JP2023173036A - Dialogue information extraction device and dialogue information extraction method - Google Patents

Abstract

To extract information on a dialogue having a viewpoint to which attention should be paid from a dialogue made by a plurality of speakers and enable a dialogue situation to be efficiently checked.SOLUTION: A dialogue information extraction device 100 includes: an utterance information input unit 111 that for each of utterances constituting a dialogue, inputs utterance information including a written text of the utterance; a dialogue structure estimation unit 112 that using an estimation model having learned a dialogue structure, estimates the dialogue structure of the dialogue from utterance information on a plurality of utterances in the dialogue; and a viewpoint-based dialogue extraction unit 113 that determines whether or not there is a predetermined viewpoint to which attention should be paid in an utterance group composed of utterances of a plurality of speakers generated based on the dialogue structure, thereby extracting an utterance group having the viewpoint.SELECTED DRAWING: Figure 1

Description

The present invention relates to a dialogue information extraction device and a dialogue information extraction method, and is suitable for application to a dialogue information extraction device and a dialogue information extraction method that analyze dialogues between a plurality of speakers.

In recent years, various efforts have been made to transcribe spoken dialogue in order to expand the value of digital dialogue services.

For example, when dealing with a call center or at a counter, when bank staff or employees provide explanations to customers and proceed with procedures, it is necessary to proceed with the dialogue while checking the status of the dialogue, with particular emphasis on the customer's consent. . In this case, any communication without clear consent from the customer poses compliance risks, such as violating the Financial Instruments Sales Act. Therefore, it is very important to confirm the content of the utterance and the state of consent to it during the dialogue or after the fact.

As a related art related to the above background, for example, Patent Document 1 discloses a call center conversation content display system that displays the response for each speaker in a chat format.

International Publication No. 2019/003395

However, the system disclosed in Patent Document 1 relatively simply displays the response for each speaker in a chat format, and when trying to confirm the content of the response conversation with respect to a predetermined viewpoint that should be noted. required manual verification from the transcribed text. In this case, not only would the cost of manual work be enormous, but there was also the risk of omissions in confirmation.

The present invention has been made in consideration of the above points, and is a dialogue information extraction method that extracts information on dialogues that have noteworthy viewpoints from dialogues between multiple speakers, and enables efficient confirmation of the dialogue situation. This paper attempts to propose a device and a method for extracting dialogue information.

In order to solve this problem, the present invention provides a dialogue information extraction device that analyzes dialogues between a plurality of speakers based on noteworthy viewpoints, and which transcribes each utterance constituting the dialogue. an utterance information input unit that inputs utterance information including text; and a dialogue structure estimation unit that estimates the dialogue structure of a dialogue from the utterance information of a plurality of utterances in the dialogue using an estimation model that has learned the dialogue structure. and a viewpoint-specific dialogue extraction unit that extracts a group of utterances having the viewpoint by determining the presence or absence of the viewpoint from a group of utterances generated based on the dialogue structure and consisting of utterances of a plurality of speakers. Provided is a dialogue information extraction device characterized by comprising the following.

Further, in order to solve such problems, the present invention provides a dialogue information extraction method using a dialogue information extraction device that analyzes dialogues between a plurality of speakers based on viewpoints to be focused on, the dialogue information extraction device comprising: an utterance information input step of inputting utterance information including a text transcribed from the utterance for each utterance constituting the dialogue; a dialogue structure estimating step of estimating a dialogue structure of the dialogue from the utterance information of a plurality of utterances in the dialogue input in the input step; and a dialogue information extraction device, based on the estimation result of the dialogue structure estimation step. The system is characterized by comprising a perspective-specific dialogue extraction step of extracting a group of utterances having the viewpoint by determining the presence or absence of the viewpoint from the generated utterance group consisting of utterances of a plurality of speakers. A method for extracting dialogue information is provided.

According to the present invention, it is possible to extract dialogue information having a noteworthy viewpoint from dialogues between a plurality of speakers, and to efficiently check the dialogue situation.

FIG. 1 is a block diagram showing a configuration example of a dialogue information extraction device 100 according to a first embodiment of the present invention. 3 is a flowchart illustrating an example of a processing procedure of dialogue information extraction processing. It is a figure which shows an example of utterance information. It is a figure which shows an example of a past utterance information group. FIG. 3 is a diagram illustrating an example of a dialog structure estimation result. FIG. 7 is a diagram illustrating an example of dialogue information regarding an utterance set having a viewpoint. FIG. 6 is a diagram illustrating a display example of a result of extracting dialogue information. FIG. 2 is a block diagram showing a configuration example of a dialogue information extraction device 101 according to a second embodiment of the present invention. 7 is a flowchart illustrating an example of a processing procedure of dialogue information extraction processing in the second embodiment. 1 is a block diagram showing an example of a hardware configuration of dialogue information extraction devices 100 and 101. FIG.

