JP2016207171A - Parameter estimation method, evaluation value estimation method, apparatus, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To allow a parameter to be obtained so as to accurately estimate an evaluation value of review.SOLUTION: An estimation input part 10 receives a plurality of learning reviews about a plurality of objects to be evaluated. A parameter estimation part 26, on the basis of the plurality of learning reviews received by the estimation input part 10, estimates a model parameter according to a topic model which shows that a parameter in which an aspect overall evaluation value of each aspect with respect to the object to be evaluated and a total evaluation value of the learning reviews depend on a total overall evaluation value and an aspect evaluation value of each aspect of the learning reviews depends on the aspect overall evaluation value of each aspect with respect to objects to be evaluated and the total evaluation value of the learning reviews and which shows distribution of respective words appearing in review sentences to which aspects are assigned with respect to each aspect and a parameter in which aspects with respect to the respective aspects and the respective evaluation values are assigned and which shows distribution of respective words appearing in review sentences where an aspect evaluation value of the aspect becomes an evaluation value are calculated.SELECTED DRAWING: Figure 3

Description

  The present invention relates to a parameter estimation method, an evaluation value estimation method, an apparatus, and a program.

  Review data describing evaluations and impressions on products has been widely shared and released, and for users who are considering purchasing and using such products, overall evaluation of the entire target (for example, restaurants and hotels) Evaluation information for individual aspects (eg meals, services, places, prices) is regarded as useful. Evaluation value estimation refers to estimating an evaluation value for an entire object or individual aspects from review data described in a natural language.

  The following Non-Patent Document 1 to Non-Patent Document 3 are known as conventional techniques related to evaluation value estimation. Non-Patent Document 1 describes a potential Dirichlet distribution method using a Dirichlet distribution as a prior distribution of model parameters. The techniques described in Non-Patent Document 2 and Non-Patent Document 3 are all improvements of the method of Non-Patent Document 1. Non-Patent Document 2 describes a method called “Local-LDA” in which the latent Dirichlet allocation method is handled in sentence units, and can be used to find an aspect that can be evaluated in a sentence. Non-Patent Document 3 describes a method for predicting an evaluation value by modeling an evaluation value corresponding to each review using a supervised LDA model. These methods aim to obtain the evaluation value for each aspect directly from the review text. For example, by modeling the relationship between an evaluation value such as “meal” and “service” from a review of a hotel, and predicting the evaluation value for each individual aspect for reviews with unknown evaluation values Is performed.

David Blei, Andrew Ng and Michael Jordan, “Latent Dirichlet Allocation.”, In JMLR 2003. In JMLR. Samuel Brody and Noemie Elhadad, “An unsupervised aspect-sentiment model for online reviews.”, In NAACL 2010 D. Blei and J. McAuliffe, “Supervised topic models.”, In NIPS 2007.

  However, in the methods of Non-Patent Documents 1 to 3, the evaluation value for each aspect is considered to be independent. On the other hand, for subjects with a high overall evaluation, the evaluation value for each aspect is high, and if the overall evaluation value for a review is high, the evaluation value for each aspect in the review is also expected to be high. However, this is not taken into account in the conventional method. For this reason, it is likely that an evaluation value for each aspect that is excessively low is assigned to an object with a high overall evaluation, and prediction with low consistency of evaluation may occur.

  The present invention has been made in view of the above circumstances, and an object thereof is to provide a parameter estimation method, an evaluation value estimation method, an apparatus, and a program that can accurately estimate an evaluation value of a review.

  In order to achieve the above object, a parameter estimation method according to the present invention is a parameter estimation method in a parameter estimation apparatus including an estimation input unit and a parameter estimation unit, and the estimation input unit includes a plurality of aspects. Receiving a plurality of reviews for learning, which are a plurality of reviews for learning related to a plurality of evaluation targets, and in which an overall evaluation value for the evaluation target and a mode evaluation value for each aspect of the evaluation target are given in advance; Based on the plurality of learning reviews received by the estimation input unit by the parameter estimation unit, the overall evaluation value of each aspect with respect to the evaluation object and the overall evaluation value of the learning review with respect to the evaluation object Depending on the overall overall evaluation value, the aspect evaluation value of each aspect of the review for learning is different from each aspect of the evaluation object. Depending on the overall evaluation value and the overall evaluation value of the review for learning, for each aspect, a parameter indicating the distribution of each word appearing in the review sentence to which the aspect is assigned, and each aspect and each evaluation value , Each assigned to a review sentence according to a topic model that indicates that an aspect is assigned and a parameter representing a distribution of each word appearing in the review sentence whose aspect evaluation value is the evaluation value is calculated. A parameter indicating the distribution of the aspect, a parameter indicating the distribution of each word appearing in the review sentence to which the aspect is assigned, and an aspect for each aspect and each evaluation value are assigned, and A parameter representing the distribution of each word appearing in the review sentence in which the aspect evaluation value of the aspect becomes the evaluation value, and all of the plurality of evaluation objects The average and standard deviation of the overall evaluation values, the standard deviation of the overall evaluation values for the plurality of evaluation targets for each aspect, the standard deviation of the overall evaluation values of the plurality of learning reviews, and the plurality for each aspect Estimating a model parameter including a standard deviation of an aspect evaluation value of the review for learning.

