JP6755633B2 - Message judgment device, message judgment method and program - Google Patents

Description

The present invention relates to a message determining device, a message determining method, and a program for determining whether or not an utterance in a dialogue by a plurality of speakers is a message utterance indicating a message of the dialogue.

Detecting utterances (message utterances) that indicate the content that the customer wants to inquire of the receptionist from the dialogue between the customer and the person in charge at the call center or window clarifies the content of the inquiry. It is useful for improving the efficiency of responses by the person in charge of response and for realizing automatic response by the system to the request.

As a method of determining whether or not an utterance in a dialogue by multiple speakers, such as a dialogue between a customer and a person in charge of reception, is a message utterance, information on whether or not the utterance in the dialogue is a message utterance is added. There is a method of using the learned learning data (message utterance learning data) (see Non-Patent Document 1). In this method, learning data to which information on whether or not the utterance in the dialogue is a message utterance is added is used. Then, by machine learning using the learning data, a message determination model for determining whether or not the utterance in the dialogue is a message utterance is generated.

R.-E. Fan, K.-W. Chang, C.-J. Hsieh, X.-R. Wang, and C.-J. Lin. LIBLINEAR: A library for large linear classification Journal of Machine Learning Research 9 (2008), 1871-1874.

In general, there are differences in the expression of message utterances in each field. Therefore, in the method disclosed in Non-Patent Document 1, when the message determination model generated by using the message utterance learning data in a certain field is applied to another field, whether or not the utterance is a message utterance. It may not be possible to determine with high accuracy. It is conceivable to prepare the message utterance learning data for each field in which the determination is to be performed, but preparing the message utterance learning data for each field causes an increase in cost.

An object of the present invention made in view of the above problems is a message determination device and a message capable of determining whether or not an utterance in a dialogue is a message utterance while suppressing an increase in cost. The purpose is to provide a determination method and a program.

In order to solve the above problem, the message determination device according to the present invention is a message determination device that determines whether or not the utterance in the dialogue by a plurality of speakers is a message utterance indicating the message of the dialogue. Therefore, from the utterances in the dialogue, it is detected whether the utterance is a message confirmation utterance, which is a utterance for confirming the message in each utterance, and a story different from the speaker of the detected message confirmation utterance. It is provided with a determination unit that determines that the utterance immediately before the message confirmation utterance is the utterance of the message, and the determination unit is different from the speaker of the detected message confirmation utterance. Of a plurality of consecutive utterances of the other speaker including the utterance immediately before the message confirmation utterance, the similarity is toward the utterance immediately before the message confirmation utterance. speech in going up the range it determined that the requirements utterance.
Further, in order to solve the above problem, the message determination device according to the present invention determines whether or not the utterance in the dialogue by a plurality of speakers is a message utterance indicating the message of the dialogue. The device is a message confirmation utterance detection unit that detects whether the utterance is a message confirmation utterance that is a utterance for confirming a message in each utterance from the utterances in the dialogue, and the detected message confirmation. The utterance is made by a speaker different from the speaker of the utterance, and at least the utterance immediately before the message confirmation utterance is determined to be the message utterance, and the determination unit includes the detection unit. The message confirmation utterance is made by a speaker different from the speaker of the message confirmation utterance, and among a plurality of consecutive utterances of the other speaker including the utterance immediately before the message confirmation utterance, the message confirmation speech in the range of similarities towards the utterance of the previous utterance goes up you determines that the requirements utterance.

Further, in order to solve the above problem, the requirement determination method according to the present invention determines whether or not the utterance in the dialogue by a plurality of speakers is a requirement utterance indicating the requirement of the dialogue. It is a method, and it is detected from the utterances in the dialogue whether it is a message confirmation utterance that is a utterance for confirming a message in each utterance, and is different from the speaker of the detected message confirmation utterance. a speech by the speaker viewed including the determining that the speech immediately before at least the requirements confirmed utterance is the business utterance, the detected another speaker to the speaker of the requirements confirmation speech Of a plurality of consecutive utterances of another speaker including the utterance immediately before the message confirmation utterance, the degree of similarity increases toward the utterance immediately before the message confirmation utterance. The utterance within the range is determined to be the above-mentioned utterance .

