JP2013045363A - Context dependency estimation device, speech clustering device, method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To estimate degrees of depending on context for dialog data.SOLUTION: Feature quantity of each speech is extracted from dialog data which is a time series of a plurality of speeches by a feature quantity extraction unit 30. The plurality of speeches are clustered by using a technique of a CRP on the basis of the extracted feature quantity of each speech by a CRP clustering unit 31. The plurality of speeches are clustered by using a technique of infinite an HMM on the basis of the extracted feature quantity of each speech by an infinite HMM clustering unit 32. Context dependency of the speeches is calculated on the basis of clustering results of the CRP and clustering results of the HMM by a context dependency calculation unit 23.

Description

  The present invention relates to a context dependency estimation apparatus, an utterance clustering apparatus, a method, and a program, and more particularly to a context dependency estimation apparatus, an utterance clustering apparatus, a method, and a program for clustering utterances with respect to conversation data.

  When constructing a dialogue system in a specific dialogue domain (where the domain represents the dialogue content / field / genre handled by the dialogue system, for example, flight reservation or conference room reservation), collect the dialogue data of the domain, Researchers and developers need to model the interaction domain interaction. For example, the vocabulary set determines what is good and what kind of utterance should be handled.

  What is particularly important in constructing a dialogue system is the latter phase of determining the type of utterance (also called dialogue action type or utterance action type), and there are many studies as shown in Non-Patent Document 1.

  However, in these studies, humans determine in advance the types of dialogue. In general, deciding what utterances exist in a domain and how many interaction actions are required requires detailed expert analysis and is expensive. Therefore, a technique is known in which utterances are clustered to automatically determine from the data what kind of utterances are clustered and how many interactive actions are required (Non-patent Document 2).

  The technique of Non-Patent Document 2 uses a technique called Chinese Restaurant Process (CRP) to cluster utterances and estimate the optimal number of dialogue actions. In this method, utterances during dialogue are regarded as independent, clustering is performed, and at the same time, the number of clusters (that is, the number of dialogue actions) is determined.

  There are several methods other than CRP that do not determine the number of clusters in advance. For example, a method called Affinity Propagation and a method called X-Means are known (Non-patent Documents 3 and 4). It is also possible to find the optimum number of clusters by repeatedly using a method for determining the number of clusters in advance (for example, K-Means). For example, with respect to a certain evaluation set, the number of clusters is increased little by little, and the number of clusters that provides the highest clustering accuracy is optimized. Here, the accuracy may be a general purity or F value (F-measure) in clustering evaluation.

A. Stolcke, N. Coccaro, R. Bates, P. Taylor, CV Ess-Dykema, K. Ries, E. Shriberg, D. Jurafsky, R. Martin, and M. Meteer, “Dialogue act modeling for automatic tagging and recognition of conversational speech, "Computational Linguistics, vol. 26, no. 3, pp. 339-373, 2000. N. Crook, R. Granell, and S. Pulman, “Unsupervised classification of dialogue acts using a Dirichlet process mixture model,” in Proc. SIGDIAL, 2009, pp. 341-348. Clustering by Passing Messages Between Data Points. Brendan J. Frey and Delbert Dueck, Science 315, 972--976, 2007. Dan Pelleg and Andrew Moore: X-means: Extending K-means with Efficient Estimation of the Number of Clusters.In Proc.ICML, 2000.

  In the method described in Non-Patent Document 2 described above, utterances during dialogue are regarded as independent during clustering. However, it is customary for dialogue data to consist of continuous utterances. In the prior art, context information important for such dialogue is not used, and the accuracy of clustering and the estimation of the number of dialogue actions are not sufficient.

  For example, “yes” can be both affirmative and affirmative, and it cannot be determined whether it is affirmative or affirmative unless it is from the context, but the prior art treats it as the same.

  In addition, knowing how much the utterances in the domain depend on the context is useful for constructing a dialogue system, but since the conventional technology regards utterances as independent, such knowledge is I can't get it.

  The present invention has been made in view of the above circumstances, and provides a context-dependent estimation device, method, and program capable of estimating the degree of dependence of conversation data on the context. Objective. It is a second object of the present invention to provide an utterance clustering apparatus and method capable of accurately clustering utterances with respect to dialogue data in consideration of the context.

  In order to achieve the above object, a context-dependent estimation apparatus according to the present invention includes a feature amount extracting unit that extracts a feature amount of each utterance from dialogue data that is a time series of a plurality of utterances, and the feature amount extracting unit. First clustering means for clustering the plurality of utterances based on the feature amount of each utterance extracted by the above, and context information of the utterance based on the feature amount of each utterance extracted by the feature amount extraction means And a second clustering unit that clusters the plurality of utterances, a clustering result obtained by the first clustering unit, and an estimation unit that estimates a degree depending on a context based on the clustering result obtained by the second clustering unit. And.

  A context dependency estimation method according to the present invention is a context dependency estimation method in a context dependency estimation apparatus including a feature amount extraction unit, a first clustering unit, a second clustering unit, and an estimation unit, wherein the context dependency The estimation apparatus extracts feature amounts of each utterance from dialogue data that is a time series of a plurality of utterances by the feature amount extraction unit, and each utterance extracted by the feature amount extraction unit by the first clustering unit. Clustering the plurality of utterances based on the feature amount of the utterance, and using the context information of the utterance based on the feature amount of each utterance extracted by the feature amount extraction means by the second clustering means, Clustering a plurality of utterances, and by the estimating means, the clustering result by the first clustering means, and the second Based on the clustering result by rastering means, and estimates the degree that depends on the context.

