JP5663205B2 - Dialog learning device, dialog analysis device, dialog learning method, dialog analysis method, program - Google Patents

Description

  The present invention relates to a dialogue learning device for creating a hidden Markov model for evaluating a dialogue composed of a plurality of dialogue actions, a dialogue learning method, a dialogue analysis device for evaluating dialogue using the created hidden Markov model, and dialogue The present invention relates to an analysis method and a program for causing a computer to function as these devices.

  As a technique for estimating user satisfaction (evaluation value) in a dialogue, there is a technology that uses an N-gram of a dialogue act described in Non-Patent Document 1. In this method, first, dialogues are classified according to satisfaction, and N-gram probabilities of dialogue actions (labels indicating utterance intentions) for each satisfaction are calculated. Then, for the unknown dialogue, the likelihood of the dialogue action sequence is calculated using the N-gram probability of the dialogue action at each satisfaction level, and with the satisfaction that has the highest likelihood, To do.

  Another similar technique uses a Hidden Markov model (HMM). In Non-Patent Document 2, first, a dialogue performed between humans is modeled using a hidden Markov model. Here, it is assumed that each state of the HMM outputs a dialogue action, and in the process of modeling the HMM, the output probability of the dialogue action in each state using a series of dialogue actions performed between humans as learning data. And optimize the transition probability between states. When a dialogue between a person and a computer is given, the dialogue action sequence of the dialogue is input to the learned HMM, and the satisfaction (naturalness) of the dialogue is estimated with the output probability.

  The above two methods estimate the degree of satisfaction of the entire dialogue, but Non-Patent Document 3 proposes a method for estimating the degree of satisfaction for each user utterance in the dialogue. In Non-Patent Document 3, a satisfaction level label is assigned in advance to each utterance of dialogue, and HMM is learned based on this data. Specifically, each state of the HMM corresponds to a satisfaction level label, and each state outputs various feature values related to speech. In the learning process, the transition probability between states is obtained from the transition probability of the satisfaction degree label given in advance. Further, the output probability of the feature value in each state is obtained from the appearance probability of the utterance feature value for each satisfaction level label. Then, when an unknown dialogue is given, by obtaining the maximum likelihood state transition, it is possible to estimate from which state each utterance is output, and it is possible to estimate the satisfaction of each utterance.

Nao Hara, Norihide Kitaoka, Kazuya Takeda, “User Satisfaction Estimation Method Using Dialog History N-gram of Spoken Dialogue System”, 79th Spoken Language Information Processing Study Group (SIG-SLP), 2009. Naoki Kashimura, Fujio Toriumi, Kenichiro Ishii, “Evaluation of non-task-oriented dialogue system using HMM”, IEICE Transactions D Vol.J92-D No.4 pp.542-551, 2009. Klaus-Peter Engelbrech, Florian Godde, Felix Hartard, Hamed Ketabdar, and Sebastian Mo ~ ller, “Modeling user satisfaction with hidden Markov models”, In Proc. SIGDial, pp.170-177, 2009. (However, “o ~” Means oumlaut)

  Rather than estimating the satisfaction of the entire dialogue, it is useful to be able to identify the point where the dialogue is not progressing smoothly if the satisfaction for each utterance (or each dialogue act) in the dialogue can be estimated. Many. However, with the methods of Non-Patent Document 1 and Non-Patent Document 2, it is only possible to obtain the satisfaction level of dialogue. In addition, the method for obtaining the satisfaction level for each utterance as in Non-Patent Document 3 requires that the satisfaction level be labeled for each utterance in the dialogue, and the cost for estimation is high.

  It is an object of the present invention to provide a dialog learning device and a dialog analysis device that do not require data that is previously labeled for the degree of satisfaction for each dialogue action and that can evaluate the degree of satisfaction for each dialogue action in the dialogue.

  First, K and N are integers of 2 or more, k is an integer of 1 to K, and n is an integer of 1 to N. In addition, “dialogue” is a collection of “speech” of multiple speakers, “speech” is a collection of one or more continuous “dialogue acts” of only one speaker, and “dialogue act” is an individual for a certain purpose. It is an oral act.

The first dialog learning device of the present invention creates a hidden Markov model for evaluating a dialog composed of a plurality of dialog actions. The dialogue learning apparatus of the present invention includes N dialogue time-series data (d ₁ ,..., D _N ) and evaluation values (r (d ₁ ),. by using the _N)), for each stage of the evaluation, hidden Markov model M ₁ with a state of outputting the dialogue act of the _speaker, ..., to learn the M _K, hidden Markov model M _1, ..., all of M _K connect a state to ergodic comprises a first model learning unit for creating a hidden Markov model M _C.

The second dialog learning device of the present invention includes a first model learning unit, a second model learning unit, and a model connection unit. The first model learning unit includes time series data (d ₁ ,..., D _N ) of _N conversations and evaluation values (r (d ₁ ) _,. )) using, for each stage of the evaluation, hidden Markov model M ₁ with a state of outputting the dialogue act of the _speaker, ..., to learn the M _K, hidden Markov model M _1, ..., all of M _K Connect the states to _ergodic and create a hidden Markov model M _E0 . The second model learning unit learns one hidden Markov model M ₀ having a state in which the dialogue action of the speaker is output from all dialogues. Model coupling portion all the states of the Markov model M _E0 and hidden Markov model M ₀ connected to ergodic, creating a hidden Markov model M _C.

The third conversation learning device of the present invention includes a first model learning unit, a second model learning unit, a first model connection unit, a re-learning unit, a second model connection unit, and an iterative determination unit. The first model learning unit evaluates each dialogue with time series data (d ₁ ,..., D _N ) of _N dialogues at the K stage (r (d ₁ ),..., R (d _N ). ), Hidden Markov Models M ₁ ,..., M _K having a state of outputting a speaker's dialogue action are learned for each evaluation stage. The second model learning unit learns one hidden Markov model M ₀ having a state in which the dialogue action of the speaker is output from all dialogues. The first model connection unit creates a hidden Markov model M _{1 + 0} ,..., M _{K + 0} by pairing each of the hidden Markov models M _k with a copy of the Markov model M ₀ and connecting all the states in the pair. . The re-learning unit uses the hidden Markov model M _{1 + 0} ,..., M _{K + 0} as an initial model, time series data (d ₁ ,..., D _N ) and evaluation values (r (d ₁ ),..., R (d _N )). Re-learn the hidden Markov model M _{1 + 0} ,..., M _{K + 0} given as learning data. Second model coupling portion hidden relearned Markov model _M 1 + 0, _..., concatenates _{M K + 0,} and one hidden Markov model _{M C.} Repetition determining unit Hidden Markov Model M _{1 + 0,} ..., the M K + _0, determines whether the process is repeated relearning portion and the second model coupling part by whether there is improvement of more than predetermined likelihood against training data .

