JP7453714B2 - Argument analysis device and method - Google Patents

Description

The present invention relates to a discussion analysis device and a discussion analysis method for analyzing discussions by a plurality of participants.

The Harkness method (also referred to as the Harkness method) is known as a method for analyzing discussions in group studies and meetings (see, for example, Non-Patent Document 1). In the Harkness method, the transitions of participants (speakers) who have made statements in a discussion are recorded with lines. This makes it possible to analyze each participant's contribution to the discussion and their relationship with others. The Harkness method can also be effectively applied to active learning, where students learn independently.

Paul Sevigny, "Extreme Discussion Circles: Preparing ESL Students for "The Harkness Method"," Polyglossia, Ritsumeikan Asia Pacific University Language Education Center, October 2012, No. 23, p. 181-191

The Harkness method requires the recorder to constantly record discussions, which places a heavy burden on the recorder. Therefore, it is conceivable to automatically detect speaker transitions by acquiring voices uttered by participants using a sound collection device and analyzing the voices using a computer. However, the computer detects irregular sounds such as the collision of objects or the chiding of other participants while the participant is speaking as the participant's utterances, and the computer correctly detects the transition of the speaker. It may not be possible.

The present invention has been made in view of these points, and an object of the present invention is to improve the accuracy of detecting speaker transitions in discussions.

A discussion analysis device according to a first aspect of the present invention includes an information acquisition unit that acquires the amount of speech of each of the plurality of participants in a discussion in which the plurality of participants participate; a maximum speaker identification unit that identifies a maximum speaker with the maximum amount of speech among the plurality of participants; and an output unit that outputs transition information indicating a transition of speakers that has occurred between speakers.

When the first participant who is the largest speaker in one time range is different from the second participant who is the largest speaker in a time range following the one time range, the output unit The transition information indicating the transition from the first participant to the second participant may be output.

The discussion analysis device may further include a phase division unit that divides the discussion into one or more phases based on chronological similarity of the transition information.

The output unit outputs the transition information indicating the number of transitions for each second time range that is longer than the first time range, and the phase division unit outputs the transition information indicating the number of transitions for each second time range that is longer than the first time range, and the phase dividing unit clustering the transition information for each second time range based on the time under discussion in the second time range corresponding to the transition information included in the generated plurality of clusters; The one or more phases that constitute the discussion may be determined.

The discussion analysis device generates a plurality of patterns indicating the presence or absence of the transition for each combination of the plurality of participants, and selects, among the plurality of patterns, a pattern whose degree of similarity with the transition information satisfies a predetermined condition. It may further include a pattern selection section for selection.

The pattern selection unit further generates a plurality of sub-patterns in which a part of the selected pattern is changed, and selects a sub-pattern whose degree of similarity with the transition information satisfies a predetermined condition from among the plurality of sub-patterns. You may.

The output unit determines the roles of the plurality of participants based on the pattern selected by the pattern selection unit, and outputs each of the plurality of participants in association with the role of each of the plurality of participants. Good too.

The output unit may output the actions of the plurality of participants as sentences based on the pattern selected by the pattern selection unit.

The output unit is configured to output information regarding the amount of speech by the one participant in the discussion that satisfies a predetermined condition from among the plurality of discussions in which one of the plurality of participants participated. It may also be output in association with other participants.

The output unit may output information regarding the amount of speech of the plurality of participants in the plurality of discussions in which the plurality of participants belonging to a predetermined group participated, in association with the group.

The output unit is configured to output a ranking of the amount of utterances of the plurality of participants belonging to the group in a first discussion, and utterances of the plurality of participants belonging to the group in a second discussion different from the first discussion. It may also be output in association with the order of quantity.

A discussion analysis method according to a second aspect of the present invention includes the steps of: acquiring the amount of speech of each of the plurality of participants in a discussion in which the plurality of participants participate, which is executed by a processor; a step of identifying a maximum speaker who has the maximum amount of speech among the plurality of participants for each time range; and outputting transition information indicating the speaker transition that occurred between.

According to the present invention, there is an effect that the detection accuracy of speaker transition in a discussion is improved.

1 is a schematic diagram of a discussion analysis system according to an embodiment. FIG. 1 is a block diagram of a discussion analysis system according to an embodiment. FIG. 2 is a schematic diagram of how the discussion analysis device detects speaker transitions in a discussion. 2 is a schematic diagram of how a discussion analyzer divides a discussion into one or more phases; FIG. FIG. 6 is a schematic diagram of a method in which the discussion analysis device selects a pattern similar to transition information. FIG. 6 is a schematic diagram of a method in which the discussion analysis device selects a pattern similar to transition information. FIG. 3 is a front view of the display unit displaying a discussion report screen. FIG. 3 is a front view of the display section displaying a personal report screen. FIG. 3 is a front view of the display unit displaying a course report screen. FIG. 3 is a diagram showing a flowchart of a discussion analysis method performed by the discussion analysis device.

[Overview of discussion analysis system SS]
FIG. 1 is a schematic diagram of a discussion analysis system SS according to this embodiment. The discussion analysis system SS includes a discussion analysis device 1, a communication terminal 2, and a sound collection device 3. The number of communication terminals 2 and sound collection devices 3 included in the discussion analysis system SS is not limited. The discussion analysis system SS may include other devices such as servers and terminals.

The sound collecting device 3 includes a microphone array including a plurality of sound collecting sections (microphones) arranged in different directions. For example, a microphone array includes eight microphones arranged at equal intervals on the same circumference in a horizontal plane relative to the ground. By using such a microphone array, the discussion analysis device 1 can determine which participant U is the speaker (sound source) based on the voices emitted by the multiple participants U surrounding the sound collection device 3. can be identified. The sound collection device 3 transmits the voice acquired using the microphone array to the discussion analysis device 1 as data.

The communication terminal 2 is a computer capable of communicating. The communication terminal 2 is, for example, a computer terminal such as a personal computer, or a mobile terminal such as a smartphone. The communication terminal 2 sets analysis conditions for the discussion analysis device 1 and displays information received from the discussion analysis device 1.

The discussion analysis device 1 is a computer that analyzes discussions using the audio acquired by the sound collection device 3. The discussion analysis device 1 is configured by, for example, a single computer or a cloud that is a collection of computer resources.

The discussion analysis device 1 is connected to a communication terminal 2 and a sound collection device 3 by wire or wirelessly via a network N such as a local area network or the Internet. The discussion analysis device 1 may be directly connected to at least one of the communication terminal 2 and the sound collection device 3 without using the network N.

An outline of the processing executed by the discussion analysis device 1 will be explained below. First, the discussion analysis device 1 acquires audio in a discussion in which a plurality of participants U participate from the sound collection device 3. The discussion analysis device 1 uses the obtained voices to obtain the amount of speech of each of the plurality of participants U in the discussion. The discussion analysis device 1 identifies the participant U with the largest amount of speech (that is, the largest speaker) for each predetermined time range. The discussion analysis device 1 then outputs transition information indicating the transition of speakers that has occurred among the plurality of participants U, based on the time-series changes in the maximum speaker for each predetermined time range.

