JP6192166B2 - Opinion type estimation device and program thereof - Google Patents

Description

  The present invention relates to an opinion type estimation device for estimating statistical information related to an opinion type from comment data representing the content of the opinion, and a program thereof.

  Conventionally, natural language processing has been required to analyze writing by human hands. Prior to the natural language processing, a technique for estimating the opinion type is required (Patent Documents 1 to 5 and Non-Patent Documents 1 and 2). This opinion type estimation technique is an algorithm that assigns a label for each opinion type. In the case of an opinion type that indicates approval or disapproval, a label “agree” or “disagree” is assigned to an opinion.

Japanese Patent No. 5224532 Japanese Patent No. 5254888 Japanese Patent No. 5278327 Japanese Patent No. 5286298 Japanese Patent No. 5221551

Jorge Nocedal, Stephen J. Wright, “Numerical Optimization”, Springer, 2006 Yuko Otsuka, Takashi Miki, Manabu Okumura, “Opinion Analysis Engine”, Corona, 2007

  However, the conventional opinion type estimation technique includes a certain amount of error such as giving an inappropriate label to an opinion. For this reason, the conventional opinion type estimation technique has a problem that, when this technique is used alone, the estimation accuracy is not improved when obtaining statistical information related to the opinion type such as the number of pros and cons.

  Here, the cause of the error included in the opinion type estimation technique is examined. In the process of forming each individual's opinion, it is often the case that one's opinion (pros and cons) is determined with reference to the opinions of others. Nevertheless, the conventional opinion type estimation technology does not reflect the actual opinion formation process, and is considered to cause errors.

  Therefore, an object of the present invention is to provide an opinion type estimation device and a program for solving the above-described problems and improving the estimation accuracy of statistical information related to the opinion type.

  In view of the above-described problems, the opinion type estimation device according to the present invention is an opinion type estimation device in which an opinion type is set in advance in comment data representing the content of an opinion, and statistical information regarding the opinion type of the comment data is estimated. And a mixed distribution model storage unit, a parameter optimization unit, and an opinion type estimation unit.

  According to this configuration, the opinion type estimation device uses the mixture distribution model storage means to represent a process of forming an opinion depending on the utterance time, with the mixture ratio, peak time, and rapidity of each opinion type as parameters. The distribution model is stored in advance, and the speech data to which the speech time is added is stored for each opinion type. Further, the opinion type estimation device applies the utterance data to the mixed distribution model of the mixed distribution model storage unit by the parameter optimization unit, and estimates the parameter by the numerical optimization method.

  The opinion type estimation device estimates the statistical information related to the opinion type using the mixed distribution model because the parameter is optimized by the mixed type distribution model representing the opinion formation process by the opinion type estimation unit. For example, the statistical information regarding the opinion type is the ratio of the comments of the opinion type included in all the comment data, or the number of comments for each opinion type.

The speech data is text data representing the content of an opinion made by an unspecified majority person on a network regarding a specific case in a field such as politics and economy.
The opinion type is a classification of the opinion content of the utterance data into a preset type. For example, as the opinion type, labels such as “agree” and “disagree” indicating approval / disapproval of the opinion can be given.
The mixing ratio for each opinion type represents the ratio of each opinion type in the mixture distribution model. For example, the mixing ratio for each opinion type represents a ratio of “agree” and a ratio of “opposite”.

The distribution model for each opinion type is a distribution model having one or more peaks depending on the utterance time (for example, Gaussian distribution model). In other words, since the distribution model for each opinion type depends on the utterance time, it can be said that the opinion formation process is reflected.
The mixed distribution model is a probability model obtained by mixing distribution models for each opinion type. For example, as the mixed distribution model, a two-mixed distribution model including a “promise” distribution model and an “opposite” distribution model can be cited.

The excitement is the speed with which the speech increases on the network. For example, if the number of statements about a particular case on the network is increasing rapidly, it is said to be exciting. Also, for example, if the state where there are few statements about a specific case continues on the network, it is said that it is not exciting.
The cause of the increase in the number of utterances on the network may be an increase in the number of speakers due to diffusion or an increase in the number of utterances repeatedly by the same person. Although these two causes are mixed, it is possible to eliminate the influence of repeated speech by the same person by referring to the speech of the same person only once. In this case, an increase in speech represents diffusion.