Interactive analysis apparatus of the present invention, by using a Hidden Markov model M _C, or output from the state of the evaluation values of each dialogue act hidden Markov model M _C throat of the time-series data d _x dialogue evaluation value is not known An evaluation value estimating unit that obtains an evaluation value of each dialogue action by estimation and estimates the evaluation value of the dialogue from the evaluation value of each dialogue action is provided.

  According to the dialogue learning device and the dialogue analysis device of the present invention, when forming a hidden Markov model for evaluating the degree of satisfaction for each dialogue action in the dialogue, it is necessary to preliminarily label the degree of satisfaction for each dialogue act. Learning costs can be reduced. As a result, it is easier to evaluate the degree of satisfaction for each dialogue action in the dialogue, so it is easier to identify the places where the conversation during the dialogue is not smooth. It leads to the realization of a dialogue system that is easy for humans to use.

FIG. 3 is a diagram illustrating an example of a functional configuration of the dialog analysis apparatus according to the first embodiment. The figure which shows the processing flow of the dialog analyzer of Example 1. FIG. The figure which shows the image of the hidden Markov model _Mk produced when there are two speakers. The figure which shows the function structural example of the dialog analyzer of Example 2, and Example 2 modification. The figure which shows the processing flow of the dialog analyzer of Example 2 and Example 2 modification. The figure which shows the topology of the ergodic 1 in the case of two speakers by two-step evaluation. The figure which shows the function structural example of the dialog analyzer of Example 3 and Example 3 modification. The figure which shows the processing flow of the dialogue analyzer of Example 3 and Example 3 modification. The figure which shows the image of the hidden Markov model produced at each step of Example 3. FIG. The figure which shows the number of dialogs of AD data, the number of dialog acts, the average number of dialog acts per dialog, and a standard deviation. The figure which shows the number of dialogues of AL data, the number of dialogue acts, the average number of dialogue acts per dialogue, and standard deviation. The figure which shows the example of the evaluation value of the satisfaction for every dialog of AD data. The figure which shows the example of the evaluation value of the satisfaction for every utterance in the dialogue of AD data. The figure which shows the dialogue act provided to U1-U10 of FIG. The figure which shows the evaluation result with respect to AD data. The figure which shows the evaluation result with respect to AL data. The figure which shows a part of HMM of Ergodic 0 obtained from AD data. The figure which shows a part of connection 1 HMM obtained from AL data. The figure which shows the example of the estimated evaluation value series of each model with respect to AD data. The figure which shows the example of the estimated evaluation value series of each model with respect to AL data.

  In this specification, “dialogue” is a collection of “utterances” of a plurality of speakers, “utterance” is a collection of one or more continuous “dialogue actions” of only a certain speaker, and “dialogue action” is for some purpose. Individual verbal acts. For example, the speaker A, "Hello. I like a lion.", Described the dialogue that the speaker B, "I am also." As an example. The whole is “dialogue”, and this dialogue is composed of “utterance” of speaker A and “utterance” of speaker B (two “utterances”). In addition, "Hello." The dialogue act of interaction management (greeting), "I like a lion." The dialogue act of self-disclosure, "it is also me." Is a dialogue act of empathy. That is, the “utterance” of the speaker A is composed of two dialogue actions, and the “utterance” of the speaker B is composed of one dialogue action. The “speaker” may be a natural person or an apparatus such as an utterance system.

  Hereinafter, embodiments of the present invention will be described in detail. In addition, the same number is attached | subjected to the structure part which has the same function, and duplication description is abbreviate | omitted. K and N are integers of 2 or more, k is an integer of 1 to K, and n is an integer of 1 to N.

FIG. 1 shows a functional configuration example of the dialog analysis apparatus according to the first embodiment, and FIG. 2 shows a processing flow of the dialog analysis apparatus according to the first embodiment. Interactive analyzer 50, hidden Markov model (Hidden Markov model, HMM) for evaluating a dialogue composed of a plurality of dialog act to create a M _C, to evaluate the interaction comprises a plurality of dialog acts. The dialog analysis device 50 includes an input / output unit 110, a first model learning unit 120, an evaluation value estimation unit 180, and a recording unit 90.

The input / output unit 110 uses time series data (d ₁ ,..., D _N ) of _N conversations and evaluation values (r (d ₁ ),..., R (d _N )) obtained by evaluating each conversation at the K stage. And received as learning data (S111). The first model learning unit 120 uses the learning data received by the input / output unit 110 to learn hidden Markov models M ₁ ,..., M _K having a state in which a speaker's dialogue action is output at each evaluation stage. Then, all the states of the hidden Markov model M ₁ ,..., M _K are connected to the _ergodic to create a hidden Markov model M _E0 and record it in the recording unit 90 (S120). In this embodiment, a hidden Markov model M _E0 recorded in the recording unit 90, a hidden Markov model M _C used when the evaluation. Further, hereinafter, the hidden Markov model M _E0 is referred to as “ergodic 0”.

The processing in step S120 will be described in detail. Since the evaluation values are classified into K stages, a set of time series data D _k = {∀d _m | r (d _m ) = k} sorted for each evaluation value. create. Then, the hidden Markov model _Mk is learned using the set of time series data. FIG. 3 shows an image of a hidden Markov model _Mk created when there are two speakers. The HMM in FIG. 3 has two states, and each state outputs the dialogue action of speaker 1 and speaker 2, respectively. Each state can transition to any state. Such an HMM is called a Speaker HMM, and is used when a dialogue action sequence is modeled by an HMM in consideration of a speaker's turn switching. For example, Non-Patent Reference 1 (Toyomi Meguro, Ryuichiro Higashinaka, Kohji Dohsaka, Yasuhiro Minami, and Hideki Isozaki, “Analysis of listening-oriented dialogue for building listening agents”, In Proc. SIGDial, pages 124-127, 2009.) Then, the Speaker HMM is used to model the listener dialogue. As shown in FIG. 3, if the dialogue action of one speaker is output from one state, it is easy to grasp the dialogue action for each speaker. However, it is also possible to model with an HMM that outputs the dialogue action of two speakers from one state. In this embodiment, the Speaker HMM has two states, but it is easy to change the number of states.

Hidden Markov models M _k as shown in FIG. 3 are created by the number of evaluation values. That is, K pieces are created. All states are connected to all other states (connected to ergodic). At this time, the transition probabilities between the states are all made equal. An expected value maximization algorithm (EM algorithm) may be used as the learning algorithm. Note that any HMM described below may use an EM algorithm for learning.