  According to the present invention, the feature quantity of each utterance is extracted from the dialogue data which is a time series of a plurality of utterances by the feature quantity extraction means. Then, the first clustering unit clusters the plurality of utterances based on the feature amount of each utterance extracted by the feature amount extracting unit. The second clustering unit clusters the plurality of utterances using the context information of the utterance based on the feature amount of each utterance extracted by the feature amount extracting unit.

  Then, the estimation unit estimates a degree depending on the context based on the clustering result by the first clustering unit and the clustering result by the second clustering unit.

  As described above, by clustering utterances without using utterance context information, and by clustering utterances using utterance context information, it is possible to estimate the degree of dependence of conversation data on the context. .

  The first clustering means according to the present invention clusters the plurality of utterances according to CRP (Chinese Restorant Process), and the second clustering means performs between the conversation data utterances according to an infinite HMM (Hidden Markov Model). The plurality of utterances can be clustered using transition information.

  The second clustering unit according to the present invention includes an additional feature amount obtained by adding the feature amount of the utterance immediately before the utterance as context information of the utterance to the feature amount of each utterance extracted by the feature amount extracting unit. The plurality of utterances can be clustered based on the generated additional feature amount of each utterance generated.

  The dialogue data is dialogue data relating to a specific domain, and the estimation means can estimate the context dependency of the utterance in the specific domain according to the following equation.

  However, the cluster number C1 is the number of clusters clustered by the first clustering means, and the cluster number C2 is the number of clusters clustered by the second clustering means.

  The dialogue data is dialogue data concerning two different domains, the first clustering means clusters a plurality of utterances of the dialogue data concerning the domain for each domain, and the second clustering means A plurality of utterances of dialogue data related to the domain are clustered, and the estimation means estimates a context dependency of the utterance for each domain and estimates a context dependency ratio of the domain according to the following equation: Can be.

  The estimation means according to the present invention can estimate the context dependency of the cluster C clustered by the first clustering means according to the following equation.

  Here, C ′ is a set of clusters clustered by the second clustering means, and c is a cluster that is an element of C ′.

  The estimation means for estimating the context dependency of the cluster estimates the context dependency of each cluster clustered by the first clustering means and is clustered by the first clustering means according to the following formula: The average context dependency of the cluster can be estimated.

  Here, c ″ is a cluster that is an element of C.

  An utterance clustering apparatus according to the present invention includes an input unit that receives dialogue data that is a time series of a plurality of input utterances, a dialogue data storage unit that stores the dialogue data received by the input unit, and a dialogue data A feature amount extracting unit for extracting a feature amount of each utterance; and a transition between utterances of the dialog data according to an infinite HMM (Hidden Markov Model) based on the feature amount of each utterance extracted by the feature amount extracting unit. And infinite HMM clustering means for clustering the plurality of utterances using information.

  The utterance clustering method according to the present invention is an utterance clustering method in an utterance clustering apparatus including an input means, an interaction data storage means, a feature amount extraction means, and an infinite HMM clustering means, wherein the utterance clustering apparatus is controlled by the input means. Receiving dialogue data which is a time series of a plurality of inputted utterances, storing the dialogue data received by the input means in the dialogue data storage means, and by using feature amount extraction means, features of each utterance from the dialogue data And the transition information between the utterances of the dialogue data is used according to the infinite HMM (Hidden Markov Model) based on the feature quantity of each utterance extracted by the feature quantity extraction means by the infinite HMM clustering means. Clustering the multiple utterances To do.

  An utterance clustering apparatus according to the present invention includes an input unit that receives dialogue data that is a time series of a plurality of input utterances, a dialogue data storage unit that stores the dialogue data received by the input unit, and a dialogue data A feature amount extracting unit that extracts a feature amount of each utterance, and an addition in which the feature amount of the utterance immediately before the utterance is added as the context information of the utterance to the feature amount of each utterance extracted by the feature amount extracting unit Context information adding means for generating each feature quantity; CRP clustering means for clustering the plurality of utterances according to CRP (Chinese Restorant Process) based on the additional feature quantity of each utterance generated by the context information adding means; , Including.

  The utterance clustering method according to the present invention is an utterance clustering method in an utterance clustering apparatus including an input means, a dialog data storage means, a feature amount extraction means, a context information addition means, and a CRP clustering means, wherein the utterance clustering device comprises: The dialogue means which is a time series of a plurality of utterances inputted by the input means is received, the dialogue data received by the input means is stored in the dialogue data storage means, and the dialogue data is obtained by the feature amount extraction means. Then, the feature amount of each utterance is extracted, and the feature amount of the utterance immediately before the utterance as the context information of the utterance is added to the feature amount of each utterance extracted by the feature information extracting unit by the context information adding unit. Each added feature is generated, and the sentence is added by the CRP clustering means. Based on the additional feature amount of each utterance generated by the information adding means, according to CRP (Chinese Restaurant Process), clustering the plurality of speech.

  The program according to the present invention is a program for causing a computer to function as each unit of the context dependency estimation apparatus.

As described above, according to the context-dependent apparatus, method, and program of the present invention, the utterances are clustered without using the utterance context information, and the utterances are clustered using the utterance context information. As a result, it is possible to estimate the degree of dependence on the conversation data depending on the context.
Further, according to the utterance clustering apparatus and method of the present invention, it is possible to obtain an effect that utterances can be clustered with high accuracy in consideration of the context of the conversation data.

