JP5139701B2 - Language analysis model learning apparatus, language analysis model learning method, language analysis model learning program, and recording medium thereof - Google Patents

Description

  The present invention relates to a language analysis model learning technique, and more particularly to a language analysis model learning technique for learning a language analysis model used for estimating a label to be assigned to a character string or a symbol string.

  Conventionally, for example, a label (also referred to as a label series or an output series) is used as a classification tag for an input (input series) having a series structure such as a character string or a symbol string related to text, DNA base sequence, web space, etc. The problem of granting is known. Such a problem is hereinafter referred to as a sequence structure prediction problem, and an apparatus (or program) that assigns an output sequence y to an input sequence x is referred to as a sequence structure predictor. Specific examples of the input series x and the output series y are shown in FIG.

  FIG. 18A relates to a problem of dividing a text into morphemes among sequence structure prediction problems, and shows an example in which a label indicating a linguistic feature is given to the text. The input series x is divided into 12 morphemes as “Real estate information registration evaluation system was established on the 13th”. Here, the morpheme is the smallest character string that is meaningless if it is made finer than this. The label “B” indicates the start position of the morpheme on the character string, and the label “I” indicates the position included in the range of the morpheme that starts with the label “B”.

  FIG. 18B relates to a problem of extracting a proper expression from text in a sequence structure prediction problem, and shows an example in which a label indicating the type of proper noun is given to a proper noun. The input sequence x is divided into 8 morphemes, such as “Ichiro Tanaka is the president of the Land Federation”. Of these, the four morphemes “Tanaka”, “Ichiro”, “Land”, and “Alliance” indicate specific expressions, so the labels “B-person name”, “I-person name”, “B-organization name”, “I-Organization name” is assigned. The label “O” indicates a morpheme other than the unique expression. Note that the two proper nouns “Ichiro Tanaka” and “National Land Federation” also show proper expressions, and thus the labels “person name” and “organization name” are assigned respectively.

  FIG. 18 (c) relates to a problem of estimating a gene region from a DNA base sequence in a series structure prediction problem. Three bases using four types of character strings (T, C, A, G) are shown. The example which provides the label which shows an amino acid with respect to a permutation (codon) is shown. Here, the codon “ATG” is given a start codon indicating the start of translation into protein and a label “M” indicating methionine. In addition, a codon “TGA” is provided with a termination codon indicating completion of translation into protein and a label “H” indicating histidine. Corresponding labels “R”, “D”, “W”, and “Q” are given to the codons sandwiched between the codon “ATG” and the codon “TGA”. Further, a label “O” is given to the character before (left side) the codon “ATG” and the character after (right side) the codon “TGA” to indicate that it is not a corresponding amino acid. .

  A conventional language analysis model learning device performs statistical learning (hereinafter referred to as supervised learning) in advance using correct data in which a correct label sequence is assigned to an input sequence for a target sequence structure prediction problem. Yes. Here, the correct answer data is data in which a set of an input sequence and an output sequence is known before learning. Then, a language analysis model indicating the correspondence between the input series and the label series is created using the parameters obtained by supervised learning. The sequence structure predictor can actually assign a label to the input sequence by using the created language analysis model.

  As a specific method of supervised learning, conventionally, a method of combining local optimum solutions has been used. In recent years, a method based on global optimization typified by a conditional random field has come to be used, and performance has been improved (for example, see Non-Patent Document 1).