The input / output unit 110 receives the time series data d _x of the dialogue whose evaluation value is not known (S112). Evaluation value estimation unit 180, a hidden Markov model using the M _C, estimate the output from state of the evaluation value of each dialogue act hidden Markov model M _C throat of the time-series data d _x dialogue evaluation value is not known Thus, the evaluation value of each dialogue action is obtained, and the evaluation value of the dialogue is estimated from the evaluation value of each dialogue action (S180). For example, by the Viterbi decoding (Viterbi decoding), it may be estimated or output from the state of the evaluation values of the Markov model M _C throat each dialog acts hidden.

  In the above description, both the learning function and the analysis function are provided in one apparatus (dialog analysis apparatus). However, the learning function and the analysis function may be separated. In the case of separation, the dialogue learning device 51 may be formed only by the dotted line portion in FIG. And the processing flow of the dialog learning apparatus 51 should just be made like the part of step S51 shown with the dotted line of FIG.

  According to the dialogue learning device and the dialogue analysis device of the first embodiment, when forming a hidden Markov model for evaluating the degree of satisfaction for each dialogue action in the dialogue, data in which the degree of satisfaction for each dialogue action is labeled in advance is required. Learning costs can be reduced. As a result, it becomes easier to evaluate the degree of satisfaction for each dialogue act in the dialogue.

In Example 1, the example estimated by ergodic 0 was demonstrated. However, for example, dialogue acts such as greetings and responses to questions appear in common in all dialogues regardless of satisfaction. Ergodic 0 assigns one of the evaluation values for such a dialogue act. Which evaluation value is assigned depends on the output distribution of dialogue actions in the dialogue for each evaluation value. For this reason, it may happen that a high evaluation value is assigned to all “greetings” because the dialogue action corresponding to “greeting” happens to be a high evaluation value dialogue. If possible, it may be desirable not to give a score or to give an evaluation value of 0 for interactive actions that appear in common in all evaluation values. Therefore, in this embodiment, one hidden Markov model M ₀ is learned from all conversations. Then, a hidden Markov model M _E1 (hereinafter referred to as “ergodic 1”) in which the hidden Markov model M ₀ and ergodic 0 are connected is created.

  FIG. 4 shows a functional configuration example of the dialog analysis apparatus of the second embodiment, and FIG. 5 shows a processing flow of the dialog analysis apparatus of the second embodiment. The dialog analysis apparatus 100 includes an input / output unit 110, a first model learning unit 120, a second model learning unit 130, a model connection unit 140, an evaluation value estimation unit 180, and a recording unit 190.

The input / output unit 110 uses time series data (d ₁ ,..., D _N ) of _N conversations and evaluation values (r (d ₁ ),..., R (d _N )) obtained by evaluating each conversation at the K stage. And received as learning data (S111). The first model learning unit 120 uses the learning data received by the input / output unit 110 to learn hidden Markov models M ₁ ,..., M _K having a state in which a speaker's dialogue action is output at each evaluation stage. Then, all the states of the hidden Markov model M ₁ ,..., M _K are connected to the _ergodic to create a hidden Markov model M _E0 and record it in the recording unit 190 (S120). The processing so far is the same as in the first embodiment.

The second model learning unit 130 learns one hidden Markov model M ₀ having a state in which the dialogue action of the speaker is output from all dialogues, and records it in the recording unit (S130). The model connecting unit 140 connects all the states of the Markov model M ₀ and the hidden Markov model M _E0 to ergodic, creates a hidden Markov model M _E1 , and records it in the recording unit 190. In this embodiment, a hidden Markov model M _E1 recorded in the recording unit 190, a hidden Markov model M _C used when the evaluation.

The input / output unit 110 receives the time series data d _x of the dialogue whose evaluation value is not known (S112). Evaluation value estimation unit 180, a hidden Markov model using the M _C, estimate the output from state of the evaluation value of each dialogue act hidden Markov model M _C throat of the time-series data d _x dialogue evaluation value is not known Thus, the evaluation value of each dialogue action is obtained, and the evaluation value of the dialogue is estimated from the evaluation value of each dialogue action (S180). For example, by the Viterbi decoding, it may be estimated or output from the state of the evaluation values of the Markov model M _C throat each dialog acts hidden.

In the case of two speakers in a two-stage evaluation, the topology of ergodic 1 is as shown in FIG. In FIG. 6, for the sake of simplicity, only HMMs with evaluation values 1 and 2 and M ₀ (corresponding to HMMs with an evaluation value of 0) are shown. For example, when a certain dialogue sequence is decoded using this HMM, if 1 → 2 → 5 → 6 → 4 → 5 is the maximum likelihood state sequence, the corresponding evaluation value is 1 → 1 → 0 → 0. → 2 → 0. Therefore, this is an estimated evaluation value series by ergodic 1. Since it can be considered that 0 does not affect the previous evaluation value, this sequence can be read as 1 → 1 → 1 → 1 → 2 → 2.

  Also in this embodiment, the learning function and the analysis function may be separated. In the case of separation, the dialogue learning device 101 may be formed only by the dotted line portion in FIG. Then, the processing flow of the conversation learning apparatus 101 may be as shown in step S101 indicated by the dotted line in FIG.

  According to the dialogue learning device and the dialogue analysis device of the second embodiment, when forming a hidden Markov model for evaluating the degree of satisfaction for each dialogue action in the dialogue, data that labels the degree of satisfaction for each dialogue action is required in advance. Learning costs can be reduced. As a result, it becomes easier to evaluate the degree of satisfaction for each dialogue act in the dialogue.

[Modification]
In the case of the hidden Markov model M ₀ of the second embodiment, the distribution becomes broad (smooth, uniform), and the output probabilities of all dialogue actions in the state of M ₀ are averaged and may be uniformly reduced. There is. Therefore, when the time series data d _x of the unknown dialogue is given, there is a possibility that the estimated state series does not pass through the hidden Markov model M ₀ at all. In order to solve such a problem, the number of states in M ₀ is increased in this modification. By doing so, M ₀ can maintain a state having a sharp (not broad, pointed) distribution. As a result, the state sequence of the maximum likelihood obtained by the Viterbi decoding there is a possibility that through the M _0. The hidden Markov model with the increased number of states is called “Ergodic 2”. Here, it is particularly assumed that the number of states is double, that is, there are two states for each speaker (four states if there are two speakers). In this case, evaluation value estimation unit uses the ergodic 2, to estimate the output from HMM state of which evaluation value is the dialogue act when the time-series data d _x of the unknown interaction is input.