  The opinion type estimation apparatus according to the present invention can also be realized by an opinion type estimation program that causes hardware resources such as a CPU, a memory, and a hard disk included in a computer to operate in cooperation with each other as described above. This program may be distributed through a communication line, or may be distributed by writing in a recording medium such as a CD-ROM or a flash memory.

According to the present invention, the following excellent effects can be obtained.
According to the present invention, since the statistical information about the opinion type is estimated by optimizing the parameters of the mixed distribution model representing the opinion formation process, the estimation accuracy can be improved. Thus, according to the present invention, since the opinion type can be automatically added to the opinion on the network, the labor for manually adding the opinion type can be reduced.

It is a block diagram which shows the structure of the speech analysis apparatus which concerns on embodiment of this invention. It is explanatory drawing explaining the data structure of the utterance data which the opinion classification addition means of FIG. 1 output. It is explanatory drawing explaining the mixture distribution model memorize | stored in the mixture distribution model management means of FIG. It is a block diagram which shows operation | movement of the speech analysis apparatus of FIG. In the modification of this invention, it is explanatory drawing explaining a distribution model.

With reference to FIG. 1, an embodiment of the present invention will be described in detail with reference to the drawings as appropriate.
As shown in FIG. 1, a speech analysis apparatus 1 according to an embodiment of the present invention analyzes a speech that exists on a network, and includes an opinion type estimation device 2 and a speech analysis means 30.

[Configuration of speech analysis device]
The speech analysis apparatus 1 receives speech data existing on the network. This comment data is, for example, an opinion written on a homepage, blog or bulletin board. Here, it is assumed that the utterance data is time-series data added with the utterance time (the time when the opinion is written) and totaled (sorted) by the utterance time. Then, the utterance analysis device 1 outputs the utterance data to the opinion type estimation device 2 in order to estimate the ratio of the opinion type as a previous stage of the speech analysis (natural language processing).

  The opinion type estimation device 2 estimates the ratio of the opinion type from the utterance data input from the utterance analysis device 1 as statistical information regarding the opinion type. Therefore, the opinion type estimation device 2 includes an analysis object selection unit 21, an opinion type addition unit 22, a mixed distribution model management unit (mixed distribution model storage unit) 23, a parameter optimization unit 24, and an opinion type estimation unit. 25.

The analysis target selection means 21 selects the speech data to be analyzed from all the speech data input from the speech analysis device 1. For example, when the analysis target selection means 21 wants to analyze the opinion of the policy of consumption tax increase, it narrows down from all the comment data to the comment data related to the consumption tax increase. More specifically, the analysis target selection unit 21 manually sets a keyword (for example, “consumption tax increase”) characteristic in the case to be analyzed, and performs processing for selecting the utterance data including the keyword. . Then, the analysis target selecting unit 21 outputs the selected message data to the opinion type adding unit 22.
Note that the analysis target selection unit 21 may combine the processing of selecting the utterance data including the keyword with the processing of regarding the set keyword notation fluctuation or synonym as the same word.

The opinion type adding unit 22 adds an opinion type to the utterance data input from the analysis target selecting unit 21. In the present embodiment, the opinion type adding unit 22 adds the approval / disapproval of “agree” or “disagree” as the opinion type to the comment data. At this time, the opinion type adding means 22 uses the conventional opinion type estimation technique to add the opinion type to the utterance data. More specifically, the opinion type adding means 22 uses the techniques described in Section 4.2 “Automatic Extraction of Sentences Containing Opinion” and 4.3 “Evaluation Analysis Elemental Techniques” in Reference Document 1, The type can be added to the message data.
Reference 1: Yuko Otsuka, Takashi Miki, Manabu Okumura, “Opinion Analysis Engine”, Corona, 2007

  Here, since the opinion type is necessary for the comment data when the optimization method is applied to the mixed distribution model, the opinion type is added to the comment data. Naturally, since the conventional opinion type estimation technique is used, an error is included in the opinion type added to the utterance data.