This method of using a conditional random field is currently the most used in natural language analysis tasks such as chunking to read while analyzing for each “chunk” that is a unit of linguistic information processing and meaning, and extraction of proper expressions. It is widely used as one of the methods showing good performance. The conditional random field method is a method classified as a method of modeling the conditional probability p (y | x). A method for predicting the sequence structure by modeling the conditional probability p (y | x) in this way is hereinafter referred to as an identification approach.
J. Lafferty, A. McCallum and F. Pereira, Conditional Random Fields: Probabilistic Models for Segmenting and Labeling Sequence Data, In Proc. Of ICML-2001, Pages 282-289, 2001

Regardless of the sequence prediction problem, it is generally recognized that in order to perform prediction with good performance, a data amount that sufficiently covers the feature space handled during supervised learning is required. However, since correct data needs to be created manually, there is a problem that the creation cost is high.
On the other hand, since each instance y _i (i = 1, 2,...) Of the label sequence (output sequence) y handled in the learning related to the sequence prediction problem is interdependent, the feature space handled in the learning related to the sequence prediction problem. Will be very big. And in order to improve prediction performance, a huge amount of correct answer data is required. Moreover, since the sequence prediction problem is characterized by the interdependence of label sequences, in order to create correct data, higher specialized ability and cost are required for the target task.

  On the other hand, since the conventional language analysis model learning apparatus uses only a limited amount of correct answer data created manually, the performance of unknown data other than correct answer data is not good. In particular, when the sequence structure of data in another domain is predicted using the same task, the prediction performance may be significantly deteriorated because features unique to that domain cannot be learned. As a specific example, in the natural language processing specific expression extraction task, correct data of newspaper articles is generally used for model learning. When the sequence structure predictor extracts the web data by using the language analysis model obtained as a result of learning using the correct data of the newspaper article, the proper expression extraction performance for the web data is as follows. It is significantly lower than the proper expression extraction performance. This is because the web data is in a different domain from the newspaper article of the learning data.

  In addition, since the correct answer data is created manually, there is only a limited amount. For example, when applying a specific expression extraction task in a certain domain, in many domains including web data, there is no correct data in the domain to be applied. Therefore, in most cases, it is difficult to create a high-performance sequence structure predictor.

  On the other hand, acquisition of unlabeled data of a domain to be applied is often relatively easy. Here, unlabeled data means data to which no correct answer label is assigned, that is, data that has not been processed. Hereinafter, correct data having a correct label in contrast to the unlabeled data is referred to as labeled data.

  Since unlabeled data is not given information on correct labels (correct output series), it is not known how unlabeled data can be used for learning. In an attempt to design a model (classification model) with classical conditional probability to use unlabeled data for learning in the identification approach represented by the conditional random field method, there is no label from the conditional expression. Data terms are deleted. In other words, the identification approach shows good performance in the supervised learning setting, but it can be said that it is a difficult approach to capture unlabeled data.

  Further, for example, in a generation approach that is a method for designing a model (generation model) based on the joint probability p (x, y), when using unlabeled data for learning, an EM (Expectation Maximization) algorithm is used. There is a framework in which unlabeled data is taken in naturally and easily by using information on the correct label series (output series) as missing information. However, in this generation approach, in most cases, the prediction performance does not reach far compared with an identification approach in a supervised learning setting such as a conditional random field.

  Therefore, in the present invention, in order to solve the above-described problem and improve the prediction performance of the language analysis model at a low cost, a language that can learn a sequence structure predictor using labeled data and unlabeled data as input. The purpose is to provide analytical model learning technology.

In order to solve the above-mentioned problem, the language analysis model learning device according to claim 1 is provided with labeled data indicating data in which a character string or symbol string is labeled, and unlabeled data indicating a character string or symbol string. Is a language analysis model learning device that learns a language analysis model used to estimate a label to be assigned to a character string or a symbol string based on an identification model and a generation model, The identification model is a model that estimates the label to be assigned using a conditional probability indicating a probability that a predetermined label candidate appears on the condition of an input character string or symbol string, and the generation model includes: The label to be added is estimated using a joint probability indicating a probability that an input character string or symbol string and the predetermined label candidate are generated simultaneously. A model, using the unlabeled data the input, advance and parameter vector set for learning identification model, using the model integration parameter set predetermined you maximize first objective function and generating model learning means for determining a parameter vector set for generating model utilizes There labels that are the input data, and parameter vector set for the pre-learning identification model determined in the product model learning means by using the generative model for parameter vector set, and model integrated learning means for determining a pre-SL model integration parameter set that maximize the second objective function comprises, parameters for the product model in the generating the model learning unit vector The process of determining the set and the process of determining the parameter set for model integration by the model integrated learning means are alternately performed. And rows, one of said model integration parameter set and the generated model parameter vector set for learning means and said generating models model integrated learning means has determined alternately determine whether or not a predetermined convergence condition is satisfied And a convergence determination means for outputting the generation model parameter vector set and the model integration parameter set at that time when it is determined that the convergence condition is satisfied, and the first objective function includes the identification Using the model parameter vector set, the generated model parameter vector set, and the model integration parameter set, the sum of output values of the discriminant function for all outputs when unlabeled data is given is calculated. And the second objective function includes the identification model parameter vector set and the generation model parameter vector. It is a function that calculates the degree to which labeled data can be correctly identified by using a model set and the model integration parameter set .