FIG. 4 shows an example of a functional configuration of a dialog analysis apparatus according to a modification of the second embodiment, and FIG. 5 shows a processing flow of the dialog analysis apparatus according to the second embodiment. The dialogue analysis apparatus 200 includes an input / output unit 110, a first model learning unit 120, a second model learning unit 230, a model connection unit 240, an evaluation value estimation unit 180, and a recording unit 190. A difference from the second embodiment is a second model learning unit 230 and a model connection unit 240. The second model learning unit 230 learns the hidden Markov model _{M 0} with by 2 state with respect to each speaker, and records in the recording unit 190 (S230). Then, the model connecting unit 240 connects all the states of the hidden Markov model M ₀ and the hidden Markov model M _E0 to ergodic, creates the hidden Markov model M _E2 , and records it in the recording unit 190 (S240). .

  Also in this modification, the learning function and the analysis function may be separated. In the case of separation, the dialogue learning device 201 may be formed only by the dotted line portion in FIG. Then, the processing flow of the dialogue learning apparatus 201 may be as shown in step S201 indicated by the dotted line in FIG.

According to the dialogue learning device and the dialogue analysis device of this modification, the same effects as those of the second embodiment can be obtained. Furthermore, even when the distribution of ergodic 1 is broad (slow, uniform), the distribution of ergodic 2 can be sharp (not broad, pointed), so that the maximum likelihood state sequence obtained by viterbi decoding but there is a possibility that passes through the M _0. Therefore, the dialogue can be evaluated more accurately.

  The present embodiment is a technique for solving the above-described problem in the ergodic 1 using connected learning. Concatenated learning is a technique used when voices are modeled without supervision by HMM. With this technique, the output distribution of dialogue actions that appear in common in each evaluation value and the dialogue as a whole is specified. Can focus on.

  FIG. 7 shows a functional configuration example of the dialog analysis apparatus according to the third embodiment, FIG. 8 shows a processing flow of the dialog analysis apparatus according to the third embodiment, and FIG. 9 shows an image of a hidden Markov model created at each step of the third embodiment. The dialog analysis apparatus 300 includes an input / output unit 110, a first model learning unit 320, a second model learning unit 130, a first model connection unit 340, a re-learning unit 350, a second model connection unit 360, an iterative determination unit 370, and an evaluation. A value estimation unit 180 and a recording unit 390 are provided.

The input / output unit 110 uses time series data (d ₁ ,..., D _N ) of _N conversations and evaluation values (r (d ₁ ),..., R (d _N )) obtained by evaluating each conversation at the K stage. And received as learning data (S111). The first model learning unit 320 uses the learning data received by the input / output unit 110 to learn hidden Markov models M ₁ ,..., M _K having a state in which a speaker's dialogue action is output at each evaluation stage. Then, it is recorded in the recording unit 390 (S320).

The second model learning unit 130 learns one hidden Markov model M ₀ having a state in which the dialogue action of the speaker is output from all dialogues, and records it in the recording unit 390 (S 130). The first model coupling unit 340 creates a hidden Markov model M _{1 + 0} ,..., M _{K + 0} by pairing each of the hidden Markov models M _k and a copy of the Markov model M ₀ and connecting all the states in the pair. (S340).

The re-learning unit 350 uses the hidden Markov model M _{1 + 0} ,..., M _{K + 0} as an initial model, time series data (d ₁ ,..., D _N ) and evaluation values (r (d ₁ ),..., R (d _N )). the given as learning data the hidden Markov model _M 1 + _0, ..., and relearn _{M K + 0,} hidden relearned Markov model _M 1 + _0, ..., a _{M K + 0} is recorded in the recording unit 390 (S350). Second model connection unit 360, hidden relearned Markov model _M 1 + 0, _..., concatenates _{M K + 0,} and one hidden Markov model _{M C} (S360). When concatenating, the probability tables of all HMMs are added, but when the copied M ₀ is combined into a new M ₀ , this output probability is divided by K and added to 1. adjust. This process is the process of “AVG” in step S360 of FIG. At this time, each state of the Hidden Markov Model M ₀ is the Hidden Markov Model M _{1 + 0,} ..., are connected to all the states of the M K + _0. On the other hand, each state of the hidden Markov model M _{k + 0} is connected only to the state of the hidden Markov model M ₀ .

Repetition determining unit 370, a hidden Markov model _M 1 + 0, _..., the _{M K + 0,} processing and re-learning unit 350 of the second model connecting portion 360 by whether there is improvement of more than predetermined likelihood respect learning data (step S350 , S360) is determined (S370). For example, if there is a likelihood improvement in any of all the hidden Markov models M _{k + 0} , the iterative process is continued (the determination in step S370 is Yes). The Hidden Markov Model M _{1 + 0, ...,} M terminates the iterative process when there is no improvement in the likelihood for K + _0, that time and learning result Hidden Markov Models M _C recorded in the recording unit 390 (step S370 Is determined as No). Final Hidden Markov Models M _C obtained in this example, hereafter referred to as "connection 1". Incidentally, if the determination in step S370 is Yes, the the processing returns to step S350, at this time, the hidden Markov model _{M C} again, Hidden Markov Models _M 1 + 0, _..., is divided into _{M K + 0.} The transition probability of _{M 0} is evenly distributed between _{M 0} and _{M k.}

The input / output unit 110 receives the time series data d _x of the dialogue whose evaluation value is not known (S112). Evaluation value estimation unit 180, a hidden Markov model using the M _C, estimate the output from state of the evaluation value of each dialogue act hidden Markov model M _C throat of the time-series data d _x dialogue evaluation value is not known Thus, the evaluation value of each dialogue action is obtained, and the evaluation value of the dialogue is estimated from the evaluation value of each dialogue action (S180). For example, by the Viterbi decoding, it may be estimated or output from the state of the evaluation values of the Markov model M _C throat each dialog acts hidden.

  Also in this embodiment, the learning function and the analysis function may be separated. In the case of separation, the dialogue learning device 301 may be formed only by the dotted line portion in FIG. And the processing flow of the dialog learning apparatus 301 should just be made like the part of step S301 shown with the dotted line of FIG.

  According to the dialogue learning device and the dialogue analysis device of the third embodiment, when forming a hidden Markov model for evaluating the degree of satisfaction for each dialogue action in the dialogue, the data for labeling the degree of satisfaction for each dialogue action in advance is required. Learning costs can be reduced. As a result, it becomes easier to evaluate the degree of satisfaction for each dialogue act in the dialogue.