Further, in order to solve the above problems, the program according to the present invention causes the computer to function as the above-mentioned requirement determination device.

According to the message determination device, the message determination method, and the program according to the present invention, it is possible to determine whether or not the utterance in the dialogue is a message utterance while suppressing an increase in cost.

It is a block diagram which shows the structural example of the requirement determination apparatus which concerns on 1st Embodiment of this invention. It is a block diagram which shows the structural example of the requirement determination apparatus which concerns on 2nd Embodiment of this invention. It is a figure for demonstrating the generation of a matter judgment model. It is a figure which shows an example of the dialogue between a customer and a person in charge of reception. It is a figure which shows the composition example of the message utterance learning data. It is a figure which shows the other composition example of the message utterance learning data. It is a block diagram which shows the structural example of the requirement determination apparatus which concerns on 3rd Embodiment of this invention. It is a figure which shows an example of the dialogue between a customer and a person in charge of reception. It is a figure which shows an example of the determination result by the requirement determination model shown in FIG. 7. It is a block diagram which shows the structural example of the requirement determination apparatus which concerns on 4th Embodiment of this invention. It is a figure for demonstrating the determination of the message confirmation utterance by the determination unit shown in FIG. It is a figure for demonstrating the determination of the message utterance by the determination unit shown in FIG. It is a figure for demonstrating the operation of the learning data generation part shown in FIG. It is a figure which conceptually shows machine learning which input natural language. It is a figure which shows the specific example of bag-of-words. It is a figure for demonstrating the problem of the machine learning method using the conventional natural language as an input. It is a figure for demonstrating the method of machine learning which input the natural language which concerns on this invention.

Hereinafter, embodiments for carrying out the present invention will be described with reference to the drawings.

(First Embodiment)
FIG. 1 is a block diagram showing a configuration example of the requirement determination device 10 according to the first embodiment of the present invention. In the message determination device 10 according to the present embodiment, the utterance in the dialogue by a plurality of speakers such as the dialogue between the customer and the person in charge of reception indicates the matter of the dialogue (the content that the customer wants to inquire of the person in charge of reception). It determines whether or not it is a message utterance.

In addition, human beings cannot always organize what they want to convey and speak without stagnation, and think or stagnate in the middle of the story. Therefore, the voice may be interrupted (silence section occurs) while the speaker is speaking. An utterance is a speech of a speaker separated by such a break in voice.

The requirement determination device 10 shown in FIG. 1 includes a determination unit 11.

The determination unit 11 determines whether or not the utterance in the dialogue is a message utterance based on the dialogue structure of the dialogue (dialogue between the customer and the person in charge of reception) by a plurality of speakers. Specifically, the determination unit 11 detects the message confirmation utterance, which is the utterance for confirming the message, from the utterances in the dialogue. In the case of a dialogue between the customer and the person in charge of reception, the utterance for confirming the matter for confirming the content (matter) that the customer wants to inquire is the utterance by the person in charge of reception. The determination unit 11 is a utterance by a speaker (speech by the customer) different from the detected utterance of the message confirmation (speech by the person in charge of response), and at least the utterance immediately before the message confirmation utterance is the utterance of the message. Is determined.

In general, in response to a customer's inquiry, the person in charge of responding often answers the customer's inquiry after confirming the content of the customer's inquiry. Therefore, the message confirmation utterance of the person in charge of response tends to include specific words and phrases (such as "-ne." And "-masune.") For confirming the content of the inquiry. Further, in order to confirm the contents of the customer's inquiry, the person in charge of responding often answers the customer's inquiry after repeating the contents of the customer's inquiry. Therefore, the message confirmation utterance of the person in charge of response tends to include many words and phrases in common with the customer's message utterance.