The language analysis model learning method according to claim 6 uses, as input data, labeled data indicating data in which a character string or a symbol string is labeled, and unlabeled data indicating a character string or a symbol string, A language analysis model learning method of a language analysis model learning device for learning a language analysis model used for estimating a label to be assigned to a character string or a symbol string based on an identification model and a generation model, The identification model is a model that estimates the label to be assigned using a conditional probability indicating a probability that a predetermined label candidate appears on the condition of an input character string or symbol string, and the generation model includes: The label to be added is estimated using a joint probability indicating a probability that an input character string or symbol string and the predetermined label candidate are generated simultaneously. A model, by generating the model learning unit, using the unlabeled data the input, using pre and parameter vector set for learning identification model, and the model integration parameter set predetermined first determining a generate models for parameter vector set you maximize 1 objective function, the model integrated learning means, there labels that are the input using the data, and the parameter vector set in advance for learning identification model , using said generated model parameter vector set for learning generated model determined by means executes determining a pre SL model integration parameter set that maximize the second objective function, the alternately convergence determination by means para for the generative model learning means and said generating models model integrated learning means has determined alternately One of said model integration parameter set and data vector set is determined whether or not a predetermined convergence condition is satisfied, when it is determined that the convergence condition is satisfied, and the parameter vector set for generating models that point look including the step of outputting said model integration parameter set, wherein the first objective function using a parameter vector set for the identification model, and parameter vector set for the generation model, and the model integration parameter set And calculating the sum of the output values of the discriminant function for all outputs when unlabeled data is given, wherein the second objective function includes the discriminant model parameter vector set and the generated model parameter Using the vector set and the parameter set for model integration, calculate the degree to which labeled data can be correctly identified It is a function to perform.

According to the language analysis model learning device according to claim 1 or the language analysis model learning method according to claim 6 , the language analysis model learning device determines a generation model parameter vector set using unlabeled data. Thus, the unlabeled data is captured by the generation approach. Then, the language analysis model learning device determines the model integration parameter set by using the determined generation model parameter vector set and the labeled data, thereby identifying the unlabeled data captured by the generation approach by the identification approach. Can learn. Then, the language analysis model learning device determines the optimum generation model parameter vector set and model integration parameter set by alternately determining until one of the generation model parameter vector set and the model integration parameter set converges. Output. Therefore, the language analysis model that consists of these output generation model parameter vector set and model integration parameter set and learned identification model parameter vector set can achieve high prediction performance at low cost. It is.