It is the schematic which shows the structure of the context dependence estimation apparatus which concerns on the 1st Embodiment of this invention. It is a flowchart which shows the content of the context dependence estimation processing routine in the context dependence estimation apparatus which concerns on the 1st Embodiment of this invention. It is a figure which shows an example of dialog data. It is a figure which shows the example of a dialogue act. It is a figure which shows an example of dialog data. It is a figure which shows the example of a dialogue act. It is the schematic which shows the structure of the context dependence estimation apparatus which concerns on the 2nd Embodiment of this invention. It is a flowchart which estimates the degree which is dependent on context about the dialog data in the context dependence estimation apparatus which concerns on the 2nd Embodiment of this invention. It is the schematic which shows the structure of the context dependence estimation apparatus which concerns on the 3rd Embodiment of this invention.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

[First Embodiment]
<System configuration>
The context dependence estimation apparatus 100 according to the first embodiment of the present invention receives dialogue data that is a time series of a plurality of utterances related to a specific domain, and estimates and outputs the context dependence. This context dependency estimation apparatus 100 is constituted by a computer including a CPU, a RAM, and a ROM storing a program for executing a context dependency degree estimation processing routine to be described later. It is configured. As illustrated in FIG. 1, the context dependency estimation apparatus 100 includes an input unit 10, a calculation unit 20, and an output unit 28.

  The input unit 10 receives a plurality of interaction data related to a specific domain as input interaction data. Each dialogue data consists of a time series of a plurality of utterances. For example, the dialogue data is dialogue data between the dialogue system and a person, or dialogue data between people, and the data is natural language data that can be sequentially processed in time series, such as a natural language text or a speech recognition result.

  The calculation unit 20 includes a dialogue data storage unit 21, an utterance clustering unit 22, and a context dependency calculation unit 23. The context dependency calculation unit 23 is an example of an estimation unit.

  The dialogue data storage unit 21 stores a plurality of dialogue data received by the input unit 10.

  The utterance clustering unit 22 includes a feature amount extraction unit 30, a CRP clustering unit 31, and an infinite HMM clustering unit 32. The CRP clustering unit 31 is an example of a first clustering unit, and the infinite HMM clustering unit 32 is an example of a second clustering unit.

  The feature amount extraction unit 30 extracts a feature amount from each utterance in the input dialogue data. For example, the feature amount of bag-of-words is extracted. Bag-of-words is a feature quantity often used in natural language processing, and is a set with a frequency of words. In order to obtain this set, the feature quantity extraction unit 30 uses morphological analysis (ChaSen is used in the present embodiment), and obtains the feature quantity of bag-of-words for each utterance. Since low-frequency words may adversely affect clustering, only words that appear 10 times or more may be used as feature amounts for all data in each domain. A standard word is used as each word in the set. As the feature quantity, the frequency of only the content word or the frequency of only the function word may be used.

  The CRP clustering unit 31 clusters each utterance of the dialog data using the CRP method based on the feature amount of each utterance extracted by the feature amount extraction unit 30.

  CRP is a method for automatically determining the number of clusters from data, and performs clustering according to the following procedure. In CRP, data (ie, utterances) are called customers and clusters are called tables.

First, the first customer is placed on the first table. Then, the next customer (c _i ) sits on a table (t _j ) that already has a customer, or is expressed in the following table (1) in a new table (t _new ; new is the index of the new table) Sit with probability.

Here, 'n (t _j )' is a function that returns the number of customers attached to t _j , and N is the number of customers that have been on the table so far. Further, α is a hyperparameter indicating the degree to which a customer gets to a new table. The larger α is, the larger the number of clusters is. As a heuristic, the reciprocal of the estimated approximate number of clusters is used for α (for example, 0.01 if the assumed number of clusters is 100). P (c _i | t _j ) is the probability that c _i is generated from t _j . This probability is calculated according to the following equation (2).

Here, W is a set of feature values, and count (*, w) represents how many times the feature value w has occurred in the customer or table. β is a hyperparameter for preventing probability 0 and may be a sufficiently small number. For example, 0.0001. A uniform distribution is used for P (c _i | t _new ). After all customers are placed in order, the customers are rearranged using a technique called Gibbs sampling. This is to move a customer away from the table and place it again on another table (including a new table) or on the table that he / she originally used, and this rearrangement is the best arrangement for all customers. Repeat as many times as you want. If the customer arrangement does not change or sampling is performed a sufficient number of times such as 1000 times for each data, it is considered that the customer has converged, and the arrangement in the customer table at that time is taken as the clustering result.

  As described above, the CRP clustering unit 31 clusters the utterances (customers) into a plurality of clusters (tables) using the CRP technique as described above based on the feature amount of each utterance. An utterance and information on a cluster to which the utterance belongs are output.

  The infinite HMM clustering unit 32 clusters utterances using the infinite HMM technique based on the feature values of each utterance extracted by the feature value extraction unit 30.

  In the present embodiment, utterance clustering is performed using context information by a technique called infinite HMM. At that time, the number of clusters is automatically determined.

  The infinite HMM method is one of non-parametric Bayesian methods for estimating parameters from data, and an HMM that handles time-series data can handle an infinite state. The meaning that an infinite state can be handled means that the number of states is not predetermined, and the number of states is determined depending on data. Details of Mugen HMM are described in non-patent literature (Y. Teh, M. Jordan, M. Beal, and D. Blei, “Sharing clusters among related groups: Hierarchical Dirichlet processes,” in Proc. NIPS, 2004.). ing.