According to the discussion analysis system SS according to the present embodiment, the discussion analysis device 1 detects the transition of the speaker based on the change of the participant U who has the largest amount of speech, so the discussion analysis device 1 It is possible to suppress the detection of speaker transitions due to sounds that are not utterances, such as exchanges, and improve the accuracy of detecting speaker transitions in discussions.

[Configuration of discussion analysis system SS]
FIG. 2 is a block diagram of the discussion analysis system SS according to this embodiment. In FIG. 2, arrows indicate main data flows, and there may be data flows that are not shown in FIG. In FIG. 2, each block shows the configuration of a functional unit rather than a hardware (device) unit. As such, the blocks shown in FIG. 2 may be implemented within a single device or may be implemented separately within multiple devices. Data may be exchanged between blocks via any means such as a data bus, a network, or a portable storage medium.

The discussion analysis device 1 includes a control section 11 and a storage section 12. The control unit 11 includes an information acquisition unit 111 , a maximum speaker identification unit 112 , a transition detection unit 113 , a phase division unit 114 , a pattern selection unit 115 , and an output unit 116 . The storage unit 12 includes a discussion information storage unit 121 and a participant information storage unit 122.

The storage unit 12 is a storage medium including a ROM (Read Only Memory), a RAM (Random Access Memory), a hard disk drive, and the like. The storage unit 12 stores in advance a program to be executed by the control unit 11. The storage unit 12 may be provided outside the discussion analysis device 1, and in that case, data may be exchanged with the control unit 11 via a network.

The discussion information storage unit 121 stores discussion information indicating information regarding discussions. The participant information storage unit 122 stores participant information indicating information regarding participants participating in the discussion. The discussion information storage section 121 and the participant information storage section 122 may each be a storage area on the storage section 12 or may be a database configured on the storage section 12.

The control unit 11 is, for example, a processor such as a CPU (Central Processing Unit), and executes a program stored in the storage unit 12 to control the information acquisition unit 111, the maximum speaker identification unit 112, the transition detection unit 113, and the phase detection unit 112. It functions as a dividing section 114, a pattern selecting section 115, and an output section 116. At least some of the functions of the control unit 11 may be performed by an electric circuit. Furthermore, at least part of the functions of the control unit 11 may be executed by a program executed via a network.

Communication terminal 2 includes a control section 21, a storage section 22, and a display section 23. The control section 21 has a receiving section 211. The display unit 23 includes a display device capable of displaying information, such as a liquid crystal display. A touch screen that can detect the position of contact by a human may be used as the display unit 23.

The storage unit 22 is a storage medium including ROM, RAM, hard disk drive, and the like. The storage unit 22 stores in advance a program to be executed by the control unit 21. The storage unit 22 may be provided outside the communication terminal 2, and in that case, data may be exchanged with the control unit 21 via a network.

The control unit 21 is, for example, a processor such as a CPU, and functions as the reception unit 211 by executing a program stored in the storage unit 22. At least some of the functions of the control unit 21 may be performed by an electric circuit. Furthermore, at least part of the functions of the control unit 21 may be executed by a program executed via a network.

The discussion analysis device 1 and the communication terminal 2 according to this embodiment are not limited to the specific configuration shown in FIG. 2. The discussion analysis device 1 and the communication terminal 2 are not each limited to one device, but may be configured by two or more physically separate devices connected by wire or wirelessly.

[Explanation of argument analysis method]
A discussion analysis method performed by the discussion analysis device 1 according to this embodiment will be described below. A plurality of participants are seated surrounding one sound collection device 3 when having a discussion. A participant in the discussion or an analyst who analyzes the discussion sets analysis conditions by operating the communication terminal 2. For example, the analysis conditions are information indicating the number of participants in the discussion to be analyzed and the direction in which each of the plurality of participants is located with respect to the sound collection device 3 (that is, the relative position). In the discussion analysis device 1, the information acquisition unit 111 receives the set analysis conditions from the communication terminal 2, and stores them in the discussion information storage unit 121 in association with identification information (for example, discussion ID) for identifying the discussion. A discussion ID may be automatically assigned to a discussion or may be entered by a participant or analyst.

Next, when starting a discussion, the participant or analyst instructs the start of the discussion by operating the communication terminal 2. In the discussion analysis device 1 , when the information acquisition unit 111 receives a signal instructing to start a discussion from the communication terminal 2 , it transmits a signal instructing to acquire audio to the sound collection device 3 . When the sound collection device 3 receives a signal instructing the acquisition of audio from the discussion analysis device 1, it starts acquiring audio.

The sound collection device 3 acquires sounds from each of the plurality of sound collection parts, and internally records the sounds as sounds of each channel corresponding to each sound collection part. Then, the sound collection device 3 transmits the acquired voices of the plurality of channels to the discussion analysis device 1. The sound collection device 3 may transmit the acquired sounds one after another, or may transmit a predetermined amount or a predetermined amount of time. Further, the sound collection device 3 may transmit all sounds from the start to the end of acquisition all at once. In the discussion analysis device 1, the information acquisition unit 111 receives audio from the sound collection device 3, and stores it in the discussion information storage unit 121 in association with the discussion ID.

When ending the discussion, the participant or analyst instructs the end of the discussion by operating the communication terminal 2. In the discussion analysis device 1, when the information acquisition unit 111 receives a signal instructing the end of the discussion from the communication terminal 2, it transmits a signal instructing the end of audio acquisition to the sound collection device 3. When the sound collection device 3 receives a signal instructing the end of audio acquisition from the discussion analysis device 1, it ends the audio acquisition.

The subsequent processing is performed when the acquisition of the voice is completed or when the analyst gives a predetermined instruction to the communication terminal 2. Further, the subsequent processing may be performed sequentially, triggered by the start of audio acquisition. The information acquisition unit 111 performs sound source localization based on multiple channels of audio received from the sound collection device 3. Sound source localization is a process of estimating the direction of a sound source included in the sound acquired by the information acquisition unit 111 at each time (for example, every 10 to 100 milliseconds). The information acquisition unit 111 associates the direction of the sound source estimated for each time with the direction of each of the plurality of participants indicated by the analysis conditions stored in the discussion information storage unit 121.

The information acquisition unit 111 can use a known sound source localization method, such as the MUSIC (Multiple Signal Classification) method or the beamforming method, as long as the direction of the sound source can be specified based on the acquired audio.

Next, the information acquisition unit 111 determines which participant makes an utterance (utterance) at a predetermined time interval (for example, every 10 to 100 milliseconds) during the discussion based on the acquired voice and the estimated direction of the sound source. Determine whether The information acquisition unit 111 specifies a continuous period from when one participant starts speaking until it ends as a speaking period. When multiple participants speak at the same time, at least a portion of the speaking periods of the multiple participants overlap. The information acquisition unit 111 stores the utterance period specified in the discussion in the discussion information storage unit 121 in association with the discussion ID.