Further, the language analysis model learning device according to claim 2 is the language analysis model learning device according to claim 1, wherein the generation model learning means has an object represented by an expression (14) described later as the first objective function. An objective function calculation means for calculating the function G (Θ | Γ), and a parameter vector set Θ that maximizes the objective function G (Θ | Γ) represented by the following formula (14) under fixed Λ and Γ Auxiliary function calculation means for performing a process for obtaining a parameter update means for updating the obtained generation model parameter vector set Θ when the convergence determination means determines that the convergence condition is not satisfied. , The model integrated learning means uses the objective function L shown in the following equation (16) as the second objective function. ^SS-Hyb Objective function calculation means for calculating (Γ | Θ) under fixed Λ and Θ, and an objective function L shown in the following equation (16) ^SS-Hyb (Γ | Θ) is the model integration parameter γ for the discrimination model _i And partial differentiation calculation means for identification model for performing partial differentiation calculation with an objective function L shown in equation (16) described later ^SS-Hyb Model integration parameter γ for generating model (Γ | Θ) _j A partial differential calculation means for a generation model that performs a partial differentiation calculation at the time, and a parameter update means that updates the obtained model integration parameter set Γ when the convergence determination means determines that the convergence condition is not satisfied And.
Further, the language analysis model learning method according to claim 7 is the language analysis model learning method according to claim 6, wherein the step of determining the generation model parameter vector set is an expression described later as the first objective function. By calculating the objective function G (Θ | Γ) shown in (14) and calculating a predetermined auxiliary function, the objective function G (Θ | Γ) shown in the following formula (14) is obtained. A step of obtaining a parameter vector set Θ that is maximized under fixed Λ and Γ, and when the convergence determination means determines that the convergence condition is not satisfied, the generated generation model parameter vector Updating the set Θ, and the step of determining the model integration parameter set is an objective function L shown in the following equation (16) as the second objective function: ^SS-Hyb A step of calculating (Γ | Θ) under fixed Λ and Θ, and an objective function L shown in the following equation (16) ^SS-Hyb (Γ | Θ) is the model integration parameter γ for the discrimination model _i And a step of performing partial differentiation with the objective function L shown in the following equation (16) ^SS-Hyb Model integration parameter γ for generating model (Γ | Θ) _j And a step of performing partial differentiation in step (a) and updating the obtained model integration parameter set Γ when the convergence determination means determines that the convergence condition is not satisfied. To do.

The language analysis model learning device according to claim 3 is the language analysis model learning device according to claim 2, wherein the auxiliary function for maximizing the first objective function and the second objective function are A joint probability of an input sequence and an output sequence estimated from the identification model parameter vector set using labeled data, and the model integration parameter set determined in advance for the identification model parameter vector set; The probability value calculated based on the identification model parameter vector set to be integrated over the identification model integration probability value indicating the result of integration, and the generation model parameter using the unlabeled data Calculated in advance for the joint probability of the input sequence and output sequence estimated from the vector set and the parameter vector set for the generation model. A probability value calculated based on the generated model integration parameter set, and a product model integration probability value indicating a result obtained by integrating the probability values calculated over the generation model parameter vector set to be integrated, The auxiliary function calculation means includes a Q function shown in the following formula (15) as the auxiliary function, which is included as a parameter set indicating the posterior probability of the label to be given to the input character string or symbol string. The parameter vector set Θ ′ that maximizes the Q function is obtained from the current parameter vector set Θ ′, and the Q function is repeatedly obtained by replacing Θ with Θ ′ until Θ ′ does not increase with respect to Θ. A parameter vector set Θ that maximizes the objective function G (Θ | Γ) is obtained.
The language analysis model learning method according to claim 8 is the language analysis model learning method according to claim 7 , wherein the auxiliary function for maximizing the first objective function and the second objective function are A joint probability of an input sequence and an output sequence estimated from the identification model parameter vector set using labeled data, and the model integration parameter set determined in advance for the identification model parameter vector set; The probability value calculated based on the identification model parameter vector set to be integrated over the identification model integration probability value indicating the result of integration, and the generation model parameter using the unlabeled data Calculated in advance for the joint probability of the input sequence and output sequence estimated from the vector set and the parameter vector set for the generation model. A probability value calculated based on the generated model integration parameter set, and a product model integration probability value indicating a result obtained by integrating the probability values calculated over the generation model parameter vector set to be integrated, It viewed including as a parameter set indicating a posteriori probability of the label to be assigned to a character string or symbol string is the input, the step of performing a process of obtaining the parameter vector set Θ will be described later as the auxiliary function (15) The parameter vector set Θ ′ that maximizes the Q function is obtained from the current parameter vector set Θ, and is repeated while replacing Θ with Θ ′ until Θ ′ does not increase with respect to Θ. By obtaining a Q function, a parameter vector set Θ that maximizes the objective function G (Θ | Γ) is obtained .