  The parameter estimation device according to the present invention is a plurality of reviews for learning related to a plurality of evaluation objects including a plurality of aspects, and an overall evaluation value for the evaluation object and an aspect evaluation value for each aspect of the evaluation object are preliminarily obtained. An estimation input unit that receives a plurality of given learning reviews, and the overall evaluation value of each aspect and the review for learning based on the plurality of learning reviews received by the estimation input unit And the overall evaluation value of each aspect of the review for learning is the overall evaluation value of each aspect of the review for evaluation and the overall evaluation of the review for learning Depending on the value, for each aspect, a parameter representing the distribution of each word appearing in the review sentence to which the aspect is assigned, and each aspect For each evaluation value, a review is performed according to a topic model indicating that an aspect is assigned and a parameter representing a distribution of each word appearing in a review sentence in which the aspect evaluation value of the aspect becomes the evaluation value is calculated. Parameters indicating the distribution of each aspect assigned to the sentence, parameters indicating the distribution of each word appearing in the review sentence to which the aspect is assigned, and the aspects of each aspect and each evaluation value are assigned. And a parameter representing the distribution of each word appearing in the review sentence in which the aspect evaluation value of the aspect becomes the evaluation value, the average and standard deviation of the overall comprehensive evaluation values for the plurality of evaluation objects, and each aspect Standard deviation of the overall evaluation value of the aspect for the plurality of evaluation objects, and standard deviation of the overall evaluation value of the plurality of learning reviews, Wherein and the plurality of configured to include a parameter estimation unit for estimating the model parameters, the containing and the standard deviation of the learning review aspects evaluation value for the.

  The parameter estimation method according to the present invention further includes a step in which an initialization unit initializes the model parameter, and the step in which the parameter estimation unit estimates the model parameter includes the model parameter and the plurality of learning parameters. Based on the overall evaluation value of each review and the aspect evaluation value of each aspect, for each of the plurality of evaluation objects, the average and standard deviation of the overall evaluation value of the evaluation object, and the evaluation object For each, the average and standard deviation of the overall evaluation values of the aspects of the evaluation object for each aspect, the parameter representing the distribution of each aspect in the learning review for each of the plurality of learning reviews, and the plurality of learnings For each review sentence for review, a parameter indicating the distribution of each aspect assigned to the review sentence; Estimating the variational parameters including, and estimated on the basis of the variational parameter to estimate the model parameters, it is possible to make repeated until iteration termination condition is satisfied predetermined.

  The evaluation value estimation method according to the present invention is an evaluation value estimation method in an evaluation value estimation device including an input unit and an evaluation value estimation unit, wherein the input unit receives a review related to an evaluation object, and the evaluation value estimation Based on the review received by the input unit and the model parameter estimated by the parameter estimation method of the present invention, the overall evaluation value for the evaluation target of the review, and the evaluation target of the review Estimating an aspect evaluation value for each aspect.

  The program according to the present invention is a program for causing a computer to function as each unit of the parameter estimation device of the present invention.

  As described above, according to the present invention, based on a plurality of learning reviews, the overall aspect evaluation value of each aspect for the evaluation object and the overall evaluation value of the learning review depend on the overall evaluation value for the evaluation object. However, the aspect evaluation value of each aspect of the learning review depends on the overall evaluation value of each aspect and the overall evaluation value of the review for the evaluation target, and appears in the review sentence to which the aspect is assigned for each aspect. A parameter representing the distribution of each word and a parameter representing the distribution of each word appearing in the review sentence in which the aspect is assigned to each aspect and each evaluation value and the aspect evaluation value of the aspect is the evaluation value. By estimating the model parameters according to the topic model representing what is calculated, the parameters can be obtained to accurately estimate the review evaluation value. , An effect that can be obtained.

It is the image figure which expressed the production | generation process of the review by the topic model based on hierarchical prior distribution by the pseudo code. It is a pseudo code representing E-step and initialization in the model parameter estimation processing according to the embodiment of the present invention. It is the schematic which shows the structure of the parameter estimation apparatus which concerns on embodiment of this invention. It is the schematic which shows the structure of the evaluation value estimation apparatus which concerns on embodiment of this invention. It is a flowchart which shows the content of the parameter estimation process routine in the parameter estimation apparatus which concerns on embodiment of this invention. It is a flowchart which shows the content of the E-step process routine in the parameter estimation apparatus which concerns on embodiment of this invention. It is a flowchart which shows the content of the M-step process routine in the parameter estimation apparatus which concerns on embodiment of this invention. It is a flowchart which shows the content of the evaluation value estimation process routine in the parameter estimation apparatus which concerns on embodiment of this invention. It is a figure which shows the experimental result of embodiment of this invention. It is a figure which shows the experimental result of embodiment of this invention.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. Below, the outline | summary and principle of embodiment of this invention are demonstrated first.

<Outline of Embodiment of the Present Invention>
The embodiment of the present invention is for accurately estimating an evaluation value for the entire evaluation object or individual aspects of the evaluation object in the review data relating to the evaluation object. Examples of evaluation targets include restaurants and hotels. The evaluation target includes a plurality of aspects, and examples of the aspect include meals, services, places, prices, and the like. In the embodiment of the present invention, the overall comprehensive evaluation value for the entire evaluation target, the overall aspect evaluation value for the individual aspect of the evaluation target, and the evaluation value of each review are expected to be related to each other. The evaluation value is estimated by using a topic model based on Latent Dirichlet Allocation (LDA).

  In the embodiment of the present invention, each evaluation value for the evaluation target is considered as a random variable, and the following conditions (1) to (4) are set.

  The overall comprehensive evaluation value for the evaluation object represents an evaluation value for the entire evaluation object obtained from a plurality of reviews related to the evaluation object. The aspect comprehensive evaluation value of each aspect with respect to the evaluation object represents an evaluation value for the aspect of the evaluation object obtained from a plurality of reviews regarding the evaluation object. The overall evaluation value of a review represents an evaluation value for an evaluation object in one review. The aspect evaluation value of each aspect of the review represents an evaluation value for each aspect to be evaluated in one review.