The information acquisition unit 111 may exclude a specific time range during the discussion and specify the utterance period. In this case, the participant or analyst performs a predetermined operation on the communication terminal 2 or the sound collection device 3 during the time range to be excluded during the discussion. The participant or the analyst may continue to perform operations such as button operations during the time range to be excluded, or may perform operations such as button operations at the beginning and end of the time range to be excluded, respectively. The communication terminal 2 or the sound collection device 3 transmits information indicating the time range to be excluded to the discussion analysis device 1.

When the information acquisition unit 111 receives information indicating a time range to be excluded, the information acquisition unit 111 causes the discussion information storage unit 121 to store audio obtained by excluding the time range from the acquired audio, and excludes the time range from the acquired audio. The utterance period is identified using the recorded audio. Thereby, the participant or the analyst can set a time range in which confidential matters are discussed so as not to be analyzed.

In the present embodiment, the information acquisition unit 111 specifies the speech period based on the sound acquired by the sound collection device 3, but the speech period may be specified using other methods. For example, the information acquisition unit 111 may identify a participant's speaking period based on the voice uttered by the participant during a voice call or a video call (also referred to as a video conference or a video chat). Further, for example, the information acquisition unit 111 may read and acquire the speech period stored in advance in the storage unit 12.

Further, for example, the information acquisition unit 111 may identify the speaking period of the participant based on an image including the face of the participant in the discussion. In this case, an imaging device is placed near the plurality of participants having a discussion instead of or in addition to the sound collection device 3. The information acquisition unit 111 acquires time-series images including the faces of a plurality of participants captured by an imaging device during a discussion. Furthermore, the information acquisition unit 111 may acquire time-series images including the faces of multiple participants that are transmitted and received between communication terminals in a video call. The information acquisition unit 111 applies known face recognition processing to the acquired images to determine whether each of the plurality of participants is speaking based on the state of the human face (for example, whether the mouth is open or not). The speech period of each of the plurality of participants is determined.

Next, a method for the discussion analysis device 1 to detect a change in speakers in a discussion will be explained. FIG. 3 is a schematic diagram of a method by which the discussion analysis device 1 detects speaker transitions in a discussion. The information acquisition unit 111 acquires the amount of time-series utterances (also referred to as utterance amount) of each of the plurality of participants in the discussion based on the specified utterance period.

Specifically, the information acquisition unit 111 divides the discussion into a first frame (ie, a first time range) of a predetermined window width w1 (for example, 30 seconds). The first frames are shifted by a predetermined shift width s1 (for example, 10 seconds) shorter than the window width w1, and adjacent first frames partially overlap each other in time series.

The information acquisition unit 111 then calculates a value obtained by dividing the length of the participant's speech period (total speech time) in the first frame by the window width w1 as the speech amount for each first frame. The information acquisition unit 111 calculates the amount of speech for each first frame from the start time to the end time of the discussion for each of the plurality of participants. The information acquisition unit 111 causes the discussion information storage unit 121 to store information indicating the amount of speech for each first frame of each of the plurality of participants in the discussion in association with the discussion ID.

The upper diagram in FIG. 3 shows a graph G of the amount of speech by a plurality of participants over time. Graph G represents the amount of speech by a plurality of participants as a stacked graph. The horizontal axis of the graph G is time, and the vertical axis is the amount of speech. In the area of graph G, different patterns are displayed depending on each of the plurality of participants.

Further, the information acquisition unit 111 calculates the amount of interruption and the amount of excitement for each of the plurality of participants based on the acquired speaking period and amount of speaking. Specifically, when the utterance periods of two participants overlap each other in chronological order, the information acquisition unit 111 calculates the length of the overlapping utterance period between the two participants. It is calculated as the amount of interruptions for the participant whose speaking period starts later. The information acquisition unit 111 calculates the amount of interruptions by each of the plurality of participants from the start to the end of the discussion.

In addition, the information acquisition unit 111 totals the amount of speech of all the participants in a predetermined time (for example, 20 seconds) before and after one speech period of one participant, and The amount obtained by subtracting the total amount of speech before the speech period from the total speech amount (that is, the increment in the total speech amount from before to after the speech period) is calculated as the excitement amount. The information acquisition unit 111 calculates the amount of excitement for each of the plurality of participants from the start to the end of the discussion. The information acquisition unit 111 may calculate the number of speech periods in which the amount of excitement is greater than 0 out of all the number of speech periods of one participant as the number of times of excitement. The information acquisition unit 111 causes the discussion information storage unit 121 to store the amount of interruptions and the amount of excitement (or the number of times of excitement) of each of the plurality of participants in association with the discussion ID.

The largest speaker identification unit 112 identifies the largest speaker with the largest amount of speech among the plurality of participants for each first frame, based on the amount of speech acquired by the information acquisition unit 111. The maximum speaker identification unit 112 outputs an array of maximum speakers from the first frame at the beginning of the discussion to the last first frame.

The middle diagram in FIG. 3 shows the maximum speaker in the time series as band S. The time-series maximum speaker band S is generated based on the time-series speech volume graph G, and the horizontal axis corresponds to the time of the graph G. The band S of the largest speaker shows different patterns depending on each of the plurality of participants, and corresponds to the pattern of the area of the graph G.

The transition detection unit 113 detects a transition in speakers that occurs between a plurality of participants, based on the change in the maximum speaker for each first frame specified by the maximum speaker identification unit 112. Specifically, the discussion is divided into a second frame (ie, a second time range) of a predetermined window width w2. The window width w2 of the second frame is longer than the window width w1 of the first frame. That is, the second frame includes a plurality of first frames. The window width w2 may be defined as a predetermined multiple of the window width w1 (for example, 100 times the window width w1), or may be defined as a predetermined time (for example, 3000 seconds).

The second frames are shifted by a predetermined shift width s2 that is shorter than the window width w2, and parts of two adjacent second frames overlap each other in time series. The shift width s2 may be defined as a predetermined multiple of the window width w1 (for example, 5 times the window width w1), or may be defined as a predetermined time (for example, 150 seconds).

The transition detection unit 113 detects when the first participant who is the largest speaker in one first frame is different from the second participant who is the largest speaker in the first frame following the first frame. , detecting a transition from the first participant to the second participant. The transition detection unit 113 repeats transition detection for one second frame from the first frame to the last frame of the second frame, and detects the combination of participants (i.e., the first participant and the first frame). A transition matrix indicating the number of transitions detected for each combination of participants (2) is generated. When the number of participants is D, the transition matrix becomes a D×D matrix.

Further, the transition detection unit 113 repeatedly generates the transition matrix from the first second frame of the discussion to the last second frame. When the number of second frames is N, the transition detection unit 113 generates N transition matrices. The transition detection unit 113 causes the discussion information storage unit 121 to store information indicating the transition matrix generated for each second frame as transition information.

The lower diagram of FIG. 3 shows an exemplary transition matrix M. In the example of FIG. 3, there are three participants, U1, U2, and U3, and a plurality of transition matrices M are generated based on the band S of the largest speaker in the time series. The rows of the transition matrix M indicate the transition source participants, and the columns indicate the transition destination participants. In this way, the discussion analysis device 1 detects the transition of speakers based on the change in the number of maximum speakers, and therefore detects the transition of speakers based on sounds that are not utterances, such as the sound of objects colliding or participants agreeing. This makes it possible to improve the accuracy of detecting speaker transitions in discussions.