[Modification]
As with Example 2, in the case of Example 3, a hidden Markov model M ₀ may have one by 2 state with respect to each speaker. FIG. 7 shows a functional configuration example of the dialog analysis apparatus of the third embodiment, and FIG. 8 shows a processing flow of the dialog analysis apparatus of the third embodiment. The dialogue analysis apparatus 400 includes an input / output unit 110, a first model learning unit 320, a second model learning unit 230, a first model connection unit 440, a re-learning unit 350, a second model connection unit 360, an iterative determination unit 370, and an evaluation. A value estimation unit 180 and a recording unit 390 are provided.

A difference from the third embodiment is a second model learning unit 230 and a first model connection unit 440. The second model learning unit 230 learns the hidden Markov model _{M 0} with by 2 state with respect to each speaker, and records in the recording unit 190 (S230). The first model connection unit 440 creates a hidden Markov model M _{1 + 0} ,..., M _{K + 0} by pairing each of the hidden Markov models M _k and a copy of the Markov model M ₀ and connecting all the states in the pair. (S440).

  Also in this modification, the learning function and the analysis function may be separated. In the case of separation, the dialogue learning device 401 may be formed only by the dotted line portion in FIG. Then, the processing flow of the dialogue learning apparatus 401 may be as shown in the part of step S401 indicated by the dotted line in FIG.

According to the dialogue learning device and the dialogue analysis device of this modification, the same effects as those of the third embodiment can be obtained. Note that the final hidden Markov model M _C obtained in the modification, hereinafter referred to as "connection 2".

[Experiment]
Ergodic 0, Ergodic 1, Ergodic 2, Concatenation 1 and Concatenation 2 models are divided into AD data (Animal Discussion) in which people and systems are talking about animal likes and dislikes, and humans are divided into listeners and speakers The model was applied to AL data (Attentive Listening), and the effectiveness of the model was evaluated. FIG. 10 and FIG. 11 show the number of dialogs of AD data and AL data, the number of dialog actions, the average number of dialog actions per dialog, and the standard deviation.

<AD data>
AD data consists of 1000 dialogues in total. Details of this data can be found in Non-Patent Reference 2 (Ryuichiro Higashinaka, Kohji Dohsaka, and Hideki Isozaki, “Effects of self-disclosure and empathy in human-computer dialogue”, In Proc. SLT, pages 109-112, 2008.) It is stated.

  Of the 1000 dialogues, 180 dialogues are given 7 levels of satisfaction (1 is the worst and 7 is the best) from three viewpoints. Of these 180 dialogs, 90 conversations are assigned satisfaction levels for each utterance in seven stages from the same three viewpoints. Here, the three viewpoints are whether the dialogue is smooth, whether the user feels familiar with the system, and whether the user can easily continue the dialogue. This labeling was done by one evaluator. Specifically, the satisfaction evaluation value for each dialogue is given as shown in FIG. Further, the evaluation value of the satisfaction level for each utterance in the dialogue is given as shown in FIG. In this data, the evaluation value for each utterance is given only immediately after the system utterance. This is because when the evaluator gives the evaluation value, the evaluation value is obtained by the question “how satisfied is the other person's speech if you are a user”.

  In addition to this evaluation value data, each utterance in this data is given an annotation of dialogue action. The dialogue actions are roughly divided into seven types as follows, and there are 29 types in total.

Dialogue on self-disclosure In this group, DISC-OTHER (disclosure of content other than the domain that the system can handle), DISC-P (disclosure of proposition), DISC-PR (disclosure of reason for proposition), DISC-R ( Disclosure of reasons), DISC-R-OTHER (disclosure of other reasons), DISC-NO-MORE-R (disclosure of no more reasons), and RES (Yes-No response) are defined as dialogue actions did. For example, DISC-P is a dialogue act representing self-disclosure of the proposition P regarding the likes and dislikes of animals. Here, the proposition is either “the speaker likes the animal X” or “the speaker does not like the animal X”. For example, “I like cats” is DISC-P. Similarly, “I hate giraffes” is also DISC-P.

Dialogue on empathy In this group, DISC-AGREE-OTHER (consent for content outside the domain that the system can handle), DISC-AGREE-P (consent to proposition), DISC-AGREE-PR (reason for consent to proposition) Disclosure), DISC-AGREE-R (disclosure of reasons for consent), EMP (consent by clear expression), and REPEAT (consent by repetition) were defined as dialogue actions. For example, DISC-AGREE-P corresponds to an utterance indicating consent for P disclosed by the conversation partner. For example, “I like it” or “I like it” are uttered in response to the utterance “I like cats”.

Dialogue on non-sympathy In this group, DISC-DISAGREE-OTHER (disagreement about contents other than the system can handle), DISC-DISAGREE-P (disagreement with proposition), DISC-DISAGREE-R (disagreement) Disclosure of reason) was defined as a dialogue act. For example, DISC-DISAGREE-P corresponds to an utterance indicating non-consent for P self-disclosure by the conversation partner. For example, “I do not like cats” or “I am different”, which is issued in response to the utterance “I like cats”.

Dialogue activities related to dialogue management In this group, GOODBYE (parting expression), GREETING (greeting), OPEN-DIALOGUE (topic start expression), CLOSE-DIALOGUE (topic end expression), Q-OPEN-DIALOGUE (topic) SHIFT-TOPIC (topic change) was defined as a dialogue act. For example, GREETING represents a greeting at the beginning of the dialogue. For example, a "Hello" or "very much".

Dialogue related to questions In this group, Q-DISC-OTHER (questions about non-domain contents that the system can handle), Q-DISC-P (questions about propositions), Q-DISC-P-OPEN (propositions about propositions) Open questions regardless of Yes / No), Q-DISC-R (questions about reasons), and Q-DISC-R-OTHER (questions about other reasons) were defined as dialogue actions. For example, Q-DISC-P represents an utterance that asks a conversation partner a Yes-No question about the proposition P. For example, “Do you like cats?”. Q-DISC-P-OPEN represents an utterance that asks the conversation partner an open question about the proposition P. For example, it is a question whose answer is not Yes-No, such as "What do you think of cats?"

Dialogue related to mutualism In this group, ACK was defined as interactive behavior. This corresponds to conflicts such as “Yes” and “Yes”.

Other Dialogue Actions In this group, OTHER (other) was defined as a dialog action. This is an interactive action given to utterances not included above.