(1) The aspect comprehensive evaluation value for each aspect to be evaluated follows a probability distribution that averages the overall evaluation value of the evaluation object.
(2) The overall evaluation value of each review follows a probability distribution that averages the overall comprehensive evaluation value of the evaluation target.
(3) The appearance evaluation value for each aspect of each review follows a probability distribution that averages the weighted sum of the overall evaluation value for each aspect to be evaluated and the overall evaluation value of the review.
(4) Each word in the review sentence follows the aspect evaluation value for each aspect of each review and the conditional probability distribution of the latent topic.

  In the embodiment of the present invention, each evaluation value conforms to the above conditions (1) to (4), and each evaluation value is included in a variable of a topic model based on LDA (Latent Dirichlet Allocation). The hierarchy of evaluation values can be considered.

  By considering the hierarchy of review evaluation values, the overall evaluation value of each review and the level of the aspect evaluation value for each aspect of the review change depending on the level of the overall comprehensive evaluation value and the overall aspect evaluation value. It can be expected that an object with a high overall evaluation value and an overall aspect evaluation value will be able to obtain a model that is likely to have an average evaluation value that is high on average and an aspect evaluation value for each aspect. As a result, it can be expected that prediction accuracy of high review evaluation values can be obtained.

<Principle>
Next, the principle of the embodiment of the present invention will be described.

[1. Formulation]
Assume that there are N evaluation objects (for example, restaurants and hotels) as review data. Each evaluation target

N _i reviews

Is given. Individual reviews

Sentence

And each sentence

Words

Is included. Furthermore, there are K evaluation aspects for evaluation.

Suppose there is. Further, V is the number of different words (vocabulary) used in the review. Each evaluation object includes an overall evaluation value R ^all _i and an evaluation value vector of K-dimensional aspects that summarizes the aspect comprehensive evaluation values for each aspect.

Shall be granted. Each evaluation value is assumed to be a non-negative real number. Each review includes an overall evaluation value r ^all _d and a K-dimensional aspect evaluation value vector.

Shall be granted. The notation “R” represents an evaluation value obtained by integrating a plurality of reviews, and the notation “r” represents an evaluation value in individual reviews.

In the present embodiment, a model is learned using a learning review data set with a known evaluation value as learning data, and the evaluation value is estimated for review data with an unknown evaluation value. Each review in the training for review data is a whole evaluation value r ^all _d and the appearance evaluation value for each aspect ^ r _i has already been granted. In the present embodiment, the evaluation value obtained for the review data whose evaluation value is unknown is:

1. Overall comprehensive evaluation value R ^all _i for each evaluation object in the review data whose evaluation value is unknown, and aspect comprehensive evaluation value ^ R _i for each aspect
2. The overall evaluation value r ^all _d for each review in the review data whose evaluation value is unknown, and the aspect evaluation value r _dk for each aspect.

[2. model]
In the embodiment of the present invention, a “topic model based on a hierarchical prior distribution” that generates a review described in a natural language in review data according to an evaluation value is used.

In the topic model based on the hierarchical prior distribution in the present embodiment, it is assumed that the aspect evaluation values for each aspect in the same evaluation object and the same review are not independent from each other. For each evaluation object E _i , the overall overall evaluation value R ^all _i of the evaluation object E _i is generated according to a normal distribution of mean μ and standard deviation Σ, as shown in the following equation (1). To do.

Then, the total evaluation value ^{r all} _d for each aspect each aspect review and overall evaluation value _{R ik} evaluated _{E i,} as shown in the following equation (2) to (3), each evaluated _{E i} It is assumed to be generated according to a normal distribution of standard deviations σ ₁ and σ _{2 with} the overall overall evaluation value R ^all _i as an average.

As shown in the following formula (4), the aspect evaluation value r _dk for each aspect in each review is the overall evaluation value r ^all _d of the review and the aspect total for each aspect in the evaluation object E _i evaluated by the review. It is assumed that the value is generated according to a normal distribution with a standard deviation σ ₃ that averages values obtained by linearly combining the evaluation values R _ik .

Here, λ is a parameter for adjusting the degree of contribution between the overall evaluation value r ^all _d of the review d and the overall aspect evaluation value R _ik for the k-th aspect of the evaluation object E _i , and is a predetermined arbitrary constant (For example, 0.5) may be used.

  FIG. 1 shows a review generation process based on a topic model based on a hierarchical prior distribution in the present embodiment.

Assume that each sentence of the review is for the same aspect, and that the aspect is considered as a latent topic in the topic model, and the review is generated based on the generation process as shown in FIG. That is, each word of the review has an aspect z _s assigned to a review sentence including the word and an aspect evaluation value for each aspect of the review.

Conditional probability distribution with

And is generated based on

Is a distribution of words appearing in the aspect z _s as shown in the following equation (5):

And the aspect evaluation value for each aspect of the review in the aspect z _s

Distribution of words that appear when

Expressed as a product of

  In addition,

Can take any positive real value as shown in the above formula (5), but the aspect evaluation value for each aspect of the review in the review data is usually a discrete value.

Uses a value rounded to a discrete value such as an integer.

In the learning of the model in the present embodiment, the model parameter す る = that maximizes the posterior probability p (w, r ^all _d , r _dk | Υ) of the word and the evaluation value in the review data for learning whose evaluation value is known. Estimate {μ, Σ, σ ₁ , σ ₂ , σ ₃ , λ, α, φ, ψ}, and use the estimated model parameter Υ to evaluate the review for an unknown evaluation value value

Is estimated.