Next, a method in which the discussion analysis device 1 divides the discussion into one or more phases will be explained. FIG. 4 is a schematic diagram of how the discussion analysis device 1 divides the discussion into one or more phases. The phase dividing unit 114 divides the discussion into one or more phases based on the similarity in time series of transition information (transition matrix) generated for each second frame. Here, the phase dividing unit 114 divides the transition information so that the transition information is similar within one phase, that is, the similarity between the transition information of two second frames included in one phase is included in two different phases. Divide the discussion into one or more phases such that the similarity between the transition information of the two second frames is higher. The phase division unit 114 uses a known method capable of dividing the discussion into one or more phases based on the similarity of the time series of transition information.

For example, the phase division unit 114 divides the discussion into one or more phases by performing clustering using a Poisson mixture model described below. First, the phase division section 114 obtains the transition matrix generated by the transition detection section 113. Here, for calculation, the phase division unit 114 arranges the elements of the transition matrix for each second frame in a vertical line, thereby dividing the number of transitions for each combination of participants into a D ² ×1 vertical column. Convert to vector. As a result, the phase division unit 114 obtains a D ² ×N matrix in ^which N two-dimensional non-negative vectors are arranged in time series.

Since each combination of participants has a different tendency of transition, the D ² ×N matrix obtained by transforming the transition matrix becomes a mixture distribution of Poisson distribution shown in Equation (1).

Here, Poi represents a function of Poisson distribution, x represents the number of times a transition occurs for each combination of participants (i.e., each element of the transition matrix), and λ _d represents the average number of times a transition occurs for each combination of participants. , and d represents the dimension (1 to D ² ) of the vertical vector.

Considering that the discussion is divided into K clusters (K is a predetermined number of 2 or more), K sets of the above-mentioned values of λ _d can be created. Thereby, the phase division unit 114 generates a mixed distribution of K Poisson distributions as shown in Equation (2).

Here, the phase division unit 114 defines a hidden variable z _nk (z _nk is 0 or 1), which is a matrix indicating to which cluster the n-th transition matrix belongs among the N transition matrices. do. The hidden variable z _nk is 1 only when the n-th transition matrix belongs to the k-th cluster, and is 0 otherwise.

Thereby, the phase dividing unit 114 combines the distribution of equation (2) into one distribution shown in equation (3).

The phase division unit 114 performs Bayesian estimation using the model of Equation (3) to determine the parameters λ and z where the probability of x satisfies a predetermined condition (for example, the probability of x is the maximum value). Calculate. Thereby, the phase division unit 114 determines to which of the K clusters each of the N transition matrices is assigned.

At this time, the phase division unit 114 may delete clusters in which the number of assigned transition matrices is equal to or less than a predetermined threshold value from among the K clusters. In this case, the transition matrix assigned to the deleted cluster is assigned to the cluster before or after the deleted cluster. As a result, the number of clusters finally generated is K or less. Thereby, the phase division unit 114 can divide the discussion into one or more phases, leaving only clusters that have been assigned a large number of transition matrices.

In this embodiment, the phase division unit 114 allocates multiple transition matrices to multiple clusters without considering the time series, so theoretically the time series of the multiple transition matrices and the time series of the clusters are different. They may not match. However, since the transition detection unit 113 shifts the second frames while overlapping them in time series, the number of detected transitions is a moving average in time series. Therefore, multiple transition matrices that are close in time are similar to each other. As a result, under normal circumstances, the time series of the plurality of transition matrices and the time series of the plurality of clusters generated by the phase division unit 114 match.

When a plurality of clusters are generated, the phase division unit 114 divides the discussion into a plurality of phases based on the time during discussion of the second frame corresponding to the transition matrix included in each of the plurality of clusters. Specifically, the phase division unit 114 specifies the end time of the last second frame among the second frames corresponding to the transition matrix included in one cluster as the end time of the phase. Determine the multiple phases that make up the process.

Furthermore, when one cluster is generated, the phase dividing unit 114 determines the entire discussion as one phase. The phase dividing unit 114 causes the discussion information storage unit 121 to store information indicating the determined phase of the discussion in association with discussion identification information.

If the discussion is simply divided into phases such as the first half, middle, and second half based on time, the content of the discussion will not be taken into account, and the positions at which the discussion will be divided will not correspond to the actual situation. On the other hand, the discussion analysis device 1 according to the present embodiment divides the discussion into one or more phases based on the similarity of the time series of transition information, so that the discussion can be divided into units that match the actual situation.

The bottom diagram of FIG. 4 shows the phases of the discussion determined by the exemplary phase divider. In the example of FIG. 4, the discussion is divided into three phases: PH1, PH2, and PH3. The trends of speaker transitions are similar in each of phases PH1, PH2, and PH3. The discussion may be divided into phases other than three.

Next, a method for the discussion analysis device 1 to select a pattern similar to transition information will be explained. 5 and 6 are schematic diagrams of a method in which the discussion analysis device 1 selects a pattern similar to transition information. First, the pattern selection unit 115 obtains the transition matrix M (transition information) generated by the transition detection unit 113 for each discussion phase determined by the phase division unit 114. The transition matrix M for each phase may be, for example, a statistical value (average value, median value, etc.) of the transition matrix M included in the phase, or it may be a statistical value (average value, median value, etc.) of the transition matrix M included in the phase, or it may be a statistical value (average value, median value, etc.) of the transition matrix M included in the phase. It may be a transition matrix M.

The pattern selection unit 115 generates a plurality of patterns indicating the presence or absence of a transition in each combination of a plurality of participants. Here, when there are or relatively many transitions in a combination of two participants, the two participants are said to be "connected," and when there are no or relatively few transitions, the two participants are said to be "connected." One participant described it as "not connected." The pattern selection unit 115 selects a pattern in which, assuming the number of multiple participants is D, a pattern in which one central person is connected to all the others, and a pattern in which i people (i = 2 to D) are connected to each other. D types of patterns consisting of the patterns are generated.

In the example of FIG. 5, the patterns generated by the pattern selection unit 115 are a pattern P1 in which one central person is connected to all the others, a pattern P2 in which two people are connected to each other, and a pattern P2 in which three people are connected to each other. It consists of patterns P3 that are connected to each other. Although not shown in FIG. 5, pattern P1 includes a pattern in which one central person is changed to U1, U2, and U3, and pattern P2 includes a pattern in which two people who are connected to each other are changed to two out of U1, U2, and U3. Contains patterns changed to all combinations of.

The pattern selection unit 115 generates a matrix for each of the plurality of generated patterns. The pattern matrix is a transition matrix in which elements of combinations of connected participants are set to 1 and elements of combinations of unconnected participants are set to 0. Furthermore, the pattern selection unit 115 normalizes each element of the transition matrix for each phase to a range of 0 to 1.

