;
Question determination means for determining whether the intention of the input sentence is a question based on the analysis result by the analysis means;
Whether the intent of the input sentence is an answer system including an affirmative negative answer and a declarative answer based on the analysis result by the analysis means when the question determination means determines that the question is not a question An answer type determination means for determining;
Whether the intention of the input sentence is either the positive / negative answer or the assertion type answer based on the analysis result by the analysis means when the answer type determination means determines that the answer type is A statement type answer determination means for determining
An understanding determination means for determining whether the intention of the input sentence is a statement or an understanding based on an analysis result by the analysis means when the answer system determination means determines that it is not an answer system; ,
An intention determination device including:   The statement system determination means compares the analysis result of the analysis means with the analysis result prepared in advance based on the analysis by the analysis means, and the intention of the input sentence is the statement type answer or the statement, The intention determination apparatus according to claim 1, wherein the intention determination apparatus determines whether the answer is an affirmative / negative answer or the approval.   The answer system determination means compares the analysis result of the analysis means with an analysis result prepared in advance based on the analysis by the analysis means, and the intention of the input sentence is determined to be the declarative answer or the positive / negative The intention determination apparatus according to claim 2, wherein the intention determination apparatus determines whether the answer is a type answer and the statement or the understanding.   The assertion type answer determination unit compares the analysis result of the analysis unit with an analysis result prepared in advance based on the analysis by the analysis unit, and the intention of the input sentence is determined by the declaration type answer and the statement The intention determination apparatus according to claim 1, wherein the intention determination apparatus determines whether the determination is any one.   The affirmative / negative answer determination means compares the analysis result of the analysis means with an analysis result prepared in advance based on the analysis by the analysis means, and the intention of the input sentence is the positive / negative answer and the The intention determination apparatus according to claim 1, wherein the intention determination apparatus determines whether it is an understanding.   The assertion type answer determination unit compares the analysis result of the analysis unit with an analysis result prepared in advance based on the analysis by the analysis unit, and the intention of the input sentence is the declaration type answer and the positive / negative The intention determination apparatus according to claim 2 or 4, wherein the intention determination apparatus determines whether the answer is a type answer.   The understanding determination means compares the analysis result of the analysis means with an analysis result prepared in advance based on the analysis by the analysis means, and the intention of the input sentence is either the statement or the understanding. The intention determination device according to claim 2, 4, or 7.   9. The analysis unit according to claim 1, wherein the analysis unit analyzes a structure of the input sentence input by the input unit and a previous input sentence that is an input sentence input at least one before the input sentence. The intention determination device according to claim 1.   The question determination means compares the analysis result of the analysis means with an analysis result prepared in advance based on the analysis by the analysis means, and determines whether the intention of the input sentence is a question or otherwise. The intention determination device according to any one of claims 1 to 9, wherein the determination is made.   The intention determination apparatus according to claim 1, wherein the input unit inputs an input sentence indicating a user's utterance. Enter the input sentence from the user,
Analyzing the structure of the input sentence,
Based on the analysis result, determine whether the intention of the input sentence is a question,
When it is determined that it is not the question, based on the analysis result, it is determined whether the intention of the input sentence is a statement system including a statement and a statement type answer,
When it is determined that the statement system is based on the analysis result, it is determined whether the intention of the input sentence is the statement or the statement type answer,
An intention determination method comprising: determining whether the intention of the input sentence is an affirmative negative answer or an understanding based on the analysis result when it is determined that the statement is not a statement system. Enter the input sentence from the user,
Analyzing the structure of the input sentence,
Based on the analysis result, determine whether the intention of the input sentence is a question,
When it is determined that it is not the question, based on the analysis result, it is determined whether the intention of the input sentence is an answer system including an affirmative negative answer and a assertion answer,
When it is determined that the answer system, based on the analysis result, it is determined whether the intention of the input sentence is the positive / negative answer or the assertion type answer,
An intention determination method comprising: determining whether the intention of the input sentence is a statement or an understanding based on the analysis result when it is determined that the answer system is not used. Computer
An input means for inputting an input sentence from the user,
Analyzing means for analyzing the structure of the input sentence input by the input means;
Question determination means for determining whether the intention of the input sentence is a question based on the analysis result by the analysis means;
A statement for determining whether the intention of the input sentence is a statement system including a statement and a statement type answer based on the analysis result by the analysis unit when the question determination unit determines that the question is not a question System determination means,
When the statement system determination unit determines that it is a statement system, it determines whether the intention of the input sentence is either the statement or the statement type answer based on the analysis result by the analysis unit A statement type answer determination unit, and when the statement type determination unit determines that the statement is not a statement type, the intention of the input sentence is either an affirmative negative type answer or an understanding based on an analysis result by the analysis unit A program for functioning as an affirmative / negative answer determination means for determining whether or not Computer
An input means for inputting an input sentence from the user,
Analyzing means for analyzing the structure of the input sentence input by the input means;
Question determination means for determining whether the intention of the input sentence is a question based on the analysis result by the analysis means;
Whether the intent of the input sentence is an answer system including an affirmative negative answer and a declarative answer based on the analysis result by the analysis means when the question determination means determines that the question is not a question Answering system judging means for judging,
Whether the intention of the input sentence is either the positive / negative answer or the assertion type answer based on the analysis result by the analysis means when the answer type determination means determines that the answer type is A statement-type answer determination unit that determines whether the input sentence is intended to be a statement or an understanding based on the analysis result of the analysis unit A program for functioning as an understanding determination means for determining whether or not.

Images