Note that μ of the model parameters Upsilon, represents the average of the entire total evaluation value R ^all _i of a plurality of evaluation target E _i, sigma represents a standard deviation of the entire total evaluation value R ^all _i of a plurality of evaluation target E _i , Σ ₁ represents the standard deviation of the overall evaluation value R _ik for each aspect, σ ₂ represents the standard deviation of the overall evaluation value r ^all _d for each review, and σ ₃ represents the aspect evaluation value r for each aspect of each review. _It represents the standard deviation of _dk . In addition, λ represents a parameter of the degree of contribution of the overall evaluation value of the review and the overall evaluation value of the aspect for each aspect in the aspect evaluation value for each aspect of each review. In addition, α represents a parameter (K dimension) of the Dirichlet distribution, φ represents an appearance distribution of the word w (expressed for each _k of the aspect Ak, and is a word number dimension), and ψ represents an appearance distribution of the word w. (represented in every aspect _{a k} of k and aspects _{a k} of k for each aspect evaluation value _{r dk,} is the number-dimensional word.) represents the.

  Based on the generation process of FIG. 1 above, the posterior probabilities of words and evaluation values obtained for the review data for learning are as shown in the following formula (6).

[3. Model derivation]
Since it is difficult to efficiently calculate the posterior probability distribution of the hidden variables necessary for maximizing the posterior probability shown in the above formula (6), in this embodiment, Non-Patent Document 4 (MICHAEL I. JORDAN , ZOUBIN GHAHRAMANI, TOMMI S. JAAKKOLA, LAWRENCE K. SAUL, “An Introduction to Variational Methods for Graphical Models”, Machine Learning, Volume 37, Issue 2, p.183-233, 1999). Is used. In the present embodiment, the posterior probability distribution of the hidden variable is approximated using the variation distribution shown in the following formula (7) by the variation method.

Here, the allocation z _s of the aspect to each review sentence is a K-dimensional multinomial distribution.

The parameter θ _d of the multinomial distribution follows the K-dimensional Dirichlet distribution Dir (γ _d ), the overall overall evaluation value R ^all _i of the evaluation object E _i follows the normal distribution of the mean η _i and the standard deviation Δ _i , and the evaluation object E _The aspect comprehensive evaluation value R _ik for the aspect of _i follows a normal distribution with an average β _ik and a standard deviation δ _ik .
The log likelihood of the posterior probability shown in the above equation (6) by the Jensen inequality

Is obtained as shown in the following equation (8).

Here, H (q) = − E [log q] is the entropy of the variational distribution q.
Based on equation (8) above, variational parameters

The process of maximizing the lower bound L for (E-step) and the process of maximizing the lower bound for the model parameters Υ = {μ, Σ, σ ₁ , σ ₂ , σ ₃ , λ, α, φ, ψ} (M The posterior probability is maximized by a method called variation EM (Expectation-Maximization). Note that eta variational parameter [Phi, for each of a plurality of evaluation target E _i, consists average eta _i overall total evaluation value R ^all _i of the evaluated E _i, delta is a plurality of evaluation target E _i It consists of the standard deviation Δ _i of the overall overall evaluation value R ^all _i of the evaluation object E _i for each. Β is composed of an average β _ik of the overall evaluation values R _ik of the combination for each of the combinations of the plurality of evaluation objects E _i and the plurality of aspects k, and δ is the plurality of evaluation objects E _i and the plurality of evaluation objects E _i For each combination of aspects k, it consists of the standard deviation δ _ik of the aspect total evaluation value R _ik of the combination. γ consists of a Dirichlet distribution parameter γ _dk representing the distribution of each aspect k in the learning review d for each of the plurality of learning reviews d,

Is a parameter of a multinomial distribution representing the distribution of each aspect k assigned to the review sentence s for a plurality of review sentences s of a plurality of review for learning.

Consists of.

[3.1 E-step]

  First, in E-step,

  Optimal variational parameter satisfying

Ask for. FIG. 2 shows the calculation of E-step of the variational EM algorithm (i is the ID of the evaluation object, d is the ID of the review, s is the ID of the sentence, and k is the ID of the aspect) and pseudo code for initialization. Show.

Η _i , Δ _i , β _ik , and δ _ik of the variation parameter Φ are updated based on the following equations (9) to (12).

Next, the variation parameter Φ γ _dk ,

Is updated based on the following equations (13) to (14) (k = 1,..., K).

Here, Ψ (·) represents the first derivative of the log Γ function. The derivative of η _i and β _ik , γ and

Since the derivative of is dependent on each other, the calculation is performed according to the iterative procedure as shown in FIG.

[3.2 M-step]
Subsequently, in M-step, the model parameter Υ = {μ, Σ, σ ₁ , σ ₂ , σ ₃ , λ, α, φ, ψ} is updated to maximize the lower bound of the log likelihood.

  First, φ and ψ are updated based on the following formulas (15) to (16).

However, 1 (·) is a function that returns 1 when the condition is satisfied and returns 0 when the condition is not satisfied. Φ _k ^wn represents the appearance probability of the n-th word w with respect to the k-th aspect A _k , and ψ _km ^wn represents the aspect evaluation value of the review d and is n with respect to the k-th aspect A _k . This represents the appearance probability of the th word w.
Next, α is updated based on the following equation (17).

As shown in the following formula (18), since the derivative with respect to k depends on α _k ′ (k ≠ k ′), the maximum value is obtained by the iterative process by the Newton-Raphson method described in Non-Patent Document 1 above. You need a way to ask. The Hessian matrix for obtaining the maximum value is calculated as shown in the following equation (18).

Next, μ, Σ, σ ₁ , σ ₂ , and σ ₃ are updated based on the following equations (19) to (23).

[4. prediction]
After the model parameter Υ is estimated by the above-described method, the evaluation value Θ is predicted for review data whose evaluation value is unknown according to the following equation (24).