Embodiments of the present invention will be described in detail below with reference to the drawings.

Note that the following description and drawings are examples for explaining the present invention, and are omitted and simplified as appropriate to clarify the explanation. Furthermore, not all combinations of features described in the embodiments are essential to the solution of the invention. The present invention is not limited to the embodiments, and any application examples that match the idea of the present invention are included within the technical scope of the present invention. Those skilled in the art can make various additions and changes to the present invention within the scope of the present invention. The present invention can also be implemented in various other forms. Unless specifically limited, each component may be plural or singular.

In the following explanation, various information may be explained using expressions such as "table", "table", "list", "queue", etc., but various information may also be expressed using data structures other than these. good. "XX table", "XX list", etc. are sometimes referred to as "XX information" to indicate that they do not depend on the data structure. When explaining the contents of each piece of information, expressions such as "identification information", "identifier", "name", "ID", and "number" are used, but these can be replaced with each other.

In addition, in the following explanation, when the same type of elements are explained without distinguishing them, reference numerals or common numbers in the reference signs are used, and when the same kind of elements are explained separately, the reference numerals of the elements are used. Alternatively, an ID assigned to the element may be used instead of the reference code.

In addition, in the following explanation, processing performed by executing a program may be explained, but the program is executed by at least one or more processors (for example, a CPU) to store predetermined processing as appropriate. Since the processing is performed using resources (for example, memory) and/or interface devices (for example, communication ports), the main body of the processing may be a processor. Similarly, the subject of processing performed by executing a program may be a controller having a processor, a device, a system, a computer, a node, a storage system, a storage device, a server, a management computer, a client, or a host. The main body (for example, a processor) that performs processing by executing a program may include a hardware circuit that performs part or all of the processing. For example, the main body of processing performed by executing a program may include a hardware circuit that performs encryption and decryption, or compression and expansion. A processor operates as a functional unit that implements a predetermined function by operating according to a program. Devices and systems that include processors are devices and systems that include these functional units.

A program may be installed on a device, such as a computer, from a program source. The program source may be, for example, a program distribution server or a computer-readable storage medium. When the program source is a program distribution server, the program distribution server includes a processor (for example, a CPU) and a storage resource, and the storage resource may further store a distribution program and a program to be distributed. Then, by the processor of the program distribution server executing the distribution program, the processor of the program distribution server may distribute the program to be distributed to other computers. Furthermore, in the following description, two or more programs may be realized as one program, or one program may be realized as two or more programs.

In the following description, a "viewpoint" is a viewpoint from which attention is focused on an utterance (viewpoint of interest), and specifically includes, for example, agreement or disagreement with the utterance. Then, if some meaning is made from that perspective through dialogue exchange (in other words, in the text transcribed from multiple utterances), in other words, in the above specific example, it can be determined that there is agreement or disagreement with the utterances. ``I have a point of view.'' The viewpoint can be set arbitrarily, and may be specified by the user or determined by a program or the like, for example. In addition, in this explanation, a series of dialogues consisting of utterances from multiple speakers is treated as a "dialogue scenario", and the dialogue information extraction device according to the present invention analyzes the dialogue based on the viewpoint to be focused on for each dialogue scenario. do.

(1) First Embodiment FIG. 1 is a block diagram showing a configuration example of a dialogue information extraction device 100 according to a first embodiment of the present invention. The dialogue information extraction device 100 shown in FIG. 1 includes a dialogue information input unit 111, a dialogue structure estimation unit 112, an utterance set generation unit 114, and a viewpoint determination unit 115 as processing units having predetermined processing functions. 113 and an extraction result display section 116, and includes a dialogue structure estimation model storage section 121, a viewpoint determination model storage section 122, a past utterance information storage section 123, a dialogue structure storage section 124, and an extraction result display section 116 as storage sections for storing data. A result storage unit 125 is provided.

The utterance information input unit 111 has a function of inputting utterance information including text transcribed from the utterance to the dialogue information extraction device 100 for each utterance. The utterance information input unit 111 may be configured to acquire text from outside the dialogue information extraction device 100 (such as a dialogue recording device that records dialogue content as text data), or may be configured to transcribe the text from the audio data of the utterance itself. But that's fine.