Then, the pattern selection unit 115 calculates the degree of similarity between the matrix of each of the plurality of generated patterns and the normalized transition matrix for each phase. The similarity is, for example, the distance between matrices, but other values may also be used. Then, the pattern selection unit 115 selects a pattern whose calculated degree of similarity satisfies a predetermined condition (for example, the distance between matrices is the minimum) from among the plurality of patterns. The pattern selection unit 115 selects a pattern for each of the one or more phases determined by the phase division unit 114.

Furthermore, the pattern selection unit 115 generates a plurality of sub-patterns by adding changes to the selected pattern for each phase. Specifically, in addition to the selected pattern itself, the pattern selection unit 115 selects a pattern in which any one connection included in the selected pattern is deleted, and one connection not included in the selected pattern. Generate the added pattern as a subpattern. The pattern selection unit 115 may generate a sub-pattern by adding other changes to the selected pattern.

FIG. 6 shows exemplary sub-patterns when pattern P1 is selected in FIG. In this case, the subpatterns generated by the pattern selection unit 115 are a subpattern SP1 that is the pattern P1 itself, a subpattern SP2 that is obtained by deleting one connection included in the pattern P1, and a subpattern SP2 that is not included in the pattern P1. It consists of a subpattern SP3 to which one connection is added. In FIG. 6, deleted connections are represented by broken lines, and added connections are represented by dashed lines. Subpattern SP2 includes a pattern in which another connection is deleted, and subpattern SP3 includes a pattern in which another connection is added.

The pattern selection unit 115 generates a matrix for each of the plurality of generated sub-patterns. The subpattern matrix is a transition matrix in which elements of combinations of connected participants are set to 1 and elements of combinations of unconnected participants are set to 0. Furthermore, the pattern selection unit 115 normalizes each element of the transition matrix for each phase to a range of 0 to 1.

Then, the pattern selection unit 115 calculates the degree of similarity between the matrix of each of the plurality of generated sub-patterns and the normalized transition matrix for each phase. The similarity is, for example, the distance between matrices, but other values may also be used. Then, the pattern selection unit 115 selects a subpattern whose calculated degree of similarity satisfies a predetermined condition (for example, the distance between matrices is the minimum) from among the plurality of subpatterns. The pattern selection unit 115 selects sub-patterns for each of the one or more phases determined by the phase division unit 114.

The pattern selection unit 115 causes the discussion information storage unit 121 to store information indicating the selected pattern and subpattern in association with discussion identification information. The pattern selection unit 115 may select only patterns and store them in the discussion information storage unit 121 without selecting sub-patterns.

If the transitions of speakers in a discussion are simply displayed as they are in a graph or the like, the interpretation of the trends in the transitions is left to the analyst, which may result in different interpretations depending on the analyst. On the other hand, the discussion analysis device 1 according to the present embodiment can automatically classify multiple participants according to transition trends by comparing and selecting transition information with patterns and sub-patterns, and can automatically classify multiple participants according to their transition tendencies. It becomes possible to automatically output the relationships between participants as text.

The output unit 116 outputs information regarding the discussion based on the information stored in the discussion information storage unit 121 by the information acquisition unit 111, transition detection unit 113, phase division unit 114, and pattern selection unit 115. For example, the output unit 116 outputs information regarding the discussion by displaying screens shown in FIGS. 7, 8, and 9 on the display unit 23 of the communication terminal 2.

The output unit 116 is triggered by the completion of the processing by the information acquisition unit 111, transition detection unit 113, phase division unit 114, and pattern selection unit 115, or when the analyst issues a predetermined instruction to the communication terminal 2. Taking this as an opportunity, display information for displaying information regarding the discussion is generated based on the information stored in the discussion information storage section 121 and transmitted to the communication terminal 2. The receiving unit 211 of the communication terminal 2 displays the screens shown in FIGS. 7, 8, and 9 on the display unit 23 based on the display information received from the discussion analysis device 1.

FIG. 7 is a front view of the display unit 23 displaying the discussion report screen A. Discussion report screen A is a screen that displays information regarding one discussion. Discussion report screen A includes summary information A1, participant information A2, phase information A3, and comprehensive evaluation information A4. The summary information A1 is information indicating the outline of the amount of speech in the discussion in chronological order, as well as the arrangement of a plurality of participants set as analysis conditions. The summary of the amount of speech is, for example, a character string representing the time range of the phase in which the total amount of speech by a plurality of participants is maximum.

Participant information A2 is information indicating participants who meet predetermined conditions. For example, participant information A2 represents, among the plurality of participants, the participant with the largest amount of speech, the participant with the largest amount of interruptions, and the participant with the largest amount of excitement (which may also be the number of times of excitement). Further, the participant information A2 represents the participant connected in the pattern selected by the pattern selection unit 115 as the person who became the center of the discussion.

The phase information A3 includes a time range A31 of a phase in the discussion and a role A32 of participants for each phase. The phase time range A31 is information that indicates the time range of each phase by an arrow superimposed on a cumulative graph of the amount of speech by a plurality of participants in the discussion.

Participant role A32 is information indicating the role of each of the plurality of participants determined based on the pattern selected by the pattern selection unit 115. A role is a behavior tendency of a participant in a discussion, for example, a leader or a follower.

Specifically, when displaying participant roles A32, the output unit 116 determines the roles of each of the plurality of participants for each discussion phase based on the pattern selected by the pattern selection unit 115. For example, the output unit 116 determines that the participant with the largest amount of speech among the multiple participants connected to each other in the pattern selected by the pattern selection unit 115 has the role of "leader", and the other participants are assigned the role of "leader". The role is determined to be "Follower". Furthermore, the output unit 116 determines that participants who are not connected in the pattern selected by the pattern selection unit 115 have “no role”. The output unit 116 may determine other roles based on the pattern selected by the pattern selection unit 115.

Then, the output unit 116 transmits information indicating the roles of each of the plurality of participants for each phase to the communication terminal 2. The receiving unit 211 of the communication terminal 2 displays the roles of the plurality of participants for each phase received from the discussion analysis device 1 as participant roles A32 near the time range A31 of the phase. In the example of FIG. 7, below the phase time range A31, a solid line is displayed in association with a participant determined to be a leader, and a broken line is displayed in association with a participant determined to be a follower. Participant role A32 may represent the participant role in other ways. Thereby, the analyst can know the roles of each of the plurality of participants automatically determined by the discussion analysis device 1 based on the pattern of transition trends.

The comprehensive evaluation information A4 is information that is generated based on the pattern selected by the pattern selection unit 115 and represents the behavior of the participants in each phase of the discussion as a sentence. Specifically, when displaying the comprehensive evaluation information A4, the output unit 116 acquires the pattern selected by the pattern selection unit 115. The output unit 116 then generates a sentence corresponding to the pattern based on a predetermined rule. The predetermined rule is a template that is predefined in the storage unit 12 and corresponds to a pattern.