Θ represents an evaluation value, and Υ represents a model parameter. Although p (w | Θ, Υ) is obtained by integrating and eliminating the parameters z _s , α, and θ, it is difficult to calculate directly. Therefore, in the present embodiment, z _s is first assigned to each review sentence.
Note that maximization of p (w | Θ, Υ) is replaced with maximization of p (w | Θ, Υ, z), and maximization is performed as shown in the following equation (25).

_{p (w∈E i | Θ i,} Υ, z) by theta formula below are obtained as the derivative becomes zero for each element of the (26) to (29), the entire evaluation target Ei of the review The overall evaluation value R ^all _i , the aspect overall evaluation value ＲR _i for each aspect, the overall evaluation value r ^all _d for the review, and the aspect evaluation value r _dk for each aspect are obtained.

Here, T (r _dk ) is calculated by the following equation (30).

  Since the differentiation for each element of Θ is not independent, in this embodiment, all elements of Θ are set to 0, and are updated repeatedly until convergence is achieved by the update expressions shown in the above equations (26) to (29).

  Next, the parameter estimation device for estimating the model parameter relating to the review of the evaluation target and the evaluation value estimation device for estimating the evaluation value of the review of the evaluation value estimation target using the model parameter estimated by the parameter estimation device are A case where the invention is applied will be described as an example.

<System configuration of parameter estimation device>
As shown in FIG. 3, parameter estimation apparatus 100 according to the present embodiment is configured by a computer including a CPU, a RAM, and a ROM that stores a program for executing a learning processing routine described later, and functions. Specifically, it is configured as follows. The parameter estimation apparatus according to the present embodiment includes an estimation input unit 10, an estimation calculation unit 20, and a model parameter storage unit 30.

  The estimation input unit 10 receives a plurality of learning reviews related to a plurality of evaluation targets. In the review for learning, the overall evaluation value for the evaluation object and the aspect evaluation value for each aspect of the evaluation object are given in advance.

  The estimation calculation unit 20 includes a review database 22, an initialization unit 24, and a parameter estimation unit 26.

  The review database 22 stores a plurality of learning reviews received by the estimation input unit 10.

The initialization unit 24 sets initial values for each parameter. Specifically, the initialization unit 24 sets predetermined initial values for {μ, Σ, σ ₁ , σ ₂ , σ ₃ , λ, α, φ, ψ} of the model parameter 、, and the model Initialize parameters. The initialization unit 24 sets an initial value for the variation parameter Φ according to the following equation.

The parameter estimation unit 26 estimates the model parameter Υ according to the above-described topic model based on the plurality of learning reviews stored in the review database 22. The model parameter Υ includes a parameter α representing the distribution of each aspect assigned to the review sentence, a parameter φ representing the distribution of each word appearing in the review sentence to which the aspect is assigned, and each aspect and each A parameter ψ representing the distribution of each word appearing in a review sentence in which an aspect is assigned to the evaluation value and the aspect evaluation value of the aspect becomes the evaluation value, and an average μ of the overall comprehensive evaluation values for a plurality of evaluation objects Standard deviation Σ, standard deviation σ ₁ of overall evaluation values for multiple evaluation targets for each aspect, standard deviation σ ₂ of overall evaluation values for multiple learning reviews, and multiple learning reviews for each aspect The standard deviation σ ₃ of the aspect evaluation value is included.

  In the present embodiment, as shown in FIG. 1 above, in the topic model, the aspect comprehensive evaluation value of each aspect for the evaluation object and the overall evaluation value of the review for learning depend on the overall evaluation value for the evaluation object, It is shown that the aspect evaluation value of each aspect of the review for learning depends on the overall evaluation value of each aspect for the evaluation object and the overall evaluation value of the review for learning. Further, as shown in the above equation (5), it is expressed that the parameter φ and the parameter ψ are calculated in the topic model from the dependency relationship.

  The parameter estimation unit 26 uses the variational EM algorithm to estimate the variation parameter Φ in E-step and the model parameter に お い て in M-step. The parameter estimation unit 26 repeats estimating the model parameter Υ based on the variation parameter Φ estimated in E-step until a predetermined iteration end condition is satisfied.

First, the parameter estimator 26 sets a plurality of evaluation targets E based on the model parameter Υ that has been initialized or previously estimated, and the overall evaluation values and the aspect evaluation values of each aspect of the plurality of learning reviews. for each _i, the average eta _i and standard deviation delta _i overall total evaluation value of the evaluation object E _i, for each of a plurality of evaluation target E _i, aspects comprehensive evaluation value for the evaluation object E _i for each aspect k Reviews for the average β _ik and standard deviation δ _ik , the parameter γ _dk representing the distribution of each aspect k in the learning review d for each of the plurality of learning reviews d, and the review sentences s of the plurality of learning reviews d Parameter indicating the distribution of each aspect k assigned to sentence s

A variational parameter Φ including

Specifically, the parameter estimation unit 26 updates the variation parameter Δ ² _i for each evaluation target IDi in E-step according to the above equation (10). Next, the parameter estimation unit 26 updates the variation parameter δ ² _ik according to the above equation (12) for each evaluation target IDi and each aspect IDk.

Next, the parameter estimation unit 26, based on the above formulas (13) and (14), the parameter γ _dk representing the distribution of each aspect k in the learning review d for each of the plurality of learning reviews d, and a plurality of parameters γ _dk Representing the distribution of each aspect k assigned to the review sentence s for each review sentence s of the review for learning

And update. The parameter estimation unit 26 repeats the update based on the above formulas (13) and (14) until the update process converges.

And the parameter estimation part 26 updates (eta) _i about each evaluation object IDi based on the said Formula (9). Moreover, the parameter estimation part 26 updates (beta) _ik about each evaluation object IDi and each aspect IDk based on the said Formula (11). The parameter estimation unit 26 repeats the update based on the above formulas (9) and (11) until the update process converges.

  Next, the parameter estimation unit 26 performs parameter estimation in E-step in M-step.