  In addition, the speech time is added to the speech data when it is input to the opinion type estimation device 2. Therefore, as shown in FIG. 2, the utterance data includes the utterance time, the utterance content, and the opinion type. FIG. 2 illustrates remark data regarding economic policies. For example, the first opinion is written at 10:34 on September 19, 2013, and is an opinion “I want you to take measures against the economy” and indicates that you agree with economic policy. The second opinion, written at 10:35 on September 19, 2013, is an opinion that “I want the Prime Minister to quit economic measures” and that it is “opposite” for economic policy.

  Thereafter, the opinion type adding unit 22 outputs the utterance data to which the opinion type is added to the mixed distribution model management unit 23 for each utterance type. At this time, the opinion type adding means 22 instructs the parameter optimization means 24 via the mixture distribution model management means 23 (parameter optimization instruction).

  The mixed distribution model management unit 23 stores and manages the message data input from the opinion type adding unit 22 and the mixed distribution model.

<Mixed distribution model>
The mixed distribution model will be described with reference to FIG. 3 (see FIG. 1 as appropriate).
The mixed distribution model is a mixture of distribution models set for each opinion type. In the present embodiment, the mixed distribution model is a two-mixed distribution model obtained by mixing the “agree” distribution model 90 and the “opposite” distribution model 91 as shown in FIG.

In the present embodiment, the distribution models 90 and 91 are Gaussian distribution models having one or more peaks depending on the utterance time. The Gaussian distribution model is represented by the following formula (1). In Equation (1), t ₀ is the time (peak time) when the opinion peaks for each opinion type, and α ₀ represents the rapidity of the opinion.

Accordingly, each of the distribution models 90 and 91 can be expressed by the following formula (1) ′. In Expression (1) ′, n represents the number of opinion types (where n is a natural number). Further, in the expression (1) ′, the peak time t _n of the opinion type n and the rapidity α _n of the opinion type n are expressed. In this embodiment, n = 1 and 2, and when n = 1, the “agree” distribution model 90 is represented, and when n = 2, the “opposite” distribution model 91 is represented.

Since the speech analysis apparatus 1 is intended to analyze speech data in which “agree” and “opposite” are mixed, the “agree” and “opposite” distribution models 90 and 91 are mixed (distribution model). Is called a mixture distribution model). This two-mix distribution model can be expressed by the following formula (2). Since this expression (2) includes the peak time t ₁ of “agree” and the peak time t ₂ of “opposite”, it can be said that the formation process of the opinion is reflected depending on the speech time. .

In this equation (2), the rapidity α ₁ of “promotion” is expressed, and the rapidity α ₂ of “opposite” is expressed. Further, in the formula (2), the mixing ratio β between “agree” and “opposite” is expressed (where 0 <β <1). In other words, the expression (2) represents the “β” proportion “β” and the “opposite” proportion (1-β). A semicolon “;” included in the argument of the expression (2) indicates that a variable included thereafter is a parameter (argument) of the expression (1) ′.

Here, the convergence condition in the first term of Equation (2) is examined. Substituting the parameters t ₁ and α ₁ of equation (2) into equation (1) ′ and moving α ₁ forward, the following equation (1) ″ is obtained. In this case, a convergence condition is required in which the integral of the expression (1) ″ is 1 (that is, the accumulation of the expression (1) ″ is 1). This convergence condition is also true parameter alpha _2.

Returning to FIG. 1, the explanation of the opinion type estimation device 2 will be continued.
The parameter optimization unit 24 applies speech data to the mixture distribution model stored in the mixture distribution model management unit 23 and estimates an optimum parameter using a numerical optimization method. Here, when the parameter optimization command is input from the opinion type adding unit 22, the parameter optimization unit 24 starts parameter optimization. Further, the parameter optimizing unit 24 can use a steepest descent method, a maximum likelihood estimation method such as BFGS (quasi-Newton method), or a Bayesian estimation method as a numerical optimization method.