According to the language analysis model learning process according to the language analysis model learning apparatus or claim 8 as claimed in claim 3, the language analysis model learning device, the auxiliary function to maximize the first objective function and the second object Since the function includes the product of the identification model integration probability value using the labeled data and the generation model integration probability value using the unlabeled data as the posterior probability of the label to be given to the input series, The generation model parameter vector set and model integration parameter set determined by maximizing the first objective function and the second objective function respectively reflect the result of learning unlabeled data and labeled data. It becomes. Here, the unlabeled data is used as an auxiliary function for maximizing the first objective function, and the unlabeled data is not directly used as the first objective function. That is, it is possible to combine the good prediction performance of the identification approach while capturing unlabeled data by the generation approach. Here, the joint probability estimated from the identification model parameter vector is configured by, for example, a conditional random field (CRF). Moreover, the joint probability estimated from the parameter vector for generation | occurrence | production models is comprised by the hidden Markov model (HMM: Hidden Markov Model), for example.

The language analysis model learning device according to claim 4 is the language analysis model learning device according to claim 3 , wherein the output generation model parameter vector set and the model integration parameter set to be output are learned in advance. The method further comprises parameter integration means for integrating the identified identification model parameter vector set into the parameter set indicating the posterior probability.

The language analysis model learning method according to claim 9 is the language analysis model learning method according to claim 8 , wherein the output model parameter vector set and the model integration parameter set output by the parameter integration unit are provided. And the step of integrating the previously learned parameter vector set for the identification model into the parameter set indicating the posterior probability.

According to the language analysis model learning device according to claim 4 or the language analysis model learning method according to claim 9 , the language analysis model learning device has a label to be added to the input character string or symbol string. As a posterior probability, the generated model parameter vector set, the model integration parameter set, and the identification model parameter vector set are integrated into a single parameter vector (parameter set). Therefore, the integrated parameter vector can be easily used as a language analysis model.

Moreover, the language analysis model learning device according to claim 5 is the language analysis model learning device according to any one of claims 1 to 4 , wherein the input labeled data using the identification model. It further comprises an identification model learning means for creating the identification model parameter vector set by learning.

Also, the language analysis model learning method according to claim 10, at the language analysis model learning process according to any one of claims 6 to 9, by the identification model learning unit, using the identification model And learning the input labeled data to create the identification model parameter vector set.

According to the language analysis model learning device according to claim 5 or the language analysis model learning method according to claim 10 , the language analysis model learning device creates a parameter vector set for the identification model by learning the labeled data. To do. Therefore, it is possible to share the configuration for creating the identification model parameter vector set and the configuration for creating the model integration parameter set.

Also, language analysis model learning program according to claim 11 is a program for executing the language analysis model learning process according to a computer in any one of claims 6 to 10. By being configured in this way, a computer in which this program is installed can realize each function based on this program.

A computer-readable recording medium according to a twelfth aspect is characterized in that the language analysis model learning program according to the eleventh aspect is recorded. By being configured in this way, a computer equipped with this recording medium can realize each function based on a program recorded on this recording medium.

  According to the present invention, it is possible to learn a sequence structure predictor using labeled data and unlabeled data as inputs. As a result, the prediction performance of the language analysis model can be improved at low cost.

  Hereinafter, the best mode (hereinafter referred to as “embodiment”) for carrying out the language analysis model learning device and language analysis model learning method of the present invention will be described in detail with reference to the drawings. In the present embodiment, the language analysis model learning apparatus will be described with respect to the problem of specific expression extraction using text as input and label string as output.