The dialogue structure estimation unit 112 has a function of estimating the dialogue structure in the current dialogue scenario using an estimation model learned by a statistical method.

The viewpoint-specific dialogue extraction unit 113 has a function of extracting information regarding a dialogue having a predetermined viewpoint in the dialogue scenario, based on the estimation result of the dialogue structure in the dialogue scenario. Of the viewpoint-specific dialogue extracting unit 113, the utterance set generating unit 114 has a function of generating a utterance set indicating a combination of utterances having a strong dialogue relationship based on the estimation result of the dialogue structure. Further, the viewpoint determination unit 115 has a function of determining whether or not there is a viewpoint in the utterance set generated by the utterance set generation unit 114.

The extraction result display unit 116 has a function of outputting dialogue information regarding the utterance set that corresponds to the user's specified conditions, among the utterance sets having a predetermined viewpoint extracted by the viewpoint-based dialogue extraction unit 113.

The dialogue structure estimation model storage unit 121 stores an estimation model used when the dialogue structure estimation unit 112 estimates a dialogue structure.

The viewpoint determination model storage unit 122 stores a determination model used by the viewpoint determination unit 115 to determine the presence or absence of a viewpoint in the utterance set.

The past utterance information storage unit 123 stores a past utterance information group that is a collection of utterance information (past utterance information) of a series of dialogues in each dialogue scenario. FIG. 4, which will be described later, shows a specific example of the past utterance information group.

The dialogue structure storage unit 124 stores information indicating the estimation result of the dialogue structure in the dialogue scenario. The dialogue structure is estimated by the dialogue structure estimation unit 112, and a specific example of the dialogue structure estimation result is shown in FIG. 5, which will be described later.

The extraction result storage unit 125 stores dialogue information to be extracted determined by the viewpoint-based dialogue extraction unit 113 (viewpoint determination unit 115).

FIG. 2 is a flowchart illustrating an example of a processing procedure for dialogue information extraction processing. The dialogue information extraction process shown in FIG. 2 is executed by each part of the dialogue information extraction device 100. The dialog information extraction process may be executed sequentially online while a dialog between two or more speakers is in progress, or may be executed after the dialog ends. In the following, the processing procedure of the dialog information extraction processing shown in FIG. 2 will be explained with reference to other drawings as appropriate, taking as an example the case where the dialog information extraction processing is executed sequentially. Note that the viewpoint of interest for the utterance is specified, for example, by the user before the process shown in FIG. 2 is started.

According to FIG. 2, first, the utterance information input unit 111 inputs text data (utterance information) in which the latest utterance is transcribed (step S101).

FIG. 3 is a diagram showing an example of speech information. The utterance information 210 shown in FIG. 3 includes the following items: utterance ID 211, utterance time 212, utterance user ID 213, and text 214. The utterance ID 211 indicates an identifier (utterance ID) assigned to each utterance. The utterance ID may be an identifier assigned according to the order of utterances regardless of the dialogue scenario, or may be an identifier assigned according to the order of utterances for each dialogue scenario. The utterance time 212 indicates the time at which the utterance occurred. The utterance user ID 213 indicates the identifier (user ID) of the speaker of the utterance. Text 214 indicates a text that is a transcription of the content of the utterance.

Next to step S101, the utterance information input unit 111 refers to the past utterance information group stored in the past utterance information storage unit 123, and determines whether there is utterance information regarding past utterances in the current dialogue scenario (past utterance information). It is confirmed whether or not (step S102).

FIG. 4 is a diagram showing an example of a past utterance information group. The past utterance information group 220 shown in FIG. 4 includes the following items: utterance ID 221, utterance time 222, user ID 223, and text 224. Basically, each item of the past utterance information group 220 corresponds to each item 211 to 214 of the utterance information 210 illustrated in FIG. 3. The utterance ID 221 indicates an identifier that can specify the order of utterances in the dialogue scenario corresponding to the past utterance information group 220. Note that if the utterance ID 211 of the utterance information 210 is an identifier assigned in the order of utterances regardless of the dialogue scenario, when registering the utterance information 210 in the past utterance information group 220, the value of the utterance ID 211 is set to the utterance in the dialogue scenario. It is only necessary to convert the values into corresponding values in order and then register them in the utterance ID 221.