For example, in the pattern selected by the pattern selection unit 115, when the participant U1 and the participant U2 are connected to each other and the amount of speech by the participant U1 is larger than the amount of speech by the participant U2, the output unit 116 outputs the following: The sentence "A discussion was held centered around U1, with U2 also participating." is generated. The storage unit 12 stores in advance templates corresponding to each pattern that can be generated by the pattern selection unit 115. The method of generating sentences based on the patterns shown here is an example, and the output unit 116 can use a known method that can generate sentences based on each pattern that the pattern selection unit 115 can generate.

Thereby, the analyst can know the relationships between the plurality of participants automatically generated by the discussion analysis device 1 based on the pattern of transition trends in the form of text, and it becomes easier to understand the relationships.

Furthermore, the output unit 116 may generate sentences based on sub-patterns in addition to the pattern selected by the pattern selection unit 115. For example, in the pattern selected by pattern selection section 115, participant U1, participant U2, and participant U3 are connected to each other, and in the subpattern selected by pattern selection section 115, participant U3 and participant U1 are connected to each other. If the connection is deleted, a sentence is generated saying "All members participated in the discussion. The exchange of comments was mainly centered around U1 and U2."

Thereby, the analyst can know the relationships among the plurality of participants, which are automatically generated by the discussion analysis device 1 based on sub-patterns obtained by further subdividing the transition tendency pattern, as sentences.

FIG. 8 is a front view of the display unit 23 displaying the personal report screen B. Personal report screen B is a screen that displays information regarding multiple discussions in which one participant to be displayed has participated in the past. The personal report screen B includes participant trend information B1, participant progress information B2, and discussion information B3.

The participant tendency information B1 is information representing the tendency of the participant to be displayed as a sentence generated based on the characteristics of the participant to be displayed and the behavior of the participant to be displayed in the discussion. Specifically, the participant information storage unit 122 stores in advance information indicating the characteristics of the participants. The information indicating the characteristics of the participants is, for example, the result of a psychological test conducted on the participants.

When displaying participant trend information B1, the output unit 116 acquires information indicating characteristics of the participants and patterns selected by the pattern selection unit 115 for multiple discussions in which the participants have participated in the past. Based on the pattern selected by the pattern selection unit 115, the output unit 116 determines the role (that is, behavioral tendency) of the participant in each of the plurality of discussions using the method described above. The output unit 116 selects one role (for example, the role with the highest frequency) from among the determined roles. Then, the output unit 116 generates a sentence corresponding to the participant's characteristics and the selected participant's role based on a predetermined rule. The predetermined rule is a template that is predefined in the storage unit 12 and corresponds to the characteristics of the participants and the roles of the participants.

For example, if the participant's characteristics indicate a high level of uniqueness and the selected participant is "no role", the output unit 116 may output a message that "is good at expressing his or her uniqueness, but is cold to others." Generates the sentence "There are some things that tend to touch each other." The storage unit 12 stores in advance templates corresponding to various participant characteristics and participant roles. The method of generating sentences based on the characteristics of the participants and the roles of the participants shown here is just an example, and the output unit 116 uses a known method that can generate sentences based on the characteristics of the participants and the roles of the participants. A method can be used.

This allows the analyst to compare and recognize the characteristics of the participants collected in advance and the behavioral trends of the participants in the discussion.

Participant progress information B2 is information indicating the amount of speech, amount of interruptions, amount of excitement, and role of the participant to be displayed in a plurality of discussions in which the participant to be displayed has participated in the past. In the example of FIG. 8, participant progress information B2 represents the amount of speech by participants for each discussion as a bar graph, and the average amount of speech of all participants in the discussion is superimposed on the bar graph. . Further, participant progress information B2 represents the number of discussions in which a participant took on a predetermined role (eg, leader or follower). Furthermore, participant progress information B2 represents the number of discussions in which the amount of utterances, the amount of interruptions, and the amount of excitement of participants each satisfied predetermined conditions.

Further, participant progress information B2 represents a phase of the discussion in which the amount of speech satisfies a predetermined condition (for example, a phase in which the average amount of speech of the participants is the highest). Participant progress information B2 also includes information about other participants when the display target participant's utterance increases while the display target participant is participating in the same discussion as a specific other participant. It represents a person. In addition, participant progress information B2 shows that the amount of utterances in a second discussion (for example, the last discussion) that is different from the first discussion has increased compared to the amount of utterances in the first discussion (for example, the first discussion). It shows whether it is done or not.

This allows the analyst to see a list of behavior trends for one participant in discussions he or she has participated in in the past. The participant progress information B2 shown here is just an example, and other information may be expressed based on the amount of speech, amount of interruptions, amount of excitement, and role of the participant in multiple discussions in which the participant has participated in the past. good.

Discussion information B3 is information indicating the amount of speech by a participant to be displayed in a discussion that satisfies a predetermined condition among a plurality of discussions in which the participant to be displayed has participated in the past. In the example of FIG. 8, discussion information B3 includes a graph of the amount of speech for each discussion that satisfies a predetermined condition. In the graph of discussion information B3, the shaded area represents the time-series change in the amount of speech of the displayed participant in one discussion, and the white area represents the total amount of speech of all participants in the discussion. It represents changes over time.

In the example of FIG. 8, the discussions displayed in the discussion information B3 are a plurality of discussions in chronological order (time order). This allows the analyst to view a list of how the trends in the amount of speech by participants have changed over time.

Further, the discussions displayed in the discussion information B3 may be a plurality of discussions that are similar to each other or a plurality of discussions that are not similar to each other. In this case, the output unit 116 calculates the inter-matrix distance between the plurality of transition matrices of the plurality of discussions in which the display target participant participated, and classifies the plurality of discussions whose inter-matrix distance is smaller than a predetermined value to be similar to each other. A plurality of arguments are identified, or a plurality of arguments whose inter-matrix distance is larger than a predetermined value are identified as a plurality of arguments that are dissimilar to each other. This allows the analyst to view a list of trends in the amount of speech by participants in discussions in which the participants are participating, in which the trends in speaker transitions are similar or dissimilar.

FIG. 9 is a front view of the display unit 23 displaying the course report screen C. The course report screen C is a screen that displays information regarding multiple discussions in which multiple participants belonging to the group to be displayed have participated in the past. For example, a group may be multiple participants taking the same course, multiple participants receiving instruction from the same instructor, or the like. The course report screen C includes participant distribution information C1, course progress information C2, course statistical information C3, and ranking information C4.

The participant distribution information C1 is information indicating the distribution of the amount of speech and the amount of interruptions of a plurality of participants belonging to the group to be displayed. In the example of FIG. 9, the distribution information C1 represents a combination of the amount of speech and the amount of interruptions of a plurality of participants belonging to the group to be displayed as a plot, with the horizontal axis representing the amount of speech and the vertical axis representing the amount of interruptions.

This allows the analyst to know the tendencies of multiple participants belonging to the group to be displayed. For example, participants plotted in the upper right area of the distribution information C1 tend to lead the discussion because they have a large amount of speech and a large amount of interruptions. Participants plotted in the upper left area of the distribution information C1 have a large amount of interruptions but a small amount of speech, so they tend to tune in to others in the discussion. Participants plotted in the lower right area of the distribution information C1 have a large amount of speech but a small amount of interruptions, and therefore tend to behave well in discussions. The participants plotted in the lower left area of the distribution information C1 have a small amount of speech and a small amount of interruptions, so they tend to be reluctant to participate in the discussion.