  In this embodiment, by using this infinite HMM, utterance clustering using context information is performed, and at the same time, the number of dialogue actions is estimated. Infinite HMM models a sequence of utterances. That is, this is a model in which an utterance is output from each state and transitions to the next state. It can be considered that context information (in particular, information on the immediately preceding utterance) is used to handle transitions between states (ie, transitions between utterance sets). Note that since a plurality of states are connected in the HMM, clustering is not necessarily performed depending only on the immediately preceding utterance.

  Here, a clustering method using an infinite HMM will be described. Since the infinite HMM performs clustering by a process similar to CRP, data will be referred to as a customer and the cluster will be referred to as a table.

In the infinite HMM, the customer c _i sits on a table t _j that has already arrived at the customer or a new table (t _{j = new} ) according to the probability expressed by the following equation (3).

Here, t _c represents a table where _c is seated. In infinite HMM, the customer may order a customer before and after the c _i, respectively and c _i-1 and c _{i + 1.} This corresponds to the order of utterances in the dialog data.

P (t _j , t _k ) is a transition probability between tables, and is obtained by the following equation (4).

Here, α is a hyperparameter that represents the degree of arrival of a customer at a new table, and K represents the number of tables that already have customers. transitions (t _j , t _k ) is the number of transitions from t _j to t _k , and γ is a hyperparameter for avoiding probability 0. It is sufficient if the number is sufficiently small. For example, 0.00001. The probability that the customer will arrive at the new table is expressed by the following equation (5).

Here, a uniform distribution is used for P (c _i | t _new ).

  As in the case of CRP, the customer arrangement is optimized using Gibbs sampling, and the finally obtained customer arrangement is used as the clustering result. As described above, the customer determines the table that the customer should arrive by looking at the table that the customer in front of him / her arrives, and infinite HMM performs clustering using context information.

  As described above, the infinite HMM clustering unit 32 clusters the utterances (customers) into a plurality of clusters (tables) using the infinite HMM technique as described above based on the feature amount of each utterance. And each utterance and information of the cluster to which the utterance belongs.

  As will be described below, the context dependency calculation unit 23 calculates the context dependency of the utterance and the context dependency of the cluster, which indicate the degree of dependency on the context.

  The clustering result using CRP is the result when the context is not seen, and the clustering result using the infinite HMM is the result when the context is seen. Therefore, it is possible to calculate how much the utterance depends on the context in the domain by looking at the difference in the number of clusters (estimated number of dialogue actions). Specifically, the context dependency calculating unit 23 calculates the context dependency of the utterance related to the domain of the conversation data according to the following equation (6).

  Further, by comparing the cluster by clustering using CRP and the cluster by clustering using infinite HMM, it is possible to know what kind of utterance is more dependent on the context. Specifically, it can be realized by examining how each data (utterance) of each cluster by clustering using CRP is allocated to each cluster by clustering using infinite HMM. If each data of one cluster of CRP is allocated to only one or a small number of clusters of infinite HMM, it can be said that the utterance of that cluster is not so context-dependent. However, if each data of one cluster of CRP is allocated to many clusters of infinite HMM, it can be said that it is a cluster depending on the context.

  Therefore, the context dependency calculation unit 23 calculates the context dependency of each cluster C of the clustering result by the CRP clustering unit 31 according to the following equation (7).

  Here, C ′ is a set of clusters resulting from clustering by the infinite HMM clustering unit 32, and c is each element (cluster) of the set of clusters C ′. P (c) is obtained according to the following equation (8).

  The above equation (7) is similar to the entropy equation in information theory, and takes a large value when the data (utterances) in the CRP cluster is scattered in many clusters of the infinite HMM. As a result, the context dependency of each cluster can be obtained. In other words, if this value is large, the utterances included in the cluster are considered to be highly context-dependent, and analyzing these utterances can lead to the construction of a dialogue system that is robust against context-dependent utterances. Become.

  For example, if multiple utterances belonging to a cluster with CRP correspond to many clusters of infinite HMM, the utterances in that cluster are similar on the surface but are likely to have different meanings depending on the context. .

  By taking only such utterances and analyzing them intensively, it is possible to connect to an understanding unit of a dialog system that can understand user utterances with high accuracy according to the context.

  Also, by taking the average of the context dependency of all clusters, the overall context dependency can also be calculated and used for analysis. Therefore, the context dependency calculation unit 23 calculates the average context dependency according to the following equation (9).

  Here, c ″ is each element of the cluster set C in the clustering result by the CRP clustering unit 31.

  The output unit 28 outputs the utterance context dependency, the cluster context dependency, and the average context dependency calculated by the context dependency calculation unit 23.

<Operation of context dependency estimation device>
Next, the operation of the context dependency estimation apparatus 100 according to the present embodiment will be described. First, when a plurality of time series of a plurality of utterances related to a certain domain are inputted as dialogue data to the context dependence estimation device 100, the plurality of dialogue data inputted by the context dependence estimation device 100 is converted into the dialogue data storage unit 21. Stored in Then, the context dependency estimation apparatus 100 executes the context dependency estimation processing routine shown in FIG.

  First, in step S101, bag-of-words feature amounts are extracted for all utterances of a plurality of dialogue data. In step S102, each utterance is classified into a plurality of clusters by clustering using a CRP method based on the feature amount of each utterance extracted in step S101.

  In the next step S103, each utterance is classified into a plurality of clusters by clustering using the infinite HMM method based on the feature amount of each utterance extracted in step S101.