  For example, a dialogue act as shown in FIG. 14 is given to U1-U10 in FIG. From the data shown above, it is possible to obtain a dialogue action and an evaluation value associated with it in each dialogue. For example, in the utterance sequence (U1-U10) shown in FIG. 13 and FIG. OPEN, DISC-P, ACK, Q-DISC-R}, the smoothness corresponding to each is {4, 4, 4, 4, 4, 4, 4, 4, 4, 4}, and the familiarity is { 4, 4, 4, 4, 4, 5, 5, 5, 5, 5}, the ease of continuing the dialogue is {4, 4, 4, 5, 5, 5, 5, 5, 5, 6} It becomes the evaluation value. Note that the evaluation value of the dialogue action corresponding to the user utterance here takes over the evaluation value for the dialogue action of the immediately preceding system. One utterance can be given one or more dialogue acts. For example, a user is "Hello. I like a lion." If you and uttered, the utterance corresponding to the {GREETING, DISC-P} dialogue act column called. In this case, it is assumed that both dialogue actions have an evaluation value given to this utterance.

  AD-SUB1, which is a subset of AD data, is data with an evaluation value for each dialog. AD-SUB2 is data in which an evaluation value for each utterance is given in addition to an evaluation value for each dialogue. It should be noted that all utterances of AD data are given a dialogue action label.

<AL data>
AL data consists of 1260 dialogues in total. This data is a hearing dialogue (a dialogue in which one hears and the other speaks and the listener listens to the dialogue) (refer to Non-Patent Reference 1 for the hearing dialogue).

  As with the AD data, the satisfaction level for each dialogue is assigned to each dialogue in seven stages from three viewpoints. Also, 100 dialogs (AL-SUB1) are assigned labels for each utterance in seven stages from the same three viewpoints. Here, the three viewpoints are whether the dialogue is smooth, whether the speaker feels close to the listener and whether it is a good listener. The procedure for assigning evaluation values is the same as that for AD data. There are 40 kinds of dialogues such as greetings, self-disclosure, questions, etc., which are assigned to all utterances. Specific actions include GREETING, DISC-FACT (self-disclosure of facts), INFO (communication of objective information), Q-FACT (question about facts), DISC-EVAL-NEG (negative) Self-disclosure of evaluation), DISC-EVAL-POS (self-disclosure of positive evaluation), CONFIRM (confirmation), SYNPATHY (expression of empathy), etc.

<Experimental procedure>
First, an AD conversation HMM (ergodic 0, ergodic 1, ergodic 2, concatenation 1, and concatenation 2 model) is learned from AD-SUB 1 to which only a label for each conversation is assigned. Using these models, the evaluation value is estimated for each dialogue action in AD-SUB1, and the effectiveness of the proposed model is verified by evaluating whether this evaluation value matches the evaluation value for each utterance obtained in AD-SUB2. To do. Also for the AL dialogue, a model based on the proposed method is learned from all AL data and compared with the evaluation value for each utterance possessed by the AL-SUB1.

For ergodic 1, ergodic 2, connection 1, and connection 2, M ₀ represents an evaluation value of 0 and represents the same value as the previous evaluation value. If the maximum likelihood state at the start of dialogue is M ₀ , the evaluation value is 4 which is an average of 7 levels. For example, if the estimated evaluation value series is 0 → 1 → 2 → 7 → 0 → 5 → 0 → 6 → 0, this series is 4 → 1 → 2 → 7 → 7 → 5 → 5 → 6 → It is considered 6 and compared with the correct answer series.

  In this evaluation experiment, a 10-fold cross-validation is performed. That is, if it is AD data, AD-SUB1 is divided into 10 parts, learning HMM from 9 fold data among them, and evaluating using the evaluation value for each utterance included in the remaining one fold, Repeat 10 times. The same procedure is used to evaluate AL data. In this evaluation experiment, ergodic 0 is used as the baseline. This is because no studies have been found in the past to estimate the evaluation value for each dialogue action (or utterance) based only on the evaluation value for each dialogue, and Ergodic 0 has the simplest structure in the present invention. .

  Moreover, the method of nonpatent literature 3 is used as an upper limit (best result) of this method. This method uses the evaluation value for each dialogue action as correct answer data, and learns the HMM to associate the dialogue action with the evaluation value. That is, as described in “Problems to be Solved by the Invention”, it is necessary to label each utterance in the dialogue, and there is a problem that the cost for estimation is high. Since the present invention is intended to maintain the estimation accuracy while eliminating the need for labeling the degree of satisfaction for each dialogue action, the estimation accuracy of the method of Non-Patent Document 3 is set as the upper limit.

  The methods of ergodic 0, ergodic 1, ergodic 2, concatenation 1, concatenation 2 and non-patent document 3 output an estimated evaluation value sequence for the dialogue action sequence, but the comparison with the correct evaluation value sequence is an evaluator. Will only conduct dialogue actions that are actually rated. As described above, the evaluation value by the evaluator is given only for the interactive action of the system in the AD data, and the evaluation value by the evaluator only for the interactive action of the hearing person in the AL data. This is because the evaluation values for the dialogue action and the dialogue action of the speaker are approximate and may not be accurate. Actually, these evaluation values are given as padding because of the need for modeling in Non-Patent Document 3, and in order to make an accurate evaluation, the evaluator actually gave the evaluation value. It would be reasonable to do only for actions.

<Evaluation scale>
In this evaluation experiment, the estimated evaluation value sequence for the dialogue action sequence is evaluated using four evaluation scales. Here, R is a correct evaluation value series in the meaning of reference (correct answer), and H is an evaluation value series estimated in the sense of hypothesis. L is the length of the dialogue action sequence. Note that MR is evaluated by simply matching the evaluation values. RMSE evaluates the coincidence of evaluation values as a distance. DIST and KL consider the series as a distribution of evaluation values and evaluate the match.

Match Rate (MR)

Here, match is a function that returns 1 when the elements match and returns 0 when the elements do not match.

Root Mean Squared Error (RMSE)

Distribution Distance (DIST)

Here, maxscore is a maximum score value (7 in this evaluation experiment), and count is a function that returns the number of evaluation values i in the evaluation value series. For example, when R is <4, 2, 2, 1, 2, 3>, count (2, R) is 3.

Kullback-Leibler Divergence (KL)

Here, μ _R and μ _H are average values of the evaluation value series of R and H, respectively, and σ _R ² and σ _H ² are variances of R and H, respectively.