Based on these tendencies, the determination unit 11 determines whether or not the utterance of the person in charge of response is a message confirmation utterance. The determination unit 11 may determine whether or not the utterance of the person in charge of response is a message confirmation utterance by machine learning using various methods such as deep learning and a support vector machine. The pattern of the message confirmation utterance is fixed as compared with the message utterance. Therefore, it is relatively easy to determine the message confirmation utterance.

Then, the determination unit 11 determines whether or not the customer's utterance is a message utterance based on the determination result of the message confirmation utterance. Specifically, the determination unit 11 is a utterance of a speaker (customer) different from the speaker (response person in charge) of the message confirmation utterance, and the utterance immediately before the message confirmation utterance is the message utterance. Is determined.

In addition, the customer may talk about the matter so as to straddle a plurality of consecutive utterances. Therefore, in addition to the utterance immediately before the message confirmation utterance, for example, the determination unit 11 may determine the utterances of a plurality of consecutive customers immediately before the message confirmation utterance as the message utterance.

In addition, in a series of utterances, the customer may first talk about content that is not related to the business, and then talk about the business. Therefore, the determination unit 11 may determine the range of utterances of the plurality of consecutive customers immediately before the message confirmation utterance as the message utterance. For example, the degree of similarity between each of a plurality of consecutive customer utterances immediately before the message confirmation utterance and the message confirmation utterance is calculated, and the degree of similarity increases toward the utterance immediately before the message confirmation utterance. The utterance of may be determined to be a message utterance.

As described above, in the present embodiment, the message determination device 10 detects the message confirmation utterance, which is the utterance for confirming the message, from the utterances in the dialogue, and the speaker of the detected message confirmation utterance. It is provided with a determination unit 11 that determines that the utterance by a speaker different from the above is at least immediately before the message confirmation utterance as the message utterance.

The dialogue structure in which the message utterance exists immediately before the message confirmation utterance is common regardless of the field in which the dialogue is taking place. By determining whether or not the utterance in the dialogue is a message utterance using this dialogue structure, the dialogue does not cause an increase in cost such as preparing the message utterance learning data for each field to be judged. It is possible to determine whether or not the utterance in is a message utterance.

(Second Embodiment)
FIG. 2 is a diagram showing a configuration example of the requirement determination device 10A according to the second embodiment of the present invention. The message determination device 10A according to the present embodiment generates a message determination model for determining whether or not an utterance in a dialogue between a plurality of speakers, such as a dialogue between a customer and a person in charge of reception, is a message utterance. It is used to generate utterance learning data for the purpose. In FIG. 2, the same components as those in FIG. 1 are designated by the same reference numerals, and the description thereof will be omitted.

The condition determination device 10A shown in FIG. 2 is different from the condition determination device 10 shown in FIG. 1 in that the learning data generation unit 12 is added.

The learning data generation unit 12 generates the message utterance learning data 13 for generating a message determination model for determining whether or not the utterance in the dialogue is a message utterance by machine learning, based on the determination result of the determination unit 11. Generate. The generated message utterance learning data 13 is input to the determination model generation unit 14, for example, as shown in FIG. 3, and whether or not the utterance in the dialogue is the message utterance by machine learning by the determination model generation unit 14. The message determination model 15 for determining whether or not is generated. The message determination model 15 determines, for example, whether or not the utterance in the dialogue between the customer and the person in charge of reception at the call center is a message utterance. The requirement determination model 15 and the determination model generation unit 14 may be provided by the requirement determination device 10A, or may be provided by an external device other than the requirement determination device 10A.

FIG. 4 is a diagram showing an example of a dialogue between the customer and the person in charge of reception.

In the example shown in FIG. 4, the customer makes utterance # 1 "About time deposits" and then utterance # 2 "Do I need to go to the store to cancel the contract?" Customer utterances # 1 and # 2 correspond to utterances in which the customer inquires of the person in charge of the reception about the cancellation of the time deposit.