  In step S104, based on the number of clusters in the clustering result in step S102 and the number of clusters in the clustering result in step S103, the context dependency of the utterance related to the domain is calculated according to the equation (6).

  In step S105, based on the data (utterance) belonging to each cluster in the clustering result in step S102 and the data (utterance) belonging to each cluster in the clustering result in step S103, according to the above equation (7), the CRP The context dependency of each cluster C is calculated by clustering using the technique. Further, based on the calculated context dependency of each cluster, the average context dependency of each cluster by clustering using the CRP technique is calculated according to the above equation (9).

  In step S106, the calculation results of steps S104 and S105 are output, and the context dependency calculation processing routine is terminated.

<Example>
Examples are shown below. An explanation will be given by using an example of clustering dialogue data between a dialogue system and a human and dialogue data between humans. The data used here is data collected through the chat interface, and is data of text dialogue.

  The dialogue data between the dialogue system and the human is data in which the dialogue system and the human are talking, and a total of 1000 pieces of dialogue data. In the dialogue, the system and humans discuss animal likes and dislikes.

  This domain will be referred to as an Animal Discussion (AD) domain. An example of interaction is shown in FIG. In FIG. 3, U represents a user utterance and S represents a system utterance. In parentheses are dialogue action types in this domain. In this domain, 29 interactive actions as shown in FIG. 4 are defined manually.

  For details of each dialogue act, please refer to non-patent literature (Ryuichiro Higashinaka, Koji Dosaka, Hideki Amagasaki, “Effects of empathy and self-disclosure in dialogue system”, Proc. -449, 2009.).

  In addition, the dialogue data between humans is a collection of dialogues of listeners. This domain is referred to as an Attentive Listening (AL) domain. The listener dialogue is a dialogue in which two people are divided into a listener and a speaker, and one of them becomes a listener and listens to the story of the speaker. As human interaction data, 1260 such interaction data were collected. An example of interaction is shown in FIG. In FIG. 5 above, S represents a talking role, and L represents a listening role. In the parentheses are interactive action types in this domain, and 38 interactive actions as shown in FIG. 6 are manually defined.

  For details of each dialogue act, see non-patent literature (T. Meguro, R. Higashinaka, Y. Minami, and K. Dohsaka, “Controlling listening-oriented dialogue using partially observable Markov decision processes,” in Proc. COLING, 2010, pp. 761-769.) for details.

  As a comparison target, a clustering technique called K-means was used. K-means is a representative clustering method used when the number of clusters is known in advance. In this method, first, clusters are randomly generated, and the clusters are updated until a locally optimal solution is obtained by the framework of the EM algorithm.

  An experiment was conducted to estimate the number of interactive actions by clustering AD domain and AL domain data using the K-means, CRP, and infinite HMM methods. Here, since K-means is a technique that cannot estimate the number of interactive actions, a direct comparison cannot be made. Therefore, for K-means, the number of correct interactive actions was given manually, and utterances were clustered. Since clustering is performed in a state where the number of dialogue actions is known in advance, it can be regarded as a very powerful baseline.

  Since the calculation cost of Gibbs sampling is relatively high, 50 dialogue data extracted at random from each domain were used for the experiment. The AD domain dialogue data was 2894 utterance data, and the AL domain dialogue data was 2470 utterance data. According to the interactive actions given manually, these subsets included 27 kinds of interactive actions and 33 kinds of interactive actions, respectively.

  Before clustering, morpheme analysis was used (ChaSen was used in this experiment), and the feature amount of bag-of-words was obtained for each utterance. However, since low-frequency words may have an adverse effect on clustering, only words that appear 10 times or more are used as feature quantities for all data in each domain, and the standard form is used for the words.

  Clustering was evaluated by contrasting the correct dialogue action labels manually assigned to the data.

  As an evaluation scale, purity and F-measure were used. Both are general indicators for clustering evaluation. “purity” indicates how much the same dialogue act is spoken in one cluster, and “F-measure” pays attention to the pair of data and quantifies how correctly the pair that should be in the same cluster is in the same cluster. Turn into. The purity is calculated by the following equation (10).

Here, C = {c ₁ ,..., C _K } is a set of clusters, D = {d 1,..., DN} is a set of dialogue actions, and N is the number of data (number of utterances). It is.

  F-measure is calculated according to the following equation (11).

  Here, TP, FP, and FN represent true positive, false positive, and false negative, respectively. true positive is the number of times a pair of utterances with the same dialogue act is in the same cluster, false positive is the number of times a pair of utterances with a different dialogue act is in the same cluster, and false negative is the same dialogue This is the number of times that a pair of utterances that are actions is in different clusters.

  Since K-means depends on a random initial value and CRP and infinite HMM operate stochastically, in this experiment, clustering was performed 100 times with each clustering method, and the average value was obtained. For CRP and infinite HMM, 0.1 was used for α and 0.01 for β and γ. The number of iterations of Gibbs sampling was 100. In other words, all customers were relocated 100 times.

  Table 1 below shows the results of clustering AD domain utterances.

＊は、Ｋ−ｍｅａｎｓに対してｔ検定により１％の有意水準で差があることを示す。＋は、ＣＲＰに対してｔ検定により１％の有意水準で差があることを示す。
また、以下の表２に、ＡＬドメインの発話のクラスタリング結果を示す。

* Indicates that there is a difference at a 1% significance level by t-test with respect to K-means. + Indicates that there is a difference at 1% significance level by t-test against CRP.
Table 2 below shows the clustering result of the utterances of the AL domain.