<Evaluation results>
FIG. 15 and FIG. 16 show the evaluation results for AD data and AL data (results of obtaining the average of each scale for each dialogue and taking the average). In addition, * in a figure shows that an average value is not inferior statistically compared with a nonpatent literature 3. FIG. +, †, and ‡ indicate that values (average values) are statistically superior to ergodic 0, ergodic 2, and connection 1. In addition, the value shown in bold represents the value with the best performance compared to other models. Here, ergodic 0 is the baseline and non-patent document 3 is the upper limit. In this evaluation experiment, the value of ergodic 1 was the same as that of ergodic 0. From these figures, it can be seen that the estimation accuracy of the links 1 and 2 is higher than that of the ergodics 0 to 2. For example, in the case of the evaluation scale DIST, the result equivalent to that of Non-Patent Document 3 is obtained for Ergodic 0. Therefore, it can be understood that a certain degree of estimation accuracy can be obtained with any model of ergodic 0 to 2 and connection 1 and 2.

  In addition, although it is thought that the method of nonpatent literature 3 is an upper limit (value which shows the best evaluation), in AD data, connection 1 and 2 have a statistically significant difference compared with an upper limit depending on an index. Absent. That is, it can be seen that the estimated performance is comparable. Although the AL data does not reach the upper limit, it can be seen that the connections 1 and 2 are approaching the upper limit as compared to ergodic 0 to 2. From this, it can be said that this invention can estimate the evaluation for every dialogue action only from the evaluation for every dialogue. Further, it can be seen that the HMM learned from all the dialogues is introduced into the model and the connected learning is effective so that the evaluation results of Ergodic 2 and Connections 1 and 2 are good.

  Ergodic 0 had exactly the same performance as Ergodic 1. This indicates that the HMM for each score has a sharp distribution, and the HMM learned from the entire dialogue has a broad distribution, so that the maximum likelihood path has passed only the HMM for each score. However, as shown in the results of Ergodic 2, it has also been found that this phenomenon can be avoided by increasing the number of HMM states learned from the entire dialogue.

  An ergodic 0 HMM obtained from AD data and a concatenated 1 HMM obtained from AL data are shown in FIGS. 17 and 18, respectively. Here, for the sake of brevity, FIG. 17 shows an output probability and transition probability of 0.05 or more, and FIG. 18 shows an output probability and transition probability of 0.1 or more. In FIG. 17, only the states corresponding to the evaluation value 1 (score: 1) and the evaluation value 7 (score: 7) are displayed. In FIG. 18, only the states corresponding to the evaluation value 0 (score: 0), the evaluation value 1 (score: 1), and the evaluation value 7 (score: 7) are displayed.

From these figures, it can be seen that transitions between states have a constant probability in ergodic 0, but that transition probabilities between states have values based on data by incorporating connected learning. Also, in FIG. 17, it can be seen that many states have the same output distribution, whereas in FIG. 18, the output distribution differs depending on the state. This is a result of the dialogue action sequence appearing in all dialogues being output from the hidden Markov model M ₀ and the output corresponding to each score being output from the corresponding state. Can think.

Finally, FIGS. 19 and 20 show examples of estimated evaluation value series of each model for AD data and AL data. In the ergodics 0 and 1, the evaluation values of the other models change more gently than the evaluation values greatly increase and decrease for each dialogue action. This is considered to be due to the fact that the evaluation value of 0 is assigned to each dialogue action without assigning any evaluation value because M ₀ functions effectively.

[Program, recording medium]
The various processes described above are not only executed in time series according to the description, but may also be executed in parallel or individually as required by the processing capability of the apparatus that executes the processes. Needless to say, other modifications are possible without departing from the spirit of the present invention.

  Further, when the above-described configuration is realized by a computer, processing contents of functions that each device should have are described by a program. The processing functions are realized on the computer by executing the program on the computer.

  The program describing the processing contents can be recorded on a computer-readable recording medium. As the computer-readable recording medium, for example, any recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, and a semiconductor memory may be used.

  The program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Furthermore, the program may be distributed by storing the program in a storage device of the server computer and transferring the program from the server computer to another computer via a network.

  A computer that executes such a program first stores, for example, a program recorded on a portable recording medium or a program transferred from a server computer in its own storage device. When executing the process, the computer reads a program stored in its own recording medium and executes a process according to the read program. As another execution form of the program, the computer may directly read the program from a portable recording medium and execute processing according to the program, and the program is transferred from the server computer to the computer. Each time, the processing according to the received program may be executed sequentially. Also, the program is not transferred from the server computer to the computer, and the above-described processing is executed by a so-called ASP (Application Service Provider) type service that realizes the processing function only by the execution instruction and result acquisition. It is good. Note that the program in this embodiment includes information that is used for processing by an electronic computer and that conforms to the program (data that is not a direct command to the computer but has a property that defines the processing of the computer).

  In this embodiment, the present apparatus is configured by executing a predetermined program on a computer. However, at least a part of these processing contents may be realized by hardware.

  INDUSTRIAL APPLICABILITY The present invention can be used for dialog analysis that analyzes dialog satisfaction in units of dialog act.

50, 100, 200, 300, 400 Dialog analyzer 51, 101, 201, 301, 401 Dialog learning device 90, 190, 390 Recording unit 110 Input / output unit 120, 320 First model learning unit 130, 230 Second model learning Units 140 and 240 Model connection unit 180 Evaluation value estimation units 340 and 440 First model connection unit 350 Re-learning unit 360 Second model connection unit 370 Iterative determination unit

Claims (12)