In response to customer utterances # 1 and # 2, the person in charge of reception makes utterance # 3 to confirm the contents of the customer's inquiry, such as "Is it possible to cancel the time deposit at a place other than the store?" In response to the utterance # 3 of the person in charge of reception, the customer makes an utterance # 4 indicating that the confirmation of the contents of the inquiry by the person in charge of reception is correct.

The responder receives the customer's utterance # 4 and makes an utterance # 5 asking for the information necessary for answering the customer's inquiry, and the customer makes an utterance # 6 as an answer to the customer's utterance # 5. Is going.

Taking the dialogue between the customer and the person in charge of reception shown in FIG. 4 as an example, the message utterance learning data 13 for generating the message determination model 15 will be described with reference to FIG.

In the dialogue between the customer and the person in charge of reception, the business dialogue is a dialogue by the customer. Therefore, as the learning data, customer utterances # 1, # 2, # 4, # 6 are extracted as shown in FIG. The message utterance learning data 13 is created by converting the dialogue voice into text by voice recognition processing. Here, fillers that are not related to the content of the dialogue, such as "ah" and "uh", may be removed. Then, information (message flag) indicating whether or not the utterance is a message utterance is added to each utterance of the customer. In the example shown in FIG. 5, when the message flag is "0", it is not a message utterance, and when the message flag is "1", it is a message utterance. Therefore, "1" is set in the message flags of utterances # 1 and # 2 indicating the dialogue message (contents that the customer wants to inquire), and "0" is set in the message flags of other utterances # 4 and # 6. Set. As described above, the message utterance learning data 13 is data in which the customer's utterance and the information indicating whether or not the utterance is the message utterance are associated with each other.

In FIG. 5, the description has been made using an example in which a message flag is set for each utterance of the customer, but the present invention is not limited to this. For example, as shown in FIG. 6, a plurality of consecutive utterances by the same speaker (customer) until the end of the utterance may be grouped as an utterance group, and a message flag may be set for each utterance group.

In the example shown in FIG. 6, one or a plurality of consecutive utterances (utterance # 1 and utterance # 2) until the customer finishes speaking are regarded as one utterance group. Then, a message flag is added to each utterance group to indicate whether or not the utterance is a message.

The message utterance learning data as shown in FIGS. 5 and 6 can be manually generated from the dialogue between the customer and the person in charge of reception. However, it is costly to generate such message utterance learning data for each field in which the message determination is desired.

Therefore, in the present embodiment, it is determined from the dialogue structure whether or not the utterance in the dialogue is a message utterance, and the determination result is used as the message utterance learning data 13. By doing so, the message utterance learning data 13 can be automatically generated from the dialogue in the field in which the message determination is desired. Then, by generating the message determination model 15 using the generated message utterance learning data 13, it is possible to determine whether or not the utterance in the dialogue is a message utterance while suppressing an increase in cost. ..

(Third Embodiment)
FIG. 7 is a diagram showing a configuration example of the requirement determination device 10B according to the third embodiment of the present invention. In FIG. 7, the same reference numerals are given to the configurations similar to those in FIGS. 2 and 3, and the description thereof will be omitted.

The condition determination device 10B shown in FIG. 7 is different from the condition determination device 10A shown in FIG. 2 in that the determination model generation unit 14 and the condition determination model 15 are added. That is, in the present embodiment, the message determination device 10B generates the message utterance learning data 13 from the determination result of whether or not the message utterance uses the dialogue structure, and the generated message utterance learning data 13 Is used to generate the message determination model 15. Then, the determination result of whether or not the utterance in the dialogue is the utterance of the message by the message determination model 15 is output.

Next, the operation of the requirement determination device 10B according to the present embodiment will be described.

As a pre-processing, a message determination model 15 for determining whether or not the utterance in the dialogue is a message utterance is generated by machine learning using the message utterance learning data 13 in which the message flag is added to the utterance in the dialogue. Will be done. As the message utterance learning data 13 used in the preprocessing, for example, data to which the message flag is manually added, data to which the message flag is added by the determination using the above-mentioned dialogue structure, and the like can be used. ..