Based on the above, the parameter φ and the parameter ψ are estimated according to the above equations (15) and (16).

  Also, the parameter estimation unit 26 estimates the parameter α representing the distribution of each aspect assigned to the review sentence according to the above equation (17) based on the parameter γ estimated in the E-step in the M-step.

Further, the parameter estimation unit 26 updates {μ, Σ, σ ₁ , σ ₂ , σ ₃ } according to the above formulas (19) to (23) based on the variation parameter Φ estimated in E-step. To do.

  Then, the parameter estimation unit 26 stores the estimated model parameter in the model parameter storage unit 30.

  The model parameter storage unit 30 stores the model parameter 推定 estimated by the parameter estimation unit 26.

<Evaluation value estimation device>
As shown in FIG. 4, evaluation value estimation apparatus 200 according to the present embodiment is configured by a computer that includes a CPU, a RAM, and a ROM that stores a program for executing an evaluation value estimation processing routine described later. Functionally, it is configured as follows.

  The evaluation value estimation apparatus according to the present embodiment includes an input unit 40, a model parameter storage unit 50, a calculation unit 60, and an output unit 70.

The input unit 40 receives a review regarding the evaluation target E _i .

  The model parameter storage unit 50 stores the same model parameters as the model parameters stored in the model parameter storage unit 30 of the parameter estimation apparatus 100.

  The calculation unit 60 includes an evaluation value estimation unit 62.

The evaluation value estimation unit 62, based on the review received by the input unit 40 and the model parameters stored in the model parameter storage unit 50, according to the above formulas (26) to (29), the overall evaluation value r ^all for the review. _d, aspects evaluation value _{r dk} for each aspect of the evaluation of the review, evaluation to estimate the appearance overall evaluation value _{R ik} for each aspect k of the entire comprehensive evaluation value ^{R all} _i, and evaluated _{E i} with respect to the target _{E i.} The evaluation value estimation unit 62 repeatedly updates until convergence is achieved by the update formulas shown in the above formulas (26) to (29).

  The output unit 70 outputs the evaluation value estimated by the evaluation value estimation unit 62.

<Operation of parameter estimation device>
Next, the operation of the parameter estimation apparatus according to this embodiment will be described. First, when a plurality of learning reviews are input to the parameter estimation device, the estimation input unit 10 receives a plurality of learning reviews and stores them in the review database 22. Then, in the parameter estimation device, a parameter estimation processing routine shown in FIG. 5 is executed.

  First, in step S100, the initialization unit 24 sets initial values for each parameter.

  In step S102, the parameter estimation unit 26, the plurality of learning reviews stored in the review database 22, the variation parameter Φ initialized in step S100, or the variable estimated last time in the previous step S102. Based on the minute parameter Φ and the model parameter 初期 initialized in step S100 or the model parameter Υ estimated last time in step S104 described later, the variation parameter Φ is estimated according to the topic model described above. Step S100 is realized by an E-step processing routine shown in FIG.

<E-step processing routine>
In step S200, the parameter estimation unit 26 determines, for each evaluation target IDi, the model parameter {Σ, σ ₁ , σ ₂ } initialized in step S100 or the model parameter {Σ, Based on σ ₁ , σ ₂ }, the variation parameter Δ ² _i is updated according to the above equation (10).

In step S202, the parameter estimation unit 26 determines, for each evaluation object IDi and each aspect IDk, the model parameters {σ ₁ , σ ₃ } initialized in step S100 or the model parameters { Based on σ ₁ , σ ₃ }, the variational parameter δ ² _ik is updated according to the above equation (12).

  In step S204, for each review sentence IDs and each aspect IDk, the parameter estimation unit 26 initializes the model parameter {φ, ψ} in step S100 or the model parameter {φ, last updated in step S104 described later. ψ} and the variation parameter {γ} initialized in step S100 or the variation parameter {γ} updated last time in step S208 described later, according to the above equation (14),

Update.

  In step S206, the parameter estimation unit 26 is updated in step S204.

On the basis of the,

Is normalized.

  In step S208, the parameter estimation unit 26 obtains each review IDd and each aspect IDk in step S206.

Then, γ _dk is updated according to the above equation (13) based on the model parameter {α} initialized in step S100 or the model parameter {α} updated last time in step S104 described later.

  In step S210, the parameter estimation unit 26 determines whether or not the processing in steps S204 to S208 has converged. If the process has converged, the process proceeds to step S212. On the other hand, if the process has not converged, the process returns to step S204.

In step S212, the parameter estimation unit 26 for each evaluation target IDi, the model parameter {μ, Σ, σ ₁ , σ ₂ } initialized in step S100 or the model parameter { μ, Σ, σ ₁ , σ ₂ }, the variation parameter {β} initialized in step S100 or the variation parameter {β} updated last time in step S214 described later, and the review database 22 are stored. The variation parameter η _i is updated based on the above formula (9) based on the overall evaluation value r _d ^all of the learning review d.

In step S214, the parameter estimation unit 26 for each evaluation object IDi and each aspect IDk, the variation parameter {η} updated in step S212 and the model parameters {σ ₁ , σ initialized in step S100. ₃ } or the model parameters {σ ₁ , σ ₃ } last updated in step S104 described later, the overall evaluation value r _d ^all of the learning review d stored in the review database 22, and the aspect of the learning review d Based on the aspect evaluation value r _dk , the variation parameter β _ik is updated according to the above equation (11).

  In step S216, the parameter estimation unit 26 determines whether or not the processing in steps S212 to S214 has converged. If the process has converged, the process proceeds to step S218. On the other hand, if the process has not converged, the process returns to step S212.

  In step S218, the variation parameter updated in steps S200 to S208 and S212 to S214 is output as a result, and the E-step processing routine is terminated.