  As can be seen from the above results, the infinite HMM has better clustering performance than other methods. That is, it was found useful to use context information for utterance clustering.

  The number of infinite HMM clusters is about 143 in the AD domain and 38 in the AL domain. This is the automatically estimated number of interactive actions.

  The number of interactive actions estimated by CRP is closer to the number given manually. From this, it can be inferred that the act of giving a dialogue act by a human being is regarded as an independent utterance. However, according to the result of this experiment, the number of dialogue actions when the context is considered is larger than the number of correct dialogue actions manually. This seems to suggest that the number of human interaction acts may be too small in view of the context. In other words, a dialog system designer can take it as a guideline for redesigning a dialog act, such as adding a dialog act that takes context into consideration, and by redesigning the dialog act according to that guideline, It will be possible to connect to a dialogue system that can properly handle user utterances.

  When the context dependency of the utterance is calculated, the context dependency of the utterance with respect to the AD domain is 143.62 / 35.03 = 4.099. The context dependency on the AL domain is 38.00 / 29.28 = 1.298. From this, it can be determined that the utterance of the AD domain is higher in context dependency in the domain. In addition, when the context dependency ratio described below is calculated by comparing the context dependency in each domain, the AD domain is 4.099 / 1.298 = 3.158 times that of the AL domain, and there are many domain-dependent dialog acts. It was found as an objective numerical value.

  As described above, according to the context dependency estimation apparatus according to the present embodiment, the utterances of conversation data are clustered and the context information of the utterance is considered using the CRP technique that does not consider the context information of the utterance. By clustering utterances of dialogue data using the infinite HMM method, and comparing the utterance clustering results, it is possible to estimate the context dependency of dialogue data of a certain domain.

  In addition, by considering the context information of utterances in utterance clustering, the performance of utterance clustering is improved, so it becomes possible to see at a glance what kind of utterances exist in the domain, making it easy to construct a dialogue system. Become. Furthermore, since the context dependency of utterances in a domain can be calculated as a numerical value, this leads to a deep understanding of the domain of dialogue data. For example, if it is known that the domain has a lot of utterances with a high degree of context dependency, it is possible to improve the system so that the understanding unit has more context information.

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

  The second embodiment is different from the first embodiment in that clustering using a CRP technique is performed based on the feature amount of an utterance to which context information is added.

  As shown in FIG. 7, the utterance clustering unit 222 of the context dependency estimation apparatus 200 according to the second embodiment includes a feature amount extraction unit 30, a CRP clustering unit 31, a context information addition unit 231, and a CRP clustering unit 232. It has. The CRP clustering unit 31 is an example of a first clustering unit, and the CRP clustering unit 232 is an example of a second clustering unit.

  The context information adding unit 231 adds the feature amount of the immediately previous utterance as context information to the feature amount of each utterance extracted by the feature amount extracting unit 30 to generate each additional feature amount. For example, if there are utterances 1,..., Utterance N, and the feature amounts are feature amounts 1,..., Feature amount N, the feature amount of the previous utterance is added to the feature amount of each utterance. That is, {start symbol, feature amount 1}, {feature amount 1, feature amount 2},..., {Feature amount N-1, feature amount N} are generated as additional feature amounts for each utterance. As a result, the dimension of the feature quantity (vector) is doubled.

  Similar to the CRP clustering unit 31, the CRP clustering unit 232 clusters the utterances using the CRP based on the additional feature amount of each utterance generated by the context information adding unit 231.

  Based on the number of clusters in the clustering result by the CRP clustering unit 31 and the number of clusters in the clustering result by the CRP clustering unit 231, the context dependency calculating unit 23 calculates the context dependency according to the same formula as the above formula (6). To do.

  In addition, the context dependence calculation unit 23 follows the same formula as the formula (7) based on the data of each cluster in the clustering result by the CRP clustering unit 31 and the data of each cluster in the clustering result by the CRP clustering unit 231. The context dependency of each cluster C of the clustering result by the RP clustering unit 31 is calculated. In addition, the context dependency calculation unit 23 calculates the average context dependency of the cluster of the clustering result by the CRP clustering unit 31 according to the same expression as the expression (9).

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

  First, in step S101, feature amounts are extracted for all utterances of a plurality of dialogue data. In step S102, each utterance is classified into a plurality of clusters by clustering using a CRP method based on the feature amount of each utterance extracted in step S101.

  In the next step S201, the feature amount of the immediately preceding utterance is added to the feature amount of each utterance extracted in step S101, and an additional feature amount is generated for each utterance.

  In step S202, each utterance is classified into a plurality of clusters by clustering using the CRP method based on the additional feature amount of each utterance generated in step S201.

  In step S104, based on the number of clusters in the clustering result in step S102 and the number of clusters in the clustering result in step S202, the context dependency of the utterance related to the domain according to the same expression as the above expression (6) Is calculated.

  In step S105, based on the data (utterance) belonging to each cluster in the clustering result in step S102 and the data (utterance) belonging to each cluster in the clustering result in step S202, an expression similar to the above expression (7) Accordingly, the context dependency of each cluster C by the clustering in step S102 is calculated. Further, based on the calculated context dependency of each cluster, the average context dependency of each cluster by the clustering in step S102 is calculated according to the same expression as the above expression (9).

  In step S106, the calculation results of steps S104 and S105 are output, and the context dependency calculation processing routine is terminated.