The machine learning method is not particularly limited as long as an appropriate model can be generated based on the training data, and various methods such as deep learning and a support vector machine can be used. The information (feature) used for determining whether or not the utterance is a message is not particularly limited, and various information can be used so that a correct determination can be made.

Next, as online processing (real-time processing according to utterance), voice data (dialogue data) of dialogue between the customer and the person in charge of response is input to the message determination device 10A, and the message determination model 15 is used. Then, it is determined whether or not the utterance in the dialogue indicated by the dialogue data is a message utterance.

In the following, it is assumed that the dialogue data between the customer and the person in charge of reception is input as the dialogue data as shown in FIG. In addition, the utterance of the customer and the utterance of the person in charge of reception are input to the message determination device 10A on different channels (2 channels).

In the example shown in FIG. 8, the customer makes an utterance # 12 inquiring whether the service can be used immediately after changing the setting after the utterance # 11 saying "Yes, I understand." Utterance # 12 corresponds to a message dialogue indicating the dialogue message (content that the customer wants to inquire).

The person in charge of reception receives the customer's utterance # 12 and confirms the customer's inquiry content. After making the utterance # 13, the customer who has opened a foreign currency deposit account can immediately use the service in response to the inquiry. I am making utterance # 14 to answer that. Then, the customer receives the utterance # 14 of the person in charge of reception and makes the utterance # 15 asking for a more detailed answer about the timing when the service becomes available. Further, the person in charge of response receives the customer's utterance # 15 and makes an utterance # 16 to answer the timing when the service becomes available.

In such a dialogue, the message determination model 15 determines that the customer utterance # 11 is not a message utterance, and that the customer utterances # 12 and # 15 are message utterances, as shown in FIG. Suppose that it is judged. The message determination device 10B outputs the determination result of the message determination model 15.

In the present embodiment, the case where the message determination device 10B includes the learning data generation unit 12 and the determination model generation unit 14 has been described, but the present invention is not limited to this, and the message determination device 10B is not limited to this. Another external device may include a learning data generation unit 12 and a determination model generation unit 14. In this case, the message determination device 10B acquires the message determination model 15 generated by the external device, and outputs the determination result of the acquired message determination model 15.

(Fourth Embodiment)
FIG. 10 is a diagram showing a configuration example of the requirement determination device 10C according to the fourth embodiment of the present invention. In FIG. 10, the same reference numerals are given to the same configurations as those in FIG. 7, and the description thereof will be omitted.

The condition determination device 10C shown in FIG. 10 is different from the condition determination device 10B shown in FIG. 7 in that the learning data generation unit 12 is changed to the learning data generation unit 12C.

The learning data generation unit 12C inputs the determination result of the determination unit 11 and the determination result of the message determination model 15, and generates the message utterance learning data 13 based on these determination results.

Next, the operation of the requirement determination device 10C according to the present embodiment will be described. The condition determination device 10C according to the present embodiment also performs pre-processing and online processing in the same manner as the condition determination device 10B according to the third embodiment.

Next, as post-processing, the determination unit 11 determines whether or not the utterance in the dialogue indicated by the input dialogue data is a message utterance based on the dialogue structure.

First, the determination unit 11 determines whether or not the utterance of the person in charge of response is a message confirmation utterance for confirming the contents of the customer's inquiry. As mentioned above, the response person's message confirmation utterance often includes specific words (such as "-ne.", "-Masu-ne.") To confirm the content of the customer's inquiry. There is a tendency. In addition, the response person's message confirmation utterance tends to contain many words and phrases in common with the customer's message utterance. Based on these tendencies, the determination unit 11 determines whether or not the utterance of the person in charge of response is a message confirmation utterance.