  Next, returning to the parameter estimation processing routine, in step S104, the model parameter Υ is estimated based on the variation parameter Φ output in step S102 and the plurality of learning reviews d stored in the review database 22. . Step S202 is realized by the M-step processing routine shown in FIG.

<M-step processing routine>
In step S300, the parameter estimation unit 26 updates the variation parameter updated in step S102.

Based on the plurality of learning reviews d stored in the review database 22, the model parameters {φ, ψ} are updated according to the above formulas (15) and (16).

  In step S302, the parameter estimation unit 26, based on the variation parameter {γ} updated in step S102 and the plurality of learning reviews d stored in the review database 22, the above equations (17) and ( According to 18), the model parameter {α} is estimated.

In step S304, the parameter estimation unit 26 calculates the above formulas (19) to (23) based on the variation parameter Φ output in step S102 and the plurality of learning reviews d stored in the review database 22. , {Μ, Σ, σ ₁ , σ ₂ , σ ₃ } are updated.

  In step S306, the model parameter 更新 updated in steps S300 to S304 is output as a result, and the M-step processing routine is terminated.

  Next, returning to the parameter estimation processing routine, in step S106, the parameter estimation unit 26 determines whether or not the processing in steps S102 to S104 has converged. If the processing from step S102 to step S104 has converged, the process proceeds to step S108. On the other hand, if the processes in steps S102 to S104 have not converged, the process returns to step S102.

  In step S108, the parameter estimation unit 26 stores the model parameter Υ obtained by the processing in steps S102 to S104 in the model parameter storage unit 30, and ends the parameter estimation processing routine.

<Operation of evaluation value estimation device>
Next, the operation of the acoustic signal analyzer according to this embodiment will be described. First, when the model parameter Υ stored in the model parameter storage unit 30 of the parameter estimation device 100 is input to the evaluation value estimation device 200, it is stored in the model parameter storage unit 50. When the review of the evaluation value estimation target is input to the evaluation value estimation device 200, the evaluation value estimation processing routine shown in FIG. 8 is executed.

  In step S400, the input unit 40 receives a review regarding the evaluation target Ei.

  In step S402, the model parameter 記憶 stored in the model parameter storage unit 50 is read.

In step S404, the evaluation value estimation unit 62 evaluates the review according to the above formulas (26) to (29) based on the review received in step S400 and the model parameter 読 み 込 ま read in step S402. overall evaluation value ^{r all} _d, appearance evaluation value for evaluating each aspect of the subject of the review _{r dk} for, the entire overall evaluation value ^{R all} _i for evaluation _{E i,} and evaluation aspects comprehensive evaluation value for each aspect k of the target _{E i} R Estimate _ik .

  In step S406, the evaluation value estimation unit 62 determines whether or not the process of step S404 has converged. If the process of step S404 has converged, the process proceeds to step S408. On the other hand, if the process of step S404 has not converged, the process returns to step S404.

  In step S408, the evaluation value estimated in step S404 is output as a result, and the evaluation value estimation processing routine is terminated.

<Example>
[data]
The embodiment of the present invention is applied to review data in which an evaluation object is a hotel. In this review data, in addition to the overall evaluation value for each hotel, one to five aspect evaluation values are assigned to seven aspects such as service, cleanliness, and location.

[Experimental settings]
(Identification of aspects)
About 300 keywords were selected by the bootstrap method, and the selected keywords were used as prior knowledge for learning the LDA model. It can be considered that the topic obtained by learning the topic model can describe the contents of each aspect.

  Then, using KL divergence, the distance between distributions of each aspect learned by the topic model and the true aspect was measured.

  FIG. 9 shows the experimental results of KL divergence between the learned aspect and the true aspect. The results obtained from the present embodiment are compared with the results obtained from LDA without prior knowledge (Non-Patent Document 1), sLDA (Non-Patent Document 3), and Local-LDA (Non-Patent Document 2). As a result, the model according to the present embodiment is closer to the true aspect than LDA and sLDA, and the result is equivalent to Local-LDA.

(Evaluation value estimation)
The performance evaluation of evaluation value estimation according to the embodiment of the present invention was performed. In addition, keywords were used as prior knowledge so that the aspects defined in the review data correspond to the aspects learned automatically. As the evaluation scale, three indices of mean square error (MSE) between the estimated evaluation value and the actual evaluation value, Mean Average Precision (MAPN), and NDCKN often used for evaluation of information retrieval were used.

  FIG. 10 shows the estimation accuracy of the evaluation value obtained from the experiment. As shown in FIG. 10, as a result of comparing the embodiment of the present invention with LDA, sLDA, Bootstrap, and LocalLDA, the embodiment of the present invention can be used for both the evaluation value for the entire object and the evaluation value for each aspect. The estimation accuracy is better than other methods.

  As described above, according to the parameter estimation device according to the embodiment of the present invention, based on a plurality of learning reviews, the overall evaluation value of each aspect for the evaluation target and the overall evaluation value of the learning review are calculated. Depending on the overall evaluation value for the evaluation target, the aspect evaluation value of each aspect of the review for learning depends on the overall evaluation value of each aspect for the evaluation object and the overall evaluation value of the learning review. Parameters that represent the distribution of each word that appears in a review sentence to which an aspect is assigned, and appearances in a review sentence in which an aspect is assigned to each aspect and each evaluation value, and the aspect evaluation value of the aspect is an evaluation value By estimating the model parameters according to the topic model representing that the parameters representing the distribution of each word are calculated, the evaluation value of the review can be accurately estimated. It is possible to obtain a parameter for.