If past utterance information does not exist in step S102 (NO in step S102), the utterance information input in step S101 is utterance information indicating the first utterance in the current dialogue scenario. In this case, no dialogue has been formed in the dialogue scenario at this point, and there is no situation in which to extract dialogue information. Therefore, the utterance information input unit 111 stores the utterance information input in step S101 in the past utterance information storage unit 123 as a group of past utterance information of a new dialogue scenario (step S110), and ends the dialogue information extraction process.

On the other hand, if past utterance information exists in step S102 (YES in step S102), the utterance information input unit 111 combines the utterance information input in step S101 with the current utterance information stored in the past utterance information storage unit 123. and the past utterance information group of the dialogue scenario are input to the dialogue structure estimation unit 112 (step S103).

Then, the dialogue structure estimation unit 112 estimates the dialogue structure in the current dialogue scenario based on the information input in step S103 (step S104). The dialogue structure estimation unit 112 then stores the dialogue structure estimation result in the dialogue structure storage unit 124.

Here, the estimation of the dialogue structure by the dialogue structure estimation unit 112 in step S104 will be explained in detail.

The dialogue structure estimation unit 112 uses an estimation model learned by a statistical method to calculate the probability of the presence or absence of a relationship between two utterances by different speakers, and based on the calculation result, there is a dialogue relationship. The dialogue structure is estimated by estimating the combination of utterances. The above estimation model is stored in the dialogue structure estimation model storage unit 121.

The dialogue information extraction device 100 (for example, the dialogue structure estimating unit 112) can learn the estimation model stored in the dialogue structure estimation model storage unit 121 using the learning data to which a dialogue structure has been added. Specifically, for example, when learning the estimation model, the following inputs are prepared. The uttered sentence (text 224) is vectorized using Word2Vec, Globe, or a pre-trained language model. Furthermore, the utterance time (utterance time 222) is treated as a vector as it is. Furthermore, the speaker (user ID 223) vectorizes the vector as a one-hot vector.

By learning an estimation model using the above training data, the feature vectors to be input during training of the estimation model or estimation using the estimation model are "two utterances targeted for relationship estimation. ``difference in utterance time between the two utterances'', ``number of utterances per speaker between the two utterances targeted for relationship estimation'', and ``similarity of uttered sentences between the two utterances targeted for relationship estimation''. have Note that the above-mentioned "two utterances targeted for relationship estimation" is a combination of one utterance and one utterance earlier than that utterance, and how many past utterances to combine is determined by Can be set arbitrarily. Further, "similarity of uttered sentences" indicates, for example, whether words or phrases that are semantically similar appear in the uttered sentences.

FIG. 5 is a diagram illustrating an example of a dialog structure estimation result. The dialogue structure data 230 shown in FIG. 5 is data showing an example of a dialogue structure estimated by the dialogue structure estimating unit 112, in which a combination of utterances having a dialogue relationship is determined using the utterance ID of each utterance. shown. The dialogue structure data 230 has a record for each dialogue structure, and each record is configured to include an utterance ID 231 for utterance A and an utterance ID 232 for utterance B. The values of the utterance IDs 231 and 232 correspond to the value of the utterance ID 221 in the past utterance information group 220 in FIG.

In this description, when one utterance is related to another utterance in a dialogue, the former utterance will be referred to as "utterance A" and the latter utterance will be referred to as "utterance B." That is, utterance B is an utterance that occurred in the past than utterance A, and the speaker of utterance A and the speaker of utterance B are different people.

Furthermore, in the dialogue structure estimation result, the combination of utterance A and utterance B does not necessarily have to be one-to-one. Specifically, for example, in the case of FIG. 5, utterance B with utterance ID "4" has a dialogue structure (dialogue relationship) with utterance A with utterance ID "5" and utterance A with utterance ID "6". ).

After the dialogue structure estimation unit 112 estimates the dialogue structure in step S104 as described above, the utterance set generation unit 114 of the viewpoint-based dialogue extraction unit 113 determines strong relationships in the dialogue based on the dialogue structure estimation result. A utterance set having the following information is generated (step S105).

An example of a method for generating an utterance set will be explained. The utterance set generation unit 114 extracts one utterance that satisfies a predetermined condition (for example, the closest utterance time) from among past utterances that have a relationship with one utterance at a probability value greater than or equal to an arbitrary threshold value, and An utterance set is a group of utterances that are a combination of utterances. According to a rule of thumb, the above predetermined condition is a past utterance with the closest utterance time, but is not limited to this. Further, the above-mentioned threshold value is set, for example, in the range of 0 to 1 by the user according to the purpose of extraction, with 0.5 as the standard. If you want to increase the recall rate (prevent missing extraction), you can set the threshold low. Furthermore, if it is desired to increase the recall rate, the utterance set may be extracted from a wider range of subjects.