In the example shown in FIG. 8, the utterance # 13 of the person in charge of response includes a phrase for confirming the content of "...". In addition, the utterance # 3 of the person in charge of reception includes many words and phrases included in the utterance # 12 of the customer before the utterance # 13, such as "foreign currency deposit service" and "usage setting". Therefore, as shown in FIG. 11, the determination unit 11 determines that the utterance # 13 of the person in charge of response is a message confirmation utterance. On the other hand, since the utterances # 14 and # 16 of the person in charge of the reception do not include words and phrases for confirming the contents such as "...", the judgment unit 11 determines the utterances # 14 and # 16 of the person in charge of the reception. Is not a message confirmation utterance.

Next, the determination unit 11 determines whether or not the customer's utterance is a message utterance based on the determination result of the message confirmation utterance. Specifically, as shown in FIG. 12, the determination unit 11 is an utterance of a speaker (customer) different from the speaker (response person) of the message confirmation utterance, which is a message confirmation utterance. The utterance # 12 immediately before # 13 is determined to be a message utterance, and the other utterances # 11 and # 15 are determined not to be a message utterance.

Note that FIG. 12 shows an example in which only the utterance immediately before the message confirmation utterance is determined to be the message utterance, but as described above, a plurality of consecutive utterances including the utterance immediately before the message confirmation utterance. The customer's utterance may be determined to be a business utterance.

The learning data generation unit 12 compares the determination result by the requirement determination model 15 with the determination result by the determination unit 11. Then, as shown in FIG. 13, the learning data generation unit 12 adds an utterance in which the determination result by the message determination model 15 and the determination result by the determination unit 11 match the message utterance learning data 13.

In the example shown in FIG. 13, the determination result of the utterance # 11 (not the utterance of the message) and the determination result of the utterance # 12 (the utterance of the message) are the same in the message determination model 15 and the determination unit 11. The learning data generation unit 12C adds the utterances # 11 and # 12 and the determination result to the message utterance learning data 13. When the determination results match between the message determination model 15 and the determination unit 11, the reliability of the determination results is considered to be high. Therefore, by setting the determination result that matches the message determination model 15 and the determination unit 11 as the message utterance learning data 13, the machine learning using the message utterance learning data 13 ensures that the message determination model 15 is reliable. It is possible to improve the sex.

Even if the learning data generation unit 12C adds the determination result by the determination unit 11 to the message utterance learning data 13 without comparing the determination result by the message determination model 15 with the determination result by the determination unit 11. Good.

As described above, in the present embodiment, the message utterance device 10C determines whether or not the utterance in the dialogue is a message utterance, and the determination unit 11 and the message determination model 15 match the determination result. A learning data generation unit 12C to be added to the learning data 13 is provided.

By setting the determination result that matches the message determination model 15 and the determination unit 11 as the message utterance learning data 13, the reliability of the message determination model 15 can be determined by machine learning using the message utterance learning data 13. It can be improved.

As described above, the message utterance learning data 13 may be generated not in utterance units but in utterance group units in which a plurality of consecutive utterances are put together. By using the learning data to which the message flag is added for each utterance group, it is possible to determine whether or not the message is uttered for each utterance group.

In FAQ search (Frequently Asked Questions), which searches for automatic responses to customers and frequently asked questions by performing message judgment for each utterance group, the message judgment model 15 can be generated in consideration of the timing of the response to the utterance. It will be possible.

In addition, the dialogue between the customer and the person in charge of reception is converted into text by voice recognition processing, and then processed by the message determination model 15. Here, an error may occur in the voice recognition process. Therefore, a process using the N-best method in which a plurality of candidates are prepared as recognition results by the voice recognition process of the word included in the utterance may be performed.

FIG. 14 is a diagram conceptually showing machine learning using natural language as an input.

In machine learning using natural language as an input, in order to perform numerical vectorization according to the input of machine learning (support vector machine (SVM)), the identity calculation for the input text is performed. That is, at the time of learning or judgment, it is necessary to perform numerical vectorization of the input text according to the input of machine learning. As such a feature calculation, for example, a model (bag-of-words) is generally used in which only whether or not a sentence contains words is considered and the arrangement of words is not considered.