  As described above, according to the context dependency estimation apparatus according to the present embodiment, the conversation data utterances are clustered using the CRP technique that does not consider the utterance context information, and the immediately preceding utterance is used as the context information. By using the additional feature amount to which the feature amount is added, the dialogue data utterances are clustered, and the utterance clustering results are compared, whereby the context dependency of the dialogue data of a certain domain can be estimated.

[Third Embodiment]
Next, a third embodiment will be described.

  In the third embodiment, utterance clustering is performed on the conversation data for each of a plurality of domains, and the context dependency ratio of the domain is calculated. The form is different. In the following, a case where the same method as that of the first embodiment is used as a method for calculating the context dependency of each domain of a plurality of domains will be described as an example, but a method of calculating the context dependency of each domain is described. As an alternative, the same method as in the second embodiment may be used.

  As illustrated in FIG. 9, the context dependency estimation apparatus 300 according to the third embodiment includes input units 10 </ b> A and 10 </ b> B, a calculation unit 20, and an output unit 28.

  The input unit 10A receives a plurality of dialogue data related to the domain A as the inputted dialogue data. The input unit 10B accepts a plurality of pieces of dialogue data related to a domain B different from the domain A as inputted dialogue data.

  The calculation unit 20 includes dialogue data storage units 21A and 21B, utterance clustering units 22A and 22B, context dependency calculation units 23A and 23B, and a context dependency ratio calculation unit 323.

  The dialogue data storage unit 21A stores a plurality of dialogue data received by the input unit 10A. The dialogue data storage unit 21B stores a plurality of dialogue data received by the input unit 10B.

  The CRP clustering unit 31 of the utterance clustering unit 22A clusters the utterances for the domain A using the CRP technique based on the feature amount of each utterance of the domain A dialogue data extracted by the feature amount extraction unit 30. . The CRP clustering unit 31 of the utterance clustering unit 22B clusters utterances for the domain B using the CRP technique based on the feature amount of each utterance of the domain B extracted by the feature amount extraction unit 30.

  The infinite HMM clustering unit 32 of the utterance clustering unit 22A performs an utterance on the domain A using the infinite HMM method based on the feature amount of each utterance of the domain A dialogue data extracted by the feature amount extraction unit 30. Clustering. The infinite HMM clustering unit 32 of the utterance clustering unit 22B performs an utterance on the domain B using the infinite HMM method based on the feature amount of each utterance of the domain B dialogue data extracted by the feature amount extraction unit 30. Clustering.

  The context dependency calculation unit 23A calculates the context dependency of the utterance related to the domain A according to the above equation (6). The context dependency calculation unit 23B calculates the context dependency of the utterance related to the domain B according to the above equation (6).

  The context-dependent ratio calculation unit 323 calculates the context-dependent ratio between domain A and domain B according to the following equation (12) based on the calculated context dependency of utterance related to domain A and the context dependency of utterance related to domain B. To do.

  In addition, the context dependency calculation unit 23A calculates the context dependency of each cluster C of the clustering result by the CRP clustering unit 31 of the utterance clustering unit 22A for the domain A according to the above equation (7). For the domain B, the context dependency calculation unit 23B calculates the context dependency of each cluster C of the clustering result by the CRP clustering unit 31 of the utterance clustering unit 22B.

  The context dependency calculation unit 23A calculates an average context dependency for the domain A according to the above equation (9). The context dependency calculation unit 23B calculates the average context dependency for the domain B according to the above equation (9).

  Note that the other configuration and operation of the context dependency estimation apparatus according to the third embodiment are the same as those of the first embodiment, and thus the description thereof is omitted.

  Thus, the context dependency ratio of the domain can be calculated by calculating and comparing the context dependency of the utterance for each domain.

  Note that the present invention is not limited to the above-described embodiment, and various modifications and applications are possible without departing from the gist of the present invention.

  For example, the case where the CRP method is used as the clustering method that does not consider the context of the entire dialogue has been described as an example. However, the clustering method is not limited to this as long as the clustering method does not consider the context of the entire dialogue. For example, utterance clustering may be performed using a method such as K-Means, Affinity Propagation, or X-Means.

  Note that the method of automatically determining the number of clusters and clustering utterances using context information can consider the relationship between utterances compared to the conventional clustering method that does not consider the context of the entire conversation. There is an advantage of high accuracy.

  The input unit 10, the dialogue data storage unit 21, the feature amount extraction unit 30, and the infinite HMM clustering unit 32 described in the first embodiment can be extracted and function as an utterance clustering device. Similarly, the input unit 10, the dialog data storage unit 21, the feature amount extraction unit 30, the context information addition unit 231, and the CRP clustering unit 232 described in the second embodiment are extracted and functioned as an utterance clustering device. Can do. By automatically determining the number of clusters and clustering utterances using context information, there is an advantage that highly accurate clustering becomes possible.

  In the present specification, the embodiment has been described in which the program is installed in advance. However, the program can be provided by being stored in a computer-readable recording medium.