The course progress information C2 is information indicating the progress of the amount of utterances, the amount of interruptions, and the amount of excitement in multiple discussions in which multiple participants belonging to the group to be displayed have participated in the past. In the example of FIG. 9, the course progress information C2 indicates the amount of speech, the amount of interruptions, and the amount of excitement in each of the first, middle, and second periods of multiple discussions in which multiple participants belonging to the group to be displayed have participated in the past. It represents a stacked bar graph. For example, the output unit 116 classifies 1/3 of the multiple discussions from the first discussion to the last into early, middle, and late stages, and calculates the total or average amount of speech, interruption amount, and excitement amount for each classification. and output. This allows the analyst to know changes in the trend of discussion in the group being displayed.

The course statistical information C3 is information indicating statistical values of the amount of speech, the amount of interruptions, the amount of excitement, and their total amount (total activity amount) of a plurality of participants belonging to the group to be displayed. In the example of FIG. 9, the course statistical information C3 is expressed as a bar graph with the horizontal axis representing the amount of speech, the amount of interruptions, the amount of excitement, and the total amount of activity, and the vertical axis representing the number of participants. Furthermore, the course statistical information C3 displays the display format (for example, color) of the bar graph corresponding to the average value of the amount of speech, the amount of interruptions, the amount of excitement, and the total amount of activity in a manner different from the display format of other bar graphs. There is. As a result, the analyst can determine the distribution of the number of people by amount of speech, amount of interruptions, amount of excitement, and total amount of activity, and the amount of speech, amount of interruptions, amount of excitement, and total amount of activity for multiple participants belonging to the group to be displayed. You can know the statistical value of.

The ranking information C4 is information indicating the ranking of the amount of speech of a plurality of participants belonging to the group to be displayed. In the example of FIG. 9, the ranking information C4 ranks a list of multiple participants in the first discussion (for example, the first discussion) according to the amount of utterances of the multiple participants (for example, in ascending order of ranking). ), and a list of multiple participants in a second discussion (for example, the last discussion) that is different from the first discussion is ranked according to the amount of utterances of the multiple participants (for example, in ascending order of rank). )represent.

Furthermore, the ranking information C4 may represent a line connecting a position corresponding to a certain participant in the first discussion and a position corresponding to the participant in the second discussion. Furthermore, the ranking information C4 may represent a change in the ranking of participants in the second discussion compared to the first discussion using a numerical value indicating the amount of change and an arrow indicating the direction of the change (up or down). This allows the analyst to know how the amount of speech by each of the multiple participants changed between the two discussions.

Discussion report screen A, personal report screen B, and course report screen C shown in FIGS. 7 to 9 are examples, and the content, appearance, and arrangement of information may be changed. Furthermore, at least a portion of the discussion report screen A, personal report screen B, and course report screen C shown in FIGS. 7 to 9 may be integrated into one screen, or may be further divided into multiple screens. good.

The output unit 116 outputs information related to the discussion not only by displaying it on the screen but also by printing it on paper using a printer, storing it as data in a storage medium, or transmitting it to the outside via a communication line. You can.

The output unit 116 may output information regarding the discussion by switching the content for each analyst (viewer). In this case, the discussion analysis device 1 receives the output content settings for each analyst in advance, and stores them in the storage unit 12 as setting information in association with the analyst. Setting the output content is performed, for example, by selecting a plug-in that indicates the output content. In the example of FIG. 7, four plug-ins are defined: summary information A1, participant information A2, phase information A3, and comprehensive evaluation information A4. The analyst or the administrator of the discussion analysis device 1 sets output content by selecting a plug-in to be output to the analyst.

In the discussion analysis device 1, when outputting information regarding the discussion, the output unit 116 specifies an analyst to be output and acquires setting information associated with the analyst. Then, the output unit 116 outputs the content indicated by the setting information (plug-in) based on the information stored in the discussion information storage unit 121. Thereby, the discussion analysis device 1 can output different types of information for each analyst.

[Flow of argument analysis method]
FIG. 10 is a diagram showing a flowchart of the discussion analysis method performed by the discussion analysis device 1. In the discussion analysis device 1, the information acquisition unit 111 acquires the amount of time-series utterances of each of the plurality of participants in the discussion (S11). The information acquisition unit 111 causes the discussion information storage unit 121 to store information indicating the amount of speech for each first frame of each of the plurality of participants in the discussion in association with the discussion ID.

At this time, the information acquisition unit 111 performs sound source localization on the discussion audio acquired by the sound collection device 3, and acquires the amount of speech by identifying the speech period of each of the plurality of participants. As another method, the information acquisition unit 111 may acquire the amount of speech by reading and acquiring the speech period stored in the storage unit 12 in advance. Alternatively, the information acquisition unit 111 may acquire the amount of speech by identifying the speech period of the participant based on an image including the face of the participant in the discussion.

The largest speaker identification unit 112 identifies the largest speaker with the largest amount of speech among the plurality of participants for each first frame based on the amount of speech acquired by the information acquisition unit 111 (S12). The transition detection unit 113 detects a transition in speakers that occurs between a plurality of participants, based on the change in the maximum speaker for each first frame specified by the maximum speaker identification unit 112 (S13). The transition detection unit 113 causes the discussion information storage unit 121 to store information indicating the transition matrix generated for each second frame as transition information.

The phase dividing unit 114 divides the discussion into one or more phases based on the similarity in the time series of the transition information indicating the transition detected by the transition detecting unit 113 (S14). The phase dividing unit 114 causes the discussion information storage unit 121 to store information indicating the determined phase of the discussion in association with discussion identification information.

The pattern selection unit 115 generates a plurality of patterns indicating the presence or absence of a transition in each combination of a plurality of participants. The pattern selection unit 115 calculates the degree of similarity between the matrix of each of the plurality of generated patterns and the normalized transition matrix (transition information) for each phase. Then, the pattern selection unit 115 selects, from among the plurality of patterns, a pattern whose calculated degree of similarity satisfies a predetermined condition for each of the one or more phases determined by the phase division unit 114 (S15).

Further, the pattern selection unit 115 may select a sub-pattern that is a modified pattern of the pattern selected for each phase. The pattern selection unit 115 causes the discussion information storage unit 121 to store information indicating the selected pattern and subpattern in association with discussion identification information.

The output unit 116 outputs information regarding the discussion based on the information stored in the discussion information storage unit 121 by the information acquisition unit 111, transition detection unit 113, phase division unit 114, and pattern selection unit 115 (S16). For example, the output unit 116 outputs information regarding the discussion by displaying screens shown in FIGS. 7, 8, and 9 on the display unit 23 of the communication terminal 2.