In addition, according to the evaluation value estimation device according to the embodiment of the present invention, the evaluation target for review based on the review related to the evaluation target and the model parameter estimated by the parameter estimation device according to the embodiment of the present invention. By estimating the overall evaluation value for and the aspect evaluation value for each aspect of the review evaluation target, the evaluation value for the review evaluation target can be accurately estimated.

  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, in the above embodiment, the case where the parameter estimation device 100 and the evaluation value estimation device 200 are configured as separate devices has been described as an example. However, the parameter estimation device 100 and the evaluation value estimation device 200 are configured as one device. May be.

  Moreover, although the parameter estimation apparatus 100 and the evaluation value estimation apparatus 200 described above have a computer system inside, if the “computer system” uses a WWW system, the homepage provision environment (or Display environment).

  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.

DESCRIPTION OF SYMBOLS 10 Estimation input part 20 Estimation operation part 22 Review database 24 Initialization part 26 Parameter estimation part 30 Model parameter storage part 40 Input part 50 Model parameter storage part 60 Operation part 62 Evaluation value estimation part 70 Output part 100 Parameter estimation apparatus 200 Evaluation value estimation device

Claims (5)

A parameter estimation method in a parameter estimation device including an estimation input unit and a parameter estimation unit,
The estimation input unit is a plurality of review reviews for a plurality of evaluation objects including a plurality of aspects, and an overall evaluation value for the evaluation object and an aspect evaluation value for each aspect of the evaluation object are given in advance. Accepting multiple learning reviews, and
Based on the plurality of learning reviews received by the estimation input unit by the parameter estimation unit, the overall evaluation value of each aspect with respect to the evaluation target and the overall evaluation value of the learning review are the evaluation target. Depending on the overall evaluation value for each aspect, the aspect evaluation value of each aspect of the review for learning depends on the overall evaluation value of each aspect for the evaluation object and the overall evaluation value of the review for learning. , A parameter indicating the distribution of each word appearing in the review sentence to which the aspect is assigned, and a review in which an aspect is assigned to each aspect and each evaluation value, and the aspect evaluation value of the aspect is the evaluation value. According to the topic model representing that the parameters representing the distribution of each word appearing in the sentence are calculated,
Parameters indicating the distribution of each aspect assigned to the review sentence, parameters indicating the distribution of each word appearing in the review sentence to which the aspect is assigned, and the aspects of each aspect and each evaluation value are assigned. A parameter representing the distribution of each word appearing in a review sentence in which the aspect evaluation value of the aspect is the evaluation value, an average and a standard deviation of the overall comprehensive evaluation values for the plurality of evaluation objects, and each aspect A standard deviation of overall evaluation values of the plurality of evaluation targets for the plurality of evaluation targets, a standard deviation of overall evaluation values of the plurality of learning reviews, and a standard deviation of the aspect evaluation values of the plurality of learning reviews for each aspect, and Estimating model parameters including:
A parameter estimation method including: The initialization unit further includes a step of initializing the model parameter, and the step of the parameter estimation unit estimating the model parameter includes the model parameter, the overall evaluation value of each of the plurality of learning reviews, and Based on the aspect evaluation value of each aspect, the average and standard deviation of the overall overall evaluation value of the evaluation object for each of the plurality of evaluation objects, and the evaluation for each aspect for each of the plurality of evaluation objects The average and standard deviation of the overall aspect evaluation values for the target, the parameter representing the distribution of each aspect in the review for learning for each of the plurality of learning reviews, and the review sentences of the plurality of review for learning, Estimating variational parameters including parameters representing the distribution of each aspect assigned to the review text And based on the estimated variation parameter, to estimate the model parameter, parameter estimation method according to claim 1 repeated until iteration termination condition is satisfied predetermined. An evaluation value estimation method in an evaluation value estimation device including an input unit and an evaluation value estimation unit,
The input unit accepting a review related to the evaluation target; and
The evaluation value estimation unit, based on the review received by the input unit, and the model parameter estimated by the parameter estimation method according to claim 1 or 2, the entire evaluation target for the review Estimating an evaluation value and an aspect evaluation value for each aspect of the evaluation target of the review;
Evaluation value estimation method including A plurality of reviews for learning related to a plurality of evaluation objects including a plurality of aspects, and a plurality of reviews for learning to which an overall evaluation value for the evaluation object and an aspect evaluation value for each aspect of the evaluation object are given in advance An input unit for estimation to accept;
Based on the plurality of learning reviews received by the estimation input unit, the overall aspect evaluation value of each aspect for the evaluation object and the overall evaluation value of the learning review become the overall evaluation value for the evaluation object. The aspect evaluation value of each aspect of the review for learning depends on the overall evaluation value of the aspect for each evaluation object and the overall evaluation value of the review for learning, and the aspect for each aspect is assigned. Parameters that represent the distribution of each word appearing in a review sentence, and each word appearing in a review sentence in which an aspect is assigned to each aspect and each evaluation value, and the aspect evaluation value of the aspect becomes the evaluation value According to the topic model representing that the parameter representing the distribution of
Parameters indicating the distribution of each aspect assigned to the review sentence, parameters indicating the distribution of each word appearing in the review sentence to which the aspect is assigned, and the aspects of each aspect and each evaluation value are assigned. A parameter representing the distribution of each word appearing in a review sentence in which the aspect evaluation value of the aspect is the evaluation value, an average and a standard deviation of the overall comprehensive evaluation values for the plurality of evaluation objects, and each aspect A standard deviation of overall evaluation values of the plurality of evaluation targets for the plurality of evaluation targets, a standard deviation of overall evaluation values of the plurality of learning reviews, and a standard deviation of the aspect evaluation values of the plurality of learning reviews for each aspect, and A parameter estimator for estimating model parameters including:
A parameter estimation apparatus including:   The program for functioning a computer as each part of the parameter estimation apparatus of Claim 4.