10, 10A, 10B Input unit 20 Arithmetic unit 21, 21A, 21B Dialog data storage unit 22, 22A, 22B, 222 Utterance clustering unit 23, 23A, 23B Context-dependent calculation unit 30 Feature quantity extraction unit 31, 232 CRP clustering unit 32 Infinite HMM clustering unit 100, 200, 300 Context dependency estimation device 231 Context information addition unit 323 Context dependency ratio calculation unit

Claims (13)

Feature amount extraction means for extracting feature amounts of each utterance from dialogue data that is a time series of a plurality of utterances;
First clustering means for clustering the plurality of utterances based on the feature quantity of each utterance extracted by the feature quantity extraction means;
Second clustering means for clustering the plurality of utterances using context information of the utterance based on the feature quantity of each utterance extracted by the feature quantity extraction means;
Estimating means for estimating the degree of dependence on the context based on the clustering result by the first clustering means and the clustering result by the second clustering means;
Context-dependent estimation device including The first clustering means clusters the plurality of utterances in accordance with CRP (Chinese Restaurant Process).
The context dependence estimation apparatus according to claim 1, wherein the second clustering unit clusters the plurality of utterances using transition information between utterances of the conversation data according to an infinite HMM (Hidden Markov Model).   The second clustering unit generates an additional feature amount obtained by adding the feature amount of the utterance immediately before the utterance as context information of the utterance to the feature amount of each utterance extracted by the feature amount extracting unit, The context dependence estimation apparatus according to claim 1, wherein the plurality of utterances are clustered based on the generated additional feature amount of each utterance. The interaction data is interaction data regarding a specific domain,
4. The context dependency estimation apparatus according to claim 1, wherein the estimation unit estimates a context dependency of the utterance in the specific domain according to the following expression. 5.

However, the cluster number C1 is the number of clusters clustered by the first clustering means, and the cluster number C2 is the number of clusters clustered by the second clustering means. The interaction data is interaction data regarding two different domains;
The first clustering means clusters, for each domain, a plurality of utterances of dialogue data related to the domain,
The second clustering means clusters, for each domain, a plurality of utterances of dialogue data related to the domain,
5. The context dependency estimation apparatus according to claim 4, wherein the estimation means estimates the context dependency of the utterance for each domain and estimates the context dependency ratio of the domain according to the following equation.

The context estimation device according to any one of claims 1 to 3, wherein the estimation unit estimates the context dependency of the cluster C clustered by the first clustering unit according to the following expression.

Here, C ′ is a set of clusters clustered by the second clustering means, and c is a cluster that is an element of C ′. The estimation means estimates the context dependence of each cluster clustered by the first clustering means, and estimates the average context dependence of the clusters clustered by the first clustering means according to the following formula: 6. The context-dependent estimation device according to 6.

Here, c ″ is a cluster which is an element of C. A context dependency estimation method in a context dependency estimation apparatus including a feature amount extraction unit, a first clustering unit, a second clustering unit, and an estimation unit,
The context-dependent estimation device includes:
The feature value extraction means extracts feature values of each utterance from dialogue data that is a time series of a plurality of utterances,
Clustering the plurality of utterances based on the feature amount of each utterance extracted by the feature amount extraction means by the first clustering means;
Based on the feature quantity of each utterance extracted by the feature quantity extraction means by the second clustering means, the utterance context information is used to cluster the plurality of utterances,
A context-dependent estimation method characterized in that the estimation unit estimates a context-dependent degree based on a clustering result by the first clustering unit and a clustering result by the second clustering unit. Input means for receiving dialogue data that is a time series of a plurality of input utterances;
Dialog data storage means for storing the dialog data received by the input means;
Feature amount extraction means for extracting feature amounts of each utterance from the dialogue data;
Infinite HMM clustering means for clustering the plurality of utterances using transition information between utterances of the dialogue data according to an infinite HMM (Hidden Markov Model) based on the feature quantity of each utterance extracted by the feature quantity extraction means. When,
Utterance clustering device. Input means for receiving dialogue data that is a time series of a plurality of input utterances;
Dialog data storage means for storing the dialog data received by the input means;
Feature amount extraction means for extracting feature amounts of each utterance from the dialogue data;
Context information adding means for generating each additional feature quantity obtained by adding the feature quantity of the utterance immediately before the utterance as context information of the utterance to the feature quantity of each utterance extracted by the feature quantity extracting means;
CRP clustering means for clustering the plurality of utterances according to a CRP (Chinese Restaurant Process) based on the additional feature amount of each utterance generated by the context information adding means;
Utterance clustering device. An utterance clustering method in an utterance clustering device including an input means, an interaction data storage means, a feature amount extraction means, and an infinite HMM clustering means,
The utterance clustering device comprises:
Accepting dialogue data that is a time series of a plurality of utterances input by the input means,
Storing the dialog data received by the input means in a dialog data storage means;
The feature amount extraction means extracts the feature amount of each utterance from the dialogue data,
Based on the feature amount of each utterance extracted by the feature amount extraction means by the infinite HMM clustering means, the plurality of utterances using transition information between utterances of the dialog data according to an infinite HMM (Hidden Markov Model). An utterance clustering method characterized by clustering. An utterance clustering method in an utterance clustering device including an input means, an interaction data storage means, a feature amount extraction means, a context information addition means, and a CRP clustering means,
The utterance clustering device comprises:
Accepting dialogue data that is a time series of a plurality of utterances input by the input means,
Storing the dialogue data received by the input means in the dialogue data storage means;
The feature amount extraction means extracts the feature amount of each utterance from the dialogue data,
The context information adding means generates an additional feature quantity obtained by adding the feature quantity of the utterance immediately before the utterance as context information of the utterance to the feature quantity of each utterance extracted by the feature quantity extracting means,
An utterance clustering method comprising: clustering the plurality of utterances according to a CRP (Chinese Restorant Process) by the CRP clustering means based on an additional feature amount of each utterance generated by the context information adding means.   The program for functioning a computer as each means of the context dependence estimation apparatus of any one of Claims 1-7.

Images