In addition, as a modification of the utterance set generation method, the utterance set generation unit 114 extracts a plurality of utterances with similar utterance times from past utterances for one utterance, and sets the size of the utterance set to 2 or more. good. Alternatively, a plurality of utterances with similar utterance times may be extracted from among past utterances that are related by a probability value greater than or equal to an arbitrary threshold value.

Returning to the explanation of FIG. 2. After generating the utterance set in step S105, the utterance set generation unit 114 inputs the generated utterance set to the viewpoint determination unit 115 (step S106).

Next, the viewpoint determination unit 115 determines the presence or absence of a viewpoint in each utterance set using the determination model learned by the statistical method (step S107). The above judgment model is stored in the viewpoint judgment model storage unit 122.

An example of a method for determining the presence or absence of a viewpoint by the viewpoint determination unit 115 in step S107 will be described in detail. First, a pre-trained language model is used as the judgment model. Then, the viewpoint determination unit 115 uses the utterances of the utterance set arranged in order of utterance time and their labels (presence or absence of viewpoint) as one sample, and fine-tunes the pre-trained language model using that sample set. The decision model is learned by For example, if the number of utterances constituting an utterance set is two, it can be treated as a sentence pair classification problem. Further, even if the number of utterances constituting the utterance set is three or more, it may be treated as a classification problem in the same way.

Next, the viewpoint-specific dialogue extraction unit 113 (for example, the viewpoint determination unit 115) determines an utterance set having a viewpoint as an extraction target based on the determination result in step S107 (step S108). Then, the viewpoint-specific dialogue extraction unit 113 stores dialogue information regarding the determined extraction target utterance set in the extraction result storage unit 125.

Note that, as described above, in this embodiment, "having a viewpoint" means having some meaning from the viewpoint. Therefore, when the viewpoint of interest is the customer's agreement or disagreement, not only the utterance set that includes the customer's utterance of agreement but also the utterance set that includes the customer's utterance of disagreement is determined as the utterance set that has the viewpoint. be done.

FIG. 6 is a diagram illustrating an example of dialogue information regarding an utterance set having a viewpoint. The dialogue information 240 shown in FIG. 6 has a record for each utterance set determined to have a viewpoint, and each record includes a user ID 241 of utterance A, an utterance ID 242 of utterance A, a utterance A 243, and a user ID 244 of utterance B. , utterance ID 245 of utterance B, utterance B 246, and viewpoint 247.

In the dialog information 240, the user ID 241 of the utterance A, the utterance ID 242 of the utterance A, and the utterance A 243 are utterance information regarding "utterance A" on the answering side that constitutes the utterance set. The utterance ID 242 of the utterance A can be obtained from the utterance ID 231 of the dialogue structure data 230 (see FIG. 5). By referring to the past utterance information group 220 using this utterance ID as a key (see FIG. 4), the user ID 241 can be obtained from the user ID 223, and the utterance A 243 can be obtained from the text 224. Further, the user ID 244 of utterance B, the utterance ID 245 of utterance B, and the utterance B 246 are utterance information regarding "utterance B" of the inquiring side that constitutes the utterance set. Therefore, in the same way as the utterance information regarding utterance A, the utterance ID 245 of utterance B is derived from the utterance ID 232 of the dialogue structure data 230, the user ID 244 is derived from the user ID 223 of the past utterance information group 220, and the utterance B 246 is the text of the past utterance information group 220. 224, respectively. Furthermore, the viewpoint 247 indicates the content of the viewpoint that the utterance set has. The content of the viewpoint is, for example, "agree" or "disagree", and these contents are determined when the viewpoint determination unit 115 determines whether there is a viewpoint.

After step S108, the extraction result display unit 116 filters the dialogue information to be extracted stored in the extraction result storage unit 125 using the filter conditions specified by the user, and sets the utterances that correspond to the filter conditions. The dialog information regarding the extraction process is output as an extraction result (step S109).

FIG. 7 is a diagram illustrating a display example of the dialogue information extraction results. A display screen 250 shown in FIG. 7 is an example of a screen displayed on a terminal that can be operated by a user using a GUI or the like (a display included in the dialog information extraction device 100 may be used), and includes an input screen 251 and an output screen 252. It is composed of The input screen 251 accepts the user's designation of filter conditions and displays the designation contents. Note that the filter conditions can be arbitrarily specified by the user based on keywords, speaking users, viewpoints, etc., but the condition items are not limited to these. The output screen 252 displays the extraction results of dialogue information based on the filter conditions.