For example, when the steepest descent method is used, initial values of parameters t ₁ , t ₂ , α ₁ , α ₂ , and β are set in advance in the mixed distribution model management unit 23. Then, the parameter optimization unit 24 sets the parameters t ₁ , t ₂ , α ₁ , α ₂ , β of the mixed distribution model management unit 23 until an optimal mixed distribution model is obtained (that is, until the parameters converge). Update repeatedly. When the parameter optimization unit 24 finishes updating the parameters t ₁ , t ₂ , α ₁ , α ₂ , and β, the opinion type estimation unit 25 sends the opinion type ratio to the opinion type estimation unit 25 via the mixed distribution model management unit 23. Command the estimation (opinion type estimation command).

  The opinion type estimation means 25 estimates the ratio of opinion types using the mixed distribution model in which the parameters are estimated by the parameter optimization means 24. For example, when the opinion type estimation command is input from the parameter optimization unit 24, the opinion type estimation unit 25 estimates the ratio of the opinion type.

Here, the opinion type estimation means 25 estimates the ratio of the opinion type at the utterance time and its vicinity by using the expression (2) in which the parameters t ₁ , t ₂ , α ₁ , α ₂ , and β are estimated. It becomes possible to do. That is, the opinion type estimation unit 25 can estimate the ratio “agree” and the ratio “1−β” of “opposite” included in the whole remark data from the expression (2).

  Moreover, the opinion type estimation means 25 can estimate the ratio of “agree” and the ratio of “opposite” at a certain time τ as shown in the following equation (3).

Furthermore, the opinion type estimation means 25 uses the equation (4) generalized from the equation (3) to calculate the proportion of “agree” and the proportion of “opposite” in the time interval (τ ₁ , τ ₂ ). Can be estimated. Here, the time interval (τ ₁ , τ ₂ ) represents the interval from time τ ₁ to time τ ₂ .

  Thereafter, the opinion type estimation means 25 outputs the ratio of “agree” and the ratio of “opposite” to the speech analysis means 30 as estimation results. At this time, the opinion type estimation unit 25 reads the utterance data from the mixed distribution model management unit 23 together with the estimation result, and outputs the utterance data to the utterance analysis unit 30.

  It should be noted that the opinion type estimation means 25 is set in advance as to which of the formulas (2) to (4) is used to calculate the opinion type ratio. For example, the user of the speech analysis apparatus 1 may set which of the above formulas (2) to (4) is used.

Hereinafter, the speech analysis means 30 of the speech analysis device 1 will be described.
The speech analysis unit 30 analyzes speech data using the estimation result input from the opinion type estimation unit 25. Here, the speech analysis means 30 can analyze the speech data to which “agree” and “opposite” are added using conventional natural language processing. And the speech analysis means 30 outputs the analysis result of speech data outside.

[Operation of speech analysis device]
With reference to FIG. 4, operation | movement of the speech analysis apparatus 1 is demonstrated (refer FIG. 1 suitably).
The analysis object selection means 21 selects the message data to be analyzed from all the message data (step S1).

The opinion type adding means 22 adds the opinion type to the comment data selected in step S1 (step S2).
The parameter optimization unit 24 applies the utterance data to which the opinion type is added in step S2 to the mixed distribution model, and estimates an optimum parameter using a numerical optimization method (step S3).

The opinion type estimation means 25 estimates the ratio of opinion types using the mixed distribution model in which the optimum parameters are estimated in step S3 (step S4).
The speech analysis means 30 analyzes the speech data using the estimation result of step S4 (step S5).

  Since the opinion type estimation device 2 according to the embodiment of the present invention optimizes the parameters of the mixed distribution model representing the opinion formation process and estimates the ratio of opinion types, the estimation result estimation accuracy can be improved. . As a result, the opinion type estimation device 2 can automatically add an opinion type to an opinion on the network, thereby reducing the labor of manually adding an opinion type. Furthermore, since the speech analysis apparatus 1 uses a highly accurate estimation result, a favorable analysis result of speech data can be obtained.

(Modification)
The opinion type estimation device according to the present invention is not limited to the above-described embodiment, and modifications can be made without departing from the gist thereof.
In the above-described embodiment, as shown in FIG. 3, the “agree” distribution model 90 has one peak t _1, but the present invention is not limited to this. In other words, the “agree” distribution model 90 indicates that once a statement about a certain case is raised, it converges. However, in reality, it is possible that the remarks about a certain case will come up again after it has converged.