A dialogue learning device for creating a hidden Markov model for evaluating a dialogue composed of a plurality of dialogue actions,
K and N are integers of 2 or more, k is an integer of 1 to K, n is an integer of 1 to N,
Evaluation is performed using time series data (d ₁ ,..., D _N ) of _N conversations and evaluation values (r (d ₁ ),..., R (d _N )) obtained by evaluating each of the conversations at the K stage. for each of the stage, hidden Markov model M ₁ with a state of outputting the dialogue act of the _speaker, ..., to learn the M _K, the hidden Markov model M _1, ..., connecting all of the state of M _K to ergodic to, interacts learning apparatus comprising a first model learning unit for creating a hidden Markov model M _C. A dialogue learning device for creating a hidden Markov model for evaluating a dialogue composed of a plurality of dialogue actions,
K and N are integers of 2 or more, k is an integer of 1 to K, n is an integer of 1 to N,
Evaluation is performed using time series data (d ₁ ,..., D _N ) of _N conversations and evaluation values (r (d ₁ ),..., R (d _N )) obtained by evaluating each of the conversations at the K stage. for each of the stage, hidden Markov model M ₁ with a state of outputting the dialogue act of the _speaker, ..., to learn the M _K, the hidden Markov model M _1, ..., connecting all of the state of M _K to ergodic A first model learning unit for creating a hidden Markov model M _E0 ;
A second model learning unit that learns one hidden Markov model M ₀ having a state of outputting a speaker's dialogue action from all dialogues;
Connect all states of the said Markov model M ₀ Hidden Markov Models M _E0 to Ergodic, interactive learning device and a model connecting portion to create a hidden Markov model M _C. A dialogue learning device for creating a hidden Markov model for evaluating a dialogue composed of a plurality of dialogue actions,
K and N are integers of 2 or more, k is an integer of 1 to K, n is an integer of 1 to N,
Evaluation is performed using time series data (d ₁ ,..., D _N ) of _N conversations and evaluation values (r (d ₁ ),..., R (d _N )) obtained by evaluating each of the conversations at the K stage. A first model learning unit for learning a hidden Markov model M ₁ ,..., M _K having a state of outputting a speaker's dialogue action at each stage;
A second model learning unit that learns one hidden Markov model M ₀ having a state of outputting a speaker's dialogue action from all dialogues;
A first model concatenation that creates a hidden Markov model M _{1 + 0} ,..., M _{K + 0} by pairing each of the hidden Markov models M _k with a copy of the Markov model M ₀ and connecting all the states in the pair. And
The hidden Markov model M _{1 + 0} ,..., M _{K + 0} is an initial model, and the time series data (d ₁ ,..., D _N ) and the evaluation values (r (d ₁ ),..., R (d _N )) are learned data. A re-learning unit that re-learns the hidden Markov model M _{1 + 0} ,..., M _{K + 0}
Hidden relearned Markov model M _{1 + 0, ...,} concatenates M K + _0, and the second model connecting portions to one hidden Markov model M _C,
Hidden Markov Models M _{1 + 0, ...,} M for K + _0, determines the repetition determining unit whether to repeat the processing of the re-learning section and the second model coupling part by whether there is improvement of more than predetermined likelihood against training data A dialogue learning device comprising: The dialogue learning apparatus according to claim 2 or 3,
The Hidden Markov Model M ₀ is interactive learning device characterized by two states in each speaker corresponds. By using Hidden Markov Models M _C created claims 1 either by interaction learning apparatus 4, an interaction analysis apparatus for evaluating a dialogue composed of a plurality of dialogue act,
Using the Hidden Markov Model M _C, each dialogue act by estimating whether output from the state of the evaluation values of each dialogue act hidden Markov model M _C throat of the time-series data d _x dialogue evaluation value is not known A dialogue analysis apparatus comprising an evaluation value estimation unit that obtains an evaluation value of the dialogue and estimates the evaluation value of the dialogue from the evaluation value of each dialogue act. A dialogue learning method for creating a hidden Markov model for evaluating a dialogue composed of a plurality of dialogue actions,
K and N are integers of 2 or more, k is an integer of 1 to K, n is an integer of 1 to N,
Dialog learning device
Evaluation is performed using time series data (d ₁ ,..., D _N ) of _N conversations and evaluation values (r (d ₁ ),..., R (d _N )) obtained by evaluating each of the conversations at the K stage. for each of the stage, hidden Markov model M ₁ with a state of outputting the dialogue act of the _speaker, ..., to learn the M _K, the hidden Markov model M _1, ..., connecting all of the state of M _K to ergodic to, interacts learning how to perform first model learning step of creating a hidden Markov model M _C. A dialogue learning method for creating a hidden Markov model for evaluating a dialogue composed of a plurality of dialogue actions,
K and N are integers of 2 or more, k is an integer of 1 to K, n is an integer of 1 to N,
Dialog learning device
Evaluation is performed using time series data (d ₁ ,..., D _N ) of _N conversations and evaluation values (r (d ₁ ),..., R (d _N )) obtained by evaluating each of the conversations at the K stage. for each of the stage, hidden Markov model M ₁ with a state of outputting the dialogue act of the _speaker, ..., to learn the M _K, the hidden Markov model M _1, ..., connecting all of the state of M _K to ergodic A first model learning step for creating a hidden Markov model M _E0 ;
A second model learning step for learning one hidden Markov model M ₀ having a state of outputting a speaker's dialogue action from all dialogues;
Connect all states of the said Markov model M ₀ Hidden Markov Models M _E0 to Ergodic, interactive learning how to perform the model connection creating a hidden Markov model M _C. A dialogue learning method for creating a hidden Markov model for evaluating a dialogue composed of a plurality of dialogue actions,
K and N are integers of 2 or more, k is an integer of 1 to K, n is an integer of 1 to N,
Dialog learning device
Evaluation is performed using time series data (d ₁ ,..., D _N ) of _N conversations and evaluation values (r (d ₁ ),..., R (d _N )) obtained by evaluating each of the conversations at the K stage. A first model learning step for learning a hidden Markov model M ₁ ,..., M _K having a state of outputting a speaker's dialogue action for each stage of
A second model learning step for learning one hidden Markov model M ₀ having a state of outputting a speaker's dialogue action from all dialogues;
A first model concatenation that creates a hidden Markov model M _{1 + 0} ,..., M _{K + 0} by pairing each of the hidden Markov models M _k with a copy of the Markov model M ₀ and connecting all the states in the pair. Steps,
The hidden Markov model M _{1 + 0} ,..., M _{K + 0} is an initial model, and the time series data (d ₁ ,..., D _N ) and the evaluation values (r (d ₁ ),..., R (d _N )) are learned data. A re-learning step for re-learning a hidden Markov model M _{1 + 0} ,..., M _{K + 0}
Hidden Markov Models M _{1 + 0} which is _relearned, ..., concatenates M K + _0, and the second model linking step of a single hidden Markov model M _C,
Iterative determination step for determining whether to repeat the process of the re-learning step and the second model connection step depending on whether or not the likelihood of the learning data is improved by a predetermined value or more with respect to the hidden Markov model M _{1 + 0} ,..., M _{K + 0} Interactive learning method that executes and. The dialogue learning method according to claim 7 or 8, comprising:
The Hidden Markov Model M ₀ is interactive learning method characterized by two states in each speaker corresponds. By using Hidden Markov Models M _C created by any of the interactive learning method of claims 6 9, a dialogue analytical method for evaluating a dialogue composed of a plurality of dialogue act,
Dialogue analysis device
Using the Hidden Markov Model M _C, each dialogue act by estimating whether output from the state of the evaluation values of each dialogue act hidden Markov model M _C throat of the time-series data d _x dialogue evaluation value is not known A dialog analysis method that executes an evaluation value estimation step of obtaining an evaluation value of the dialogue and estimating an evaluation value of the dialogue from an evaluation value of each dialogue action.   A program that causes a computer to function as the dialogue learning apparatus according to claim 1.   A program causing a computer to function as the dialog analysis apparatus according to claim 5.

Images