[Configuration of language analysis model creation device]
FIG. 1 is a configuration diagram schematically showing an outline of a language analysis model creation device according to an embodiment of the present invention. As shown in FIG. 1, the language analysis model creation device 1 includes a language analysis model learning device 2 and a parameter integration device 3. In the learning phase, the language analysis model learning device 2 uses the labeled data D ₁ , the unlabeled data _Du, and the information stored in the learning support information storage unit 4 to process the processing results in the parameter set storage unit 5. Output to.

Of labeled data D _l shows data label is applied to the string or symbol string. Here, as shown in Expression (1), the N sets of labeled samples S ₁ = (x ⁿ , y ⁿ ) are referred to as labeled data D ₁ . The unlabeled data _Du represents a character string or a symbol string. Here, as shown in Expression (2), the M sets of unlabeled samples S _u = (x ^m ) are referred to as unlabeled data D _u .

  The learning support information storage unit 4 stores, as learning support information, an identification model feature extraction template, a generated model feature extraction template, and an output label candidate, which will be described later.

Further, the language analysis model learning device 2 shows, as processing results, an identification model parameter vector set Λ shown by Expression (3), a generated model parameter vector set Θ shown by Expression (4), and Expression (5). The model integration parameter set Γ is stored in the parameter set storage means 5. Here, λ _i is an identification model parameter vector, θ _j is a generation model parameter vector, and γ _i and γ _j are model integration parameters. The identification model is a model for estimating a label to be given using a conditional probability indicating a probability that a predetermined label candidate appears on the condition of an input character string or symbol string. The generation model is a model for estimating a label to be assigned using a joint probability indicating a probability that an input character string or symbol string and a predetermined label candidate are generated at the same time.

The parameter integration device (parameter integration means) 3 uses a generated model parameter vector set Θ, a model integration parameter set Γ, and an identification model parameter vector set Λ outputted from the language analysis model learning device 2 as a single unit. It is integrated into the parameter set. R shown in Equation (6) indicates the posterior probability of the label to be assigned to the input character string or symbol string. R is a probability value for integrating an identification model indicating the result of accumulating probability values calculated by powering p _i ^D by γ _i over i, and a probability calculated by raising p _j ^G to a power of γ _j It is in the form of a product with the generation model integration probability value indicating the result of integrating the values over j. The parameter integration device 3 stores the integrated parameter set in the language analysis model storage unit 6.

In the evaluation phase, the sequence structure prediction apparatus 7, to the unlabeled sample S _u is the input data, language analysis model stored in the language analysis model storage unit 6 (parameter set integrated by the parameter integrating unit 3) Is used to estimate the label to be given and output a labeled sample S ₁ corresponding to the input data.

[Configuration of language analysis model learning device]
FIG. 2 is a functional block diagram schematically showing the configuration of the language analysis model learning device shown in FIG. The language analysis model learning device 2 includes, for example, a CPU (Central Processing Unit), a RAM (Random Access Memory), a ROM (Read Only Memory), an HDD (Hard Disk Drive), an input / output interface, and the like. As shown in FIG. 2, an identification model learning unit 10 and a sequence structure predictor learning unit 20 are provided.

<Example of input data>
FIG. 3 is a diagram illustrating an example of information input to the language analysis model learning apparatus illustrated in FIG. 2, where (a) illustrates data with labels and (b) illustrates an output label candidate set. The example shown in FIG. 3A is the same as that shown in FIG. However, the morpheme separation is assumed to be performed in advance.

  The output label candidate set shown in FIG. 3B corresponds to the specific expression extraction, and includes five predetermined output label candidates as elements. Each element is the same as that shown in FIG. This output label candidate set is automatically determined according to the target problem, and is acquired by the language analysis model learning device 2 from the output label candidate storage means 43 of the learning support information storage means 4.