In order to express this, as shown in FIG. 5, the “agree” distribution model 90a may be set to have two peaks t ₁₁ and t ₁₂ by superposing two Gaussian distribution models. The “opposite” distribution model 91a may be set to have two peaks t ₂₁ and t ₂₂ by superimposing two Gaussian distribution models. That is, the two-mix distribution model in FIG. 5 can be expressed by the following equation (5). In this formula (5), the same constraint condition as the above-described formula (1) ″ is required.

In the equation (5), the distribution model 90a of _{"agree", t 11} represents a peak time of Gaussian distribution model _p 11 (t), alpha ₁₁ rapidly mound of Gaussian distribution model _p 11 (t) K ₁ represents the ratio of the Gaussian distribution models p ₁₁ (t) and p ₁₂ (t) (where 0 <k ₁ <1).
Further, the distribution model 90a of _{"agree", t 12} represents a peak time of Gaussian distribution model _{p 12} (t), _α 12 represents the rapidity of the protrusion of Gaussian distribution model _p 12 (t).

Further, the distribution model 91a of the _{"opposite", t 21} represents the peak time of the Gaussian distribution model _{p 21} (t), _α 21 represents the rapidity of the protrusion of Gaussian distribution model _p 21 (t), k ₂ represents the ratio of the Gaussian distribution models p ₂₁ (t) and p ₂₂ (t) (where 0 <k ₂ <1).
Further, the distribution model 91a of the _{"opposite", t 22} represents a peak time of Gaussian distribution model _{p 22} (t), _α 22 represents the rapidity of the protrusion of Gaussian distribution model _p 22 (t). Here,

In this case, the mixed distribution model management means 23 stores and manages the two mixed distribution model of FIG. Then, the parameter optimization unit 24 estimates the parameters t ₁₁ , t ₁₂ , t ₂₁ , t ₂₂ , α ₁₁ , α ₁₂ , α ₂₁ , α ₂₂ , k ₁ , k ₂ , β in the equation (5). Furthermore, the opinion type estimation means 25 estimates the ratio of the opinion type using the equation (5) in which the optimum parameter is estimated.

In FIG. 3, the “agree” distribution model 90 and the “opposite” distribution model 91 may include three or more peaks. Further, the peak time t _{2 of} the “opposite” distribution model 91 may be earlier than the peak time t _{1 of the} “agree” distribution model 90. Further, the two-mix distribution model is not limited to the examples of FIGS. 3 and 5.

In the above-described embodiment, the opinion type is two types of “agree” or “opposite”, but is not limited to this.
For example, the opinion type may be three types of “agree”, “neutral”, or “opposite”. In this case, the mixed distribution model management means stores and manages the three mixed distribution models.
Also, the opinion types may be four or more types. In this case, the mixed distribution model management means stores and manages four or more types of mixed distribution models.
That is, if the opinion type is n type, the mixed distribution model management means stores and manages the n mixed distribution model (n is an integer of 2 or more).

  In the above-described embodiment, the distribution model for each opinion type has been described as a Gaussian distribution model, but the present invention is not limited to this. For example, the parameter optimization means can use a generalized hyperbolic distribution model gh (x) of the following equation (6) as a distribution model for each opinion type. The generalized hyperbolic distribution model can be treated as a distribution model that reflects the formation process of opinions even when there is no symmetry in the time transition between before reaching the peak and after reaching the peak.

In this equation (6), μ represents the peak time, δ represents the scale (vertical, horizontal), and λ represents the order of the third type modified Bessel function K _λ (x). Here, α and γ are parameters that determine the shape of the distribution, α affects the kurtosis, and γ affects the skewness (asymmetry).

  In the above-described embodiment, the opinion type estimation device is described as independent hardware, but the present invention is not limited to this. For example, the opinion type estimation apparatus includes hardware resources such as a CPU, a memory, and a hard disk included in a computer, an analysis target selection unit, an opinion type addition unit, a mixed distribution model management unit, a parameter optimization unit, an opinion type, It can also be realized by an opinion type estimation program for cooperative operation as an estimation means.