To explain specifically with reference to FIG. 7, on the input screen 251, "contract" is specified in the keyword search, "User-B" is specified in the user search, and "disagree" is specified in the extraction viewpoint. has been done. In this case, the extraction result display unit 116 displays the user ID 241 or A record in which the user ID 244 is "User-B" and the viewpoint 247 is "disagree" is searched. Note that the dialog information extraction device 100 may be configured to be able to accept input operations from the user on the input screen 251 at any timing before step S109 in FIG. Then, in step S109 of FIG. 2, the extraction result display unit 116 acquires the contents of records that correspond to the various search conditions (filter conditions) described above from the dialogue information to be extracted stored in the extraction result storage unit 125. This is displayed on the output screen 252.

The output screen 252 output as described above is a screen that extracts the content of the dialogue based on the viewpoint that the user focuses on, and the user can efficiently check the dialogue status by looking at the display screen 250. can.

Note that the output format of the extraction results by the extraction result display section 116 is not limited to the screen configuration of the display screen 250 in FIG. 7 . In addition, for example, text data of utterances in the dialogue scenario may be displayed in a region of the display screen 250 in the form of a real-time reception chat. In this case, it is displayed more clearly what kind of response the conversation with the viewpoint of interest occurred in.

After step S109, the utterance information input unit 111 adds the utterance information input in step S101 to the past utterance information group of the current dialogue scenario stored in the past utterance information storage unit 123 (step S110). ), the dialogue information extraction process ends. Note that the execution timing of step S110 for registering the utterance information may be any timing after step S101.

By executing the dialog information extraction process as described above, the dialog information extraction device 100 estimates the dialog structure and determines the viewpoint from the utterance sequences of a plurality of speakers, and sets an utterance set including utterances having a viewpoint. is extracted, and dialogue information regarding the utterance set is output. According to the dialogue information extraction device 100 according to the present embodiment, it is possible to extract dialogue information having a noteworthy viewpoint from dialogues between a plurality of speakers, and it is possible to efficiently check the dialogue situation. becomes.

In addition, the dialogue information extraction device 100 can accurately determine not only the presence or absence of a viewpoint but also the content of the viewpoint (for example, whether it is agreed or disagreed) by making a determination regarding the viewpoint in units of utterance sets. The granularity of extracted dialogue information can be improved.

Furthermore, the dialog information extraction device 100 filters and outputs the extracted dialog information using the filter conditions specified by the user, so that the user can check the dialog status more efficiently.

Furthermore, the dialogue information extraction device 100 can perform online processing by sequentially executing the dialogue information extraction process shown in FIG. 2 every time an utterance occurs (every time text utterance information is input). Become. At this time, if the screen display of the dialogue information extraction results is updated every time an utterance occurs (every time text utterance information is input), the user can check the dialogue status (for example, consent, etc.) in real time. It is possible to check the status of viewpoints such as disagreement. Being able to extract information related to a predetermined viewpoint while continuously monitoring speech status online is useful for reducing risks such as failure to confirm consent.

(2) Second Embodiment FIG. 8 is a block diagram showing a configuration example of a dialogue information extraction device 101 according to a second embodiment of the present invention. The dialogue information extraction device 101 has a configuration in which a determination target selection unit 117 is added to the dialogue information extraction device 100 shown in FIG. 1 in the first embodiment, and the configuration common to the dialogue information extraction device 100 is The explanation will be omitted.

The determination target selection unit 117 determines whether to extract dialogue information from a series of dialogues (a plurality of utterances) in a dialogue scenario based on preset determination conditions (which may be specified by the user at any timing). It has a function of selecting a target utterance (judgment target utterance).

As the above-mentioned determination conditions, for example, a "start keyword utterance" for determining the start of the dialogue information extraction process and an "end keyword utterance" for determining the end of the dialogue information extraction process are set. At this time, the determination target selection unit 117 determines whether the "start keyword utterance" or the "end keyword utterance" is the utterance text data (utterance information) that is input to the dialogue information extraction device 101 as the dialogue scenario progresses. A search is made to see if the utterance exists, and the utterances from the utterance in which the start keyword utterance first exists to the utterance in which the end keyword utterance first exists are selected as determination target utterances. In addition, as another judgment condition, for example, if one of the speakers (operator, etc.) is configured to press a predetermined switch at the start of the dialogue section to be judged, this switch operation can be used as a judgment condition. can.