<Identification model learning means>
The discrimination model learning means 10 creates the discrimination model parameter vector set Λ by learning the input labeled data D _l using the discrimination model. The discrimination model learning means 10 and the feature extraction Means 12 and parameter learning means 13 are provided. Among these, the output candidate graph generation means 11 and the feature extraction means 12 are for performing preprocessing for learning. Further, the identification model learning means 10 uses the labeled data D _{l ′} shown in Expression (7).

<< Output candidate graph generation means >>
The output candidate graph generation means 11 generates an output candidate graph from the input labeled data D _{l ′} . As shown in FIG. 4, the output candidate graph is a lattice format in which all possible output series candidates are connected by a path. FIG. 4 is a diagram schematically illustrating an example of an output candidate graph generated by the language analysis model learning device illustrated in FIG. Here, <BOS> is a fixed special label indicating the beginning of the input sequence x, and <EOS> is a fixed special label indicating the end of the input sequence x. The lattice indicates that the output series y for the input labeled data D _{l ′} (input series x) is an individual instance y _i (i = 1,..., 5) as a node, and a dependency relationship between each instance is indicated by a link. It is a thing. One path between <BOS> and <EOS> in the output candidate graph corresponds to one output, and the output candidate graph is a graph including all possible output candidates. For example, the node 401 indicates an output instance in which the label “B-organization name” is assigned to the fourth word “land” of the labeled data D _{1 ′} . Similarly, the node 402 represents an output instance in which the label “O” is assigned to the sixth word “no” of the labeled data D _{l ′} .

≪Feature extraction means≫
The feature extraction means 12 extracts features from the output candidate graph by combining the position label to be estimated and the instances in the input series described in the feature extraction template for identification model. A specific example of the feature extraction template for the identification model is shown in FIG. FIG. 5 is a diagram illustrating an example of an identification model feature extraction template. The identification model feature extraction template 500 is a template for estimating a label of “y ₄ ” as the i-th output sequence. The identification model feature extraction template 500 extracts input words corresponding to up to two labels before and after the position label to be estimated as a feature. Therefore, when the feature extraction unit 12 estimates the label at the position of “y ₄ ” shown in FIG. 5 when using the identification model feature extraction template 500, reference numerals 601 to 605 in FIG. Extract features to show. Specifically, when the label at the position of “y ₄ ” is estimated as “B-organization name” as the i-th output sequence shown in FIG. 5, the features indicated by reference numerals 611 to 615 in FIG. To extract. Similarly, when the label at the position of “y ₄ ” is estimated as “B-person name” as the i-th output sequence shown in FIG. 5, the features indicated by reference numerals 621 to 625 in FIG. 6C are extracted. .

  In addition, the feature extraction unit 12 assigns a feature vector to each node (or link) of the output candidate graph using the output candidate graph and the identification model feature extraction template, and generates an output candidate graph with a feature vector. FIG. 7 shows the relationship between the output candidate graph and the feature vector. FIG. 7 is an explanatory diagram of feature vectors created using the identification model feature extraction template shown in FIG. 5 for the nodes indicated by circles in FIG. FIG. 7A shows the characteristics given to the node 401 shown in FIG. The feature extraction unit 12 associates a value “1” with these features, and associates a value “0” with another feature not assigned to the node 401, thereby obtaining “1” and “0”. A feature vector for an element is formed. Since the feature extraction unit 12 generates a feature by combining one of the two words before and after and the output label to which the target node belongs, the feature vectors of the nodes having different output labels at the same position of the words in the input sequence are respectively Since they are orthogonal to each other, their inner product is “0”. FIG. 7B shows the characteristics given to the node 402 shown in FIG. Since this is also the same, description is omitted.

≪Parameter learning means≫
The parameter learning means 13 performs supervised learning of an identification approach using an output candidate graph with feature vectors and an initialized parameter vector (a vector in which all elements are 0). Here, the parameter learning unit 13 performs learning of the identification model parameter vector set Λ based on the conditional random field, as shown in FIG. 2, the objective function calculation unit 131 and the objective function gradient calculation unit 132. And a convergence determination unit 133 and a parameter update unit 134. The conditional random field is described in detail in, for example, “F. Sha and F. Pereira, Shallow Parsing with Conditio1al Random Fields, In Proc. Of HLT / NAACL-2003. Pages 213-220, 2003”. Therefore, explanation is omitted.