FIG. 15 is a diagram showing a specific example of bag-of-words.

In bag-of-words, if a sentence contains a word, the numerical value corresponding to the word is set to 1, and an input vector expressing whether or not the sentence contains a word is calculated. If the input text is "Can I cancel the time deposit on the Internet?", Morphological analysis is performed on the input text.

Specifically, as shown in FIG. 15, by morphological analysis of a large amount of text, a list is made so as to cover the words appearing in the text, and a word list in which a word number is assigned to each word is generated in advance. To. Then, among the words included in the word list, the value of the input vector corresponding to the word number of the word appearing in the input text becomes "1". The value of the input vector corresponding to the word number of the word that does not appear in the input text is "0".

Instead of the morphological analysis, a feature calculation method such as bag-of-pos using part of speech, or a feature calculation combining bag-of-words and bag-of-pos may be used.

For processing in the message judgment model 15, it is possible to perform morphological analysis on the voice recognition result text in which the voice of the dialogue is converted into text by voice recognition processing, and perform feature calculation by bag-of-words or the like from the result of the morphological analysis. Conceivable.

Here, it is assumed that an error occurs in the voice recognition process. For example, suppose that the voice "Can I cancel the time deposit on the Internet?" Is erroneously recognized as "Can the enemy do the harm of money on the Internet?" As shown in Fig. 16. .. When morphological analysis is performed on such an erroneous speech recognition result text and an input vector is calculated, errors are accumulated and it is not possible to perform feature calculation reflecting the correct input speech.

As described above, the conventional method in which the voice recognition result is used as the input for machine learning has a problem that it tends to be an erroneous morpheme unit. The cause of such a problem is that the speech recognition dictionary used for speech recognition processing and the morphological analysis dictionary used for morphological analysis are often different, and inconsistencies occur due to the difference in words registered in these dictionaries. May occur. Further, as another cause, since the morphological analysis targets a normal sentence that can be read by a human, an erroneous morphological analysis may be performed due to an error in the speech recognition result text.

Further, in the conventional method in which the voice recognition result is used as the input for machine learning, if an error occurs in the voice recognition process, the feature calculation is performed while including the error, so that there is a problem that the feature does not reflect the correct input voice. ..

Therefore, in the present invention, as shown in FIG. 17, the word sequence (N-best result) of the N-position candidate obtained as a result of the speech recognition process is subjected to feature calculation (bag-of-words, etc.) by machine learning. Used for.

In the speech recognition process, a process of searching for a word string closest to the input speech among the combinations of registrations (including part of speech information) registered in the speech recognition dictionary is performed. Therefore, as a result of the voice recognition process, a sequence of words (including part of speech information) can be obtained. In addition, up to N-rank candidates other than the 1-rank candidate can be obtained in the order of proximity to the input voice. Therefore, even if the 1st place candidate is incorrect, there is a high possibility that the correct word is included in the Nth place candidate.

As described above, in the present invention, the word sequence of the N-position candidate obtained as a result of the speech recognition process is used without performing the morphological analysis process. Therefore, no error occurs due to the morphological analysis, and the result of the speech recognition process including the error is directly reflected in the feature. Moreover, since morphological analysis is not performed, the amount of processing can be reduced. Moreover, since morphological analysis is not performed, it is not necessary to prepare a morphological analysis dictionary. In addition, since the voice recognition results up to the Nth candidate are reflected in the features, even if a voice recognition error occurs in the 1st candidate, there is a high possibility that the correct words are included in the Nth candidate, and the features are calculated. Can be reflected in.

Although not particularly mentioned in the embodiment, a program for executing each process performed by the computer functioning as the message determination device 10 may be provided. The program may also be recorded on a computer-readable medium. It can be installed on a computer using a computer-readable medium. Here, the computer-readable medium on which the program is recorded may be a non-transient recording medium. The non-transient recording medium is not particularly limited, but may be, for example, a recording medium such as a CD-ROM or a DVD-ROM.