FIG. 9 is a flowchart illustrating an example of a procedure for dialog information extraction processing in the second embodiment. In FIG. 9, processes similar to the dialog information extraction process in the first embodiment shown in FIG. 2 are given the same reference numerals. The processing procedure of the dialog information extraction process shown in FIG. 9 will be described below, focusing on the differences from the dialog information extraction process of FIG. 2.

According to FIG. 9, first, the utterance information input unit 111 inputs utterance information (step S101). Next, the determination target selection unit 117 determines whether the utterance information input in step S101 is a determination target utterance based on a predetermined determination condition (step S201).

If it is determined in step S201 that the utterance is not a determination target utterance (NO in step S201), there is no need to extract dialogue information, so the dialogue information extraction process ends without performing any special processing.

On the other hand, if it is determined in step S201 that the utterance is a determination target utterance (YES in step S201), the utterance information input unit 111 inputs the past utterance information group of the current dialogue scenario stored in the past utterance information storage unit 123. It is checked whether there is past utterance information that satisfies the criteria for the utterance to be determined (step S202). Note that in the second embodiment, the past utterance information storage unit 123 stores only past utterance information that satisfies the determination condition of the determination target utterance. Therefore, in reality, in step S202, the utterance information input unit 111 only needs to check whether past utterance information exists in the past utterance information group of the current dialogue scenario.

If there is past utterance information that satisfies the criteria for the determination target utterance in step S202 (YES in step S202), the processes in steps S103 to S108 are performed, so that the utterance set including the utterance with a viewpoint is Dialogue information regarding the utterance set determined as the extraction target is stored in the extraction result storage unit 125. Next, the extraction result display unit 116 outputs, as an extraction result, dialogue information that corresponds to the filter condition specified by the user from among the dialogue information stored in the extraction result storage unit 125 (step S109). Finally, in step S110, the utterance information input unit 111 adds the utterance information input in step S101 to the past utterance information group of the current dialogue scenario stored in the past utterance information storage unit 123 (step S110), the dialog information extraction process ends.

On the other hand, if there is no past utterance information that satisfies the determination condition for the determination target utterance in step S202 (YES in step S202), no dialogue from which dialogue information is to be extracted is currently formed in the dialogue scenario. Therefore, the utterance information input unit 111 stores the utterance information input in step S101 in the past utterance information storage unit 123 as a group of past utterance information of the corresponding dialogue scenario (step S110), and ends the dialogue information extraction process. Note that if the utterance information input in step S101 relates to the first utterance of a new dialogue scenario, the utterance information input unit 111 stores the utterance information in the past utterance information storage unit as a past utterance information group of the new dialogue scenario. 123.

By executing the dialogue information extraction process as described above, the dialogue information extraction device 101 according to the second embodiment can extract dialogue information having a viewpoint while limiting the determination target utterances. Compared to the first embodiment, the processing load on the processing units (for example, the dialogue structure estimation unit 112 and the viewpoint-based dialogue extraction unit 113) is reduced, and the processing load on the storage units (for example, the past utterance information storage unit 123 and the dialogue structure storage unit 124) is reduced. etc.) capacity consumption can be suppressed. In addition, by appropriately setting judgment conditions such as "start keyword utterance" and "end keyword utterance", it is possible to exclude utterances that are unrelated to the viewpoint, thereby increasing the granularity of the dialogue information extracted. You can get the effect of improving.

(3) Hardware Configuration Finally, an example of the hardware configuration of the dialog information extraction apparatus 100, 101 according to the first or second embodiment will be described.

FIG. 10 is a block diagram showing an example of the hardware configuration of the dialog information extraction devices 100 and 101. As shown in FIG. 10, the dialogue information extraction devices 100 and 101 are configured to include a CPU 11, a ROM 12, a RAM 13, an input/output interface 14, a storage device 15, a drive device 16, and a communication interface 17. The storage device 15 stores various data and programs 20 used by the dialogue information extraction devices 100 and 101. By the CPU 11 reading out the program 20 into the RAM 13 and executing it, each processing unit of the dialogue information extraction devices 100 and 101 (utterance information input unit 111, dialogue structure estimation unit 112, viewpoint-based dialogue extraction unit 113, extraction result display unit 116) , determination target selection unit 117) is realized.