  The opinion type estimation device according to the present invention analyzes, for example, a politician or administrative institution whose policy is questioned, an artist whose question is an evaluation of a work, and a company whose reputation of a product is questioned on a network. Can be used for

DESCRIPTION OF SYMBOLS 1 Comment analysis apparatus 2 Opinion classification estimation apparatus 21 Analysis object selection means 22 Opinion classification addition means 23 Mixed distribution model management means (mixed distribution model storage means)
24 Parameter optimization means 25 Opinion type estimation means 30 Speech analysis means

Claims (6)

Opinion type is preset in the utterance data representing the content of the opinion, and is an opinion type estimation device that estimates statistical information about the opinion type of the utterance data,
The utterance data to which the utterance time is added while pre-stored a mixed distribution model representing the formation process of the opinion depending on the utterance time, with the mixture ratio, peak time, and rapid rise of each opinion type as parameters. Mixed distribution model storage means for storing each opinion type;
Applying parameter data to the mixture distribution model of the mixture distribution model storage means, and parameter optimization means for estimating the parameters by a numerical optimization method;
Opinion type estimation means for estimating statistical information related to the opinion type using a mixed distribution model in which parameters are estimated by the parameter optimization means;
An opinion type estimation device comprising: The mixed distribution model storage means includes a Gaussian distribution p of the following formula (1) ′ including a peak time t _n of the nth opinion type and a rapid rise α _{n of the nth} opinion type. _2. The opinion type estimation apparatus according to claim 1, wherein a mixed distribution model in which _n (t) is mixed is stored in advance (where n is a natural number).

The mixed distribution model storage means has two types of opinions, the peak time t ₁ of the first opinion type, the rapidity α ₁ of the first opinion type, and the second opinion Using the peak time t ₂ of the type and the rapidity α ₂ of the rise of the second opinion type as an argument of the equation (1) ′, the mixing ratio β of the first opinion type and the second opinion type A two-mixed distribution model p (t) of the following formula (2) including (0 <β <1) is stored in advance;

3. The opinion type estimation device according to claim 2, wherein the parameter optimization unit estimates the parameters t ₁ , t ₂ , α ₁ , α ₂ , β (provided that ';' is included later) Indicates that the variable is an argument of expression (1) ′). The mixed distribution model storage unit superimposes two Gaussian distributions p ₁₁ (t) and p ₁₂ (t) with the first opinion type, as shown in the following formula (5). A two-mixed distribution model p (t) of the distribution model and a distribution model obtained by superimposing two Gaussian distributions p ₂₁ (t) and p ₂₂ (t) in the second opinion type,

The parameter optimization means, as the parameter,
A mixing ratio β of the first opinion type and the second opinion type;
A ratio k ₁ (0 <k ₁ <1) of the Gaussian distribution models p ₁₁ (t) and p ₁₂ (t);
A peak time t _{11 in} a Gaussian distribution model p ₁₁ (t), a rapid rise α ₁₁ ,
A peak time t ₁₂ in the Gaussian distribution model p ₁₂ (t), a rapid rise α ₁₂ ,
A ratio k ₂ (0 <k ₂ <1) of the Gaussian distribution models p ₂₁ (t) and p ₂₂ (t);
The peak time t ₂₁ in the Gaussian distribution model p ₂₁ (t), the rapidity α _{21 of the} rise,
The peak time t ₂₂ in the Gaussian distribution model p ₂₂ (t), the rapidity α _{22 of the} rise,
The opinion type estimation device according to claim 2, wherein The mixed distribution model storage means includes the following formula (6) including the peak time μ, the scale δ, the order λ of the third type modified Bessel function K _λ (x), the kurtosis α, and the skewness γ. 2. The opinion type estimation device according to claim 1, wherein a mixed distribution model in which two or more generalized hyperbolic distribution models gh (x) are mixed is stored in advance. Indicates that the included variable is an argument of equation (6)).

  An opinion type estimation program for causing a computer to function as the opinion type estimation device according to claim 1.

Images