[Effects of this embodiment]
If the speaker transition is automatically detected simply based on the direction in which the sound is generated, sounds that occur while the participant is speaking but are not utterances will be detected as the participant's utterances, and the speaker transition will be detected automatically. may not be detected correctly. In other words, when the audio of a discussion includes short sounds such as the sound of objects colliding or other participants' chiding, it is difficult to determine whether the short sounds are necessary for analysis or not. . For example, if a participant makes a sound such as tapping a pen on a desk in a situation where the voices of adjacent groups are often mixed in as background noise, the background noise will be mixed into the separated sound. In this case, it is difficult to determine that the pen sound is "unnecessary because it is not a voice." In addition, even if we try to exclude short sounds based on the length of the sound, we cannot exclude long sounds such as "um" and "ho", and conversely, we cannot exclude important meanings such as "no" and "certainly". There is a risk that certain statements may be excluded.

In contrast, the discussion analysis device 1 according to the present embodiment detects a change in speakers based on a change in the participant U who has the largest amount of speech. Therefore, the discussion analysis device 1 can suppress the detection of speaker transitions based on sounds that are not utterances, and can improve the accuracy of detecting speaker transitions in discussions.

The discussion analysis system SS according to this embodiment is suitably used for analyzing active learning conducted by students and for analyzing meetings in organizations. The discussion analysis system SS is also suitably used to analyze group discussions held among candidates during recruitment activities in an organization. Traditionally, these discussions involve a large number of participants, resulting in significant time and expense costs for analyzing the discussions. On the other hand, the discussion analysis device 1 can analyze these discussions automatically and with high accuracy, so the cost for analysis can be significantly reduced.

Although the present invention has been described above using the embodiments, the technical scope of the present invention is not limited to the scope described in the above embodiments, and various modifications and changes can be made within the scope of the gist. be. For example, all or part of the device can be functionally or physically distributed and integrated into arbitrary units. In addition, new embodiments created by arbitrary combinations of multiple embodiments are also included in the embodiments of the present invention. The effects of the new embodiment resulting from the combination have the effects of the original embodiment.

The processor of the discussion analysis device 1 is responsible for each step (process) included in the discussion analysis method shown in FIG. That is, the processor of the discussion analysis device 1 reads a program for executing the discussion analysis method shown in FIG. Implement the argument analysis method shown. Some of the steps included in the argument analysis method shown in FIG. 10 may be omitted, the order of the steps may be changed, or a plurality of steps may be performed in parallel.

SS Discussion analysis system 1 Discussion analysis device 11 Control unit 111 Information acquisition unit 112 Maximum speaker identification unit 113 Transition detection unit 114 Phase division unit 115 Pattern selection unit 116 Output unit

Claims (13)

an information acquisition unit that acquires the amount of speech of each of the plurality of participants in a discussion in which the plurality of participants participate;
In the discussion, a maximum speaker identification unit that identifies a maximum speaker with the maximum amount of speech among the plurality of participants for each first time range;
a transition detection unit that generates transition information indicating a transition of speakers that has occurred among the plurality of participants based on a change in the maximum speaker for each of the first time ranges;
Information regarding the amount of speech by the one participant in the plurality of discussions in which one of the plurality of participants participated, the degree of similarity of the transition information being greater than a predetermined value; an output unit that outputs in association with the one participant;
A discussion analysis device with. an information acquisition unit that acquires the amount of speech of each of the plurality of participants in a discussion in which the plurality of participants participate;
In the discussion, a maximum speaker identification unit that identifies a maximum speaker with the maximum amount of speech among the plurality of participants for each first time range;
a transition detection unit that generates transition information indicating a transition of speakers that has occurred among the plurality of participants based on a change in the maximum speaker for each of the first time ranges;
Information regarding the amount of speech by the one participant in the plurality of discussions in which one of the plurality of participants participated, the degree of similarity of the transition information being smaller than a predetermined value an output unit that outputs in association with the one participant;
A discussion analysis device with. The transition detection unit is configured to detect when the first participant who is the largest speaker in one time range and the second participant who is the largest speaker in a time range following the one time range are different. , generating the transition information indicating the transition from the first participant to the second participant. The discussion analysis device according to any one of claims 1 to 3, further comprising a phase division unit that divides the discussion into one or more phases based on chronological similarity of the transition information. The transition detection unit generates the transition information indicating the number of transitions for each second time range that is longer than the first time range,
The phase division unit clusters the transition information for each of the second time ranges based on the similarity of the time series of the transition information, and clusters the transition information corresponding to the transition information included in the plurality of generated clusters. The discussion analysis device according to claim 4, wherein the one or more phases making up the discussion are determined based on the time during the discussion in a second time range. a pattern selection unit that generates a plurality of patterns indicating the presence or absence of the transition for each combination of the plurality of participants, and selects a pattern whose similarity with the transition information satisfies a predetermined condition from among the plurality of patterns; The discussion analysis device according to any one of claims 1 to 5, further comprising. The pattern selection unit further generates a plurality of sub-patterns in which a part of the selected pattern is changed, and selects a sub-pattern whose degree of similarity with the transition information satisfies a predetermined condition from among the plurality of sub-patterns. The discussion analysis device according to claim 6. The output unit determines the roles of the plurality of participants based on the pattern selected by the pattern selection unit, and outputs each of the plurality of participants in association with the role of each of the plurality of participants. The discussion analysis device according to claim 6 or 7. The discussion analysis device according to any one of claims 6 to 8, wherein the output unit outputs the actions of the plurality of participants as sentences based on the pattern selected by the pattern selection unit. 10. The output unit according to claim 1, wherein the output unit outputs information regarding the amount of speech of the plurality of participants in the plurality of discussions in which the plurality of participants belonging to a predetermined group participated, in association with the group. The argument analysis device described in item (1). The output unit is configured to output rankings of utterances of the plurality of participants belonging to the group in a first discussion and utterances of the plurality of participants belonging to the group in a second discussion different from the first discussion. 11. The discussion analysis device according to claim 10, wherein the discussion analysis device outputs the information in association with the rank of quantity. The processor executes
obtaining the amount of speech of each of the plurality of participants in a discussion in which the plurality of participants participate;
In the discussion, identifying the largest speaker with the largest amount of speech among the plurality of participants for each first time range;
generating transition information indicating a speaker transition that has occurred among the plurality of participants based on the change in the maximum speaker for each of the first time ranges;
Information regarding the amount of speech by the one participant in the plurality of discussions in which one of the plurality of participants participated, the degree of similarity of the transition information being greater than a predetermined value; outputting in association with the one participant;
An argument analysis method with The processor executes
obtaining the amount of speech of each of the plurality of participants in a discussion in which the plurality of participants participate;
In the discussion, identifying the largest speaker with the largest amount of speech among the plurality of participants for each first time range;
generating transition information indicating a speaker transition that has occurred among the plurality of participants based on the change in the maximum speaker for each of the first time ranges;
Information regarding the amount of speech by the one participant in the plurality of discussions in which one of the plurality of participants participated, the degree of similarity of the transition information being smaller than a predetermined value outputting in association with the one participant;
An argument analysis method with

Images