11 CPU
12 ROM
13 RAM
14 input/output interface 15 storage device 16 drive device 17 communication interface 20 program 100, 101 dialogue information extraction device 111 utterance information input unit 112 dialogue structure estimation unit 113 viewpoint-based dialogue extraction unit 114 utterance set generation unit 115 viewpoint determination unit 116 extraction result Display section 117 Judgment target selection section 121 Dialogue structure estimation model storage section 122 Viewpoint determination model storage section 123 Past utterance information storage section 124 Dialogue structure storage section 125 Extraction result storage section 210 Utterance information 220 Past utterance information group 230 Dialogue structure data 240 Dialogue Information 250 Display screen

Claims (11)

A dialogue information extraction device that analyzes dialogue between multiple speakers based on a viewpoint to be focused on,
an utterance information input unit that inputs, for each utterance constituting the dialogue, utterance information including text transcribed from the utterance;
a dialogue structure estimation unit that estimates a dialogue structure of the dialogue from the utterance information of a plurality of utterances in the dialogue using an estimation model that has learned the dialogue structure;
a viewpoint-specific dialogue extraction unit that extracts a group of utterances having the viewpoint by determining the presence or absence of the viewpoint from a group of utterances generated based on the dialogue structure and consisting of utterances of a plurality of speakers;
A dialogue information extraction device comprising: Regarding the utterance group extracted by the viewpoint-specific dialogue extracting unit, the extraction result display unit outputs dialogue information including the text of each utterance constituting the utterance group and the content of the viewpoint that the utterance group has. The dialog information extraction device according to claim 1. The dialogue structure estimation unit calculates the degree of relationship between two utterances by different speakers using the estimation model for a plurality of utterances in the dialogue, and determines that there is a dialogue relationship based on the result of the calculation. The dialogue information extraction device according to claim 1, wherein the dialogue structure is estimated by estimating a combination of utterances. The viewpoint-specific dialogue extraction unit is
an utterance set generation unit that generates an utterance set by narrowing down the combinations of utterances estimated by the dialogue structure estimation unit under predetermined conditions as an utterance group consisting of utterances of the plurality of speakers;
and a viewpoint determination unit that determines, for each of the utterance sets, whether or not the viewpoint exists in the utterance set using a previously learned determination model, and extracts a utterance set having the viewpoint from the result of the determination. The dialog information extraction device according to claim 3, characterized in that: The utterance set generation unit generates, as the utterance set, combinations of utterances that have a dialogical relationship at or above an arbitrary specified probability value among the combinations of utterances estimated by the dialogue structure estimation unit. The dialog information extraction device according to claim 4. The utterance set generation unit selects a predetermined number of combinations of utterances from the answering side and utterances from the questioner that are past the utterance from among the combinations of utterances estimated by the dialogue structure estimating unit, starting from the one with the smallest time difference. The dialogue information extraction device according to claim 4, wherein each is generated as the utterance set. The extraction result display unit extracts a group of utterances that match a filter condition specified by the user from among the utterance groups having the viewpoint extracted by the viewpoint-based dialogue extracting unit, and displays information about the extracted utterance group. The dialogue information extraction device according to claim 2, wherein the dialogue information is output. Each time an utterance is made as the dialogue progresses, the utterance information input unit inputs the utterance information of the utterance,
In response to the input of the utterance information, processing is performed by the dialogue structure estimation unit and the viewpoint-based dialogue extraction unit, and the extraction result display unit updates the output of the dialogue information based on the results of the processing. The dialogue information extraction device according to claim 2, characterized in that: With respect to the utterance information that the utterance information input unit attempts to input, based on a predetermined determination condition, whether or not the utterance information is to be processed by the dialogue structure estimation unit and the viewpoint-based dialogue extraction unit. The dialog information extraction device according to claim 1, further comprising: a determination target selection unit that determines. The dialogue information extraction device according to claim 1, wherein the utterance of the answering side indicating agreement or disagreement with the utterance of the questioner is regarded as having the viewpoint. A dialogue information extraction method using a dialogue information extraction device that analyzes dialogues between multiple speakers based on noteworthy viewpoints, the method comprising:
an utterance information input step in which the dialogue information extraction device inputs utterance information including text transcribed from the utterance for each utterance constituting the dialogue;
A dialogue structure in which the dialogue information extraction device estimates the dialogue structure of the dialogue from the utterance information of a plurality of utterances in the dialogue input in the utterance information input step, using an estimation model whose dialogue structure has been learned. an estimation step;
The dialogue information extraction device has the viewpoint by determining whether or not the viewpoint exists for a group of utterances made up of utterances of a plurality of speakers, which is generated based on the estimation result of the dialogue structure estimation step. a perspective-specific dialogue extraction step for extracting utterance groups;
A dialogue information extraction method characterized by comprising:

Images