Although the above embodiments have been described as typical examples, it will be apparent to those skilled in the art that many modifications and substitutions are possible within the spirit and scope of the present invention. Therefore, the present invention should not be construed as being limited by the above-described embodiments, and various modifications and modifications can be made without departing from the scope of claims. For example, it is possible to combine a plurality of constituent blocks described in the configuration diagram of the embodiment into one, or to divide one constituent block into one.

10,10A, 10B, 10C Subject Judgment Device 11 Judgment Unit 12, 12C Learning Data Generation Unit 13 Subject Speaking Learning Data 14 Judgment Model Generation Unit 15 Subject Judgment Model

Claims (8)

A message determination device for determining whether or not an utterance in a dialogue by a plurality of speakers is a message utterance indicating the message of the dialogue.
From the utterances in the dialogue, it is detected whether the utterance is a message confirmation utterance, which is an utterance for confirming the message in each utterance, and a utterance by a speaker different from the speaker of the detected message confirmation utterance is detected. It is provided with a determination unit for determining at least the utterance immediately before the message confirmation utterance as the message utterance .
The determination unit is a utterance by a speaker other than the speaker of the detected message confirmation utterance, and is a continuous utterance of a plurality of the other speakers including the utterance immediately before the message confirmation utterance. of, requirements determination apparatus characterized that you determine an utterance in a range going up the similarity toward the speech immediately before the business check utters the requirements utterance. In the requirement determination device according to claim 1,
The determination unit detects a message confirmation utterance from the utterances in the dialogue by using a machine learning model for determining whether the input utterance unit is the message confirmation utterance. Judgment device. A message determination device for determining whether or not an utterance in a dialogue by a plurality of speakers is a message utterance indicating the message of the dialogue.
From the utterances in the dialogue, the message confirmation utterance detection unit that detects whether the utterance is a message confirmation utterance, which is a utterance for confirming the message in each utterance unit,
It is provided with a determination unit that determines that the utterance by a speaker different from the speaker of the detected message confirmation utterance and at least immediately before the message confirmation utterance is the message utterance .
The determination unit is a utterance by a speaker other than the speaker of the detected message confirmation utterance, and is a continuous utterance of a plurality of the other speakers including the utterance immediately before the message confirmation utterance. of, requirements determination apparatus characterized that you determine an utterance in a range going up the similarity toward the speech immediately before the business check utters the requirements utterance. In the requirement determination device according to claim 3,
The message confirmation utterance detection unit is characterized in that only the utterance of the first speaker is input from the utterances in the dialogue, and whether the utterance is the message confirmation utterance is detected in the utterance unit of the utterance. The utterance judgment device. In the requirement determination device according to claim 3 or 4,
The message confirmation utterance detection unit detects whether or not the message confirmation utterance is a message confirmation utterance based on whether or not the utterance unit includes a specific phrase for confirming the inquiry content. apparatus. In the requirement determination device according to claim 3 or 4,
The message confirmation utterance detection unit is a message determination device characterized by detecting a message confirmation utterance by using a machine learning model for determining whether or not the message confirmation utterance is an input utterance unit. A message determination method for determining whether or not an utterance in a dialogue by a plurality of speakers is a message utterance indicating the message of the dialogue.
From the utterances in the dialogue, it is detected whether the utterance is a message confirmation utterance, which is an utterance for confirming the message in each utterance, and a utterance by a speaker different from the speaker of the detected message confirmation utterance is detected. a is, seen including: determining a speech immediately before at least the requirements confirmed utterance is the business utterance,
Of a plurality of consecutive utterances of the other speaker including the utterance immediately before the message confirmation utterance, which is a utterance by a speaker different from the speaker of the detected message confirmation utterance. A matter determination method, characterized in that an utterance within a range in which the degree of similarity increases toward an utterance immediately before a matter confirmation utterance is determined as the said utterance . A program for operating a computer as a requirement determination device according to any one of claims 1 to 6 .

Images