The objective function calculation unit 131 calculates an objective function to which the identification model parameter vector λ is input. Here, as a premise, a local feature vector f _s obtained from a position s in a sequence when an input sequence x (labeled data) is given is defined by Expression (8).

  The conditional random field defines the conditional probability p (y | x) using the product of the potential function on each clique and the overall ratio. That is, the conditional probability p (y | x) of the output sequence y with respect to the input sequence x by the conditional random field is defined by Expression (9). Further, Z (x) in equation (9) corresponds to a normalization term for all outputs y as shown in equation (10).

The objective function calculation unit 131 maximizes the (logarithm) posterior probability of the parameter vector λ using the given labeled data D _{l ′} . That is, the objective function calculation means 131 maximizes logp (λ | D _{l ′} ). Specifically, the objective function calculation unit 131 calculates an objective function L ^CRF (λ) represented by Expression (11). However, p (λ) represents the prior probability distribution of λ. In order to optimize the objective function shown in Expression (11), a numerical optimization method based on a gradient such as L-BFGS can be applied. Note that L-BFGS is described in “DC Liu and J. Nocedal, On the Limited Memory BFGS Method for Large Scale Optimization Math. Programming, Ser. B, 45 (3): 503-528, 1989”. Therefore, explanation is omitted.

The objective function gradient calculation means 132 is for calculating the gradient of the objective function shown in the equation (11). The gradient ∇L ^CRF (λ) of the objective function shown in Expression (11) is expressed by Expression (12).

  Here, E indicates the expected value of the subscript. Also, the first term on the right side is an empirical expected value of the feature vector, so it can be easily calculated by counting the features appearing in the correct answer sequence from the labeled data. Further, since the value depends only on the labeled data, it may be calculated once before learning. The second term on the right side of equation (12) is an expected value at which each feature vector for all output sequences appears. Thus, all possible output sequences need to be calculated individually. Moreover, in the structure prediction problem, the total number of all possible outputs is generally very large, and it is very difficult to calculate all possible outputs individually from the viewpoint of computational complexity. However, regarding the sequence structure prediction problem, it is known that the expected value can be calculated efficiently using the forward-backward algorithm, so that it is possible to execute the process in a realistic time (non-patent) Reference 1).

The convergence determination unit 133 determines whether or not the difference between the empirical expected value of each feature and the expected value of each feature for all outputs has converged. Specifically, the convergence determination unit 133 determines whether or not the value of the gradient ∇L ^CRF (λ) of the objective function expressed by Expression (12) has converged. When the convergence determining unit 133 determines that the value of the gradient ∇L ^CRF (λ) of the objective function has converged, the convergence model parameter vector λ at that time is stored in the identification model parameter vector set storage unit 51. Output. An example of the output identification model parameter vector (column vector) is shown in FIG. The parameter updating unit 134 updates the identification model parameter vector λ when the value of the gradient ∇L ^CRF (λ) of the objective function has not converged.

In the present embodiment, the identification model learning means 10 can learn a plurality of identification models from the same labeled data. For example, a plurality of different identification models can be created by changing the identification model feature extraction template. Note that the number of identification models is not particularly limited, and the designer can freely determine, for example, 1 to several thousand according to the task or the like. In the present embodiment, all the different (I) identification model parameter vectors λ _i created by the identification model learning means 10 are represented by the above-described equation (3).

<Sequence structure predictor learning means>
The sequence structure predictor learning unit 20 receives the identification model parameter vector set Λ stored in the identification model parameter vector set storage unit 51 as an input, and generates a generation model learning unit 21 and a model integrated learning unit 22 that execute processing alternately. And convergence determining means 23 for performing this alternate processing until a predetermined end condition is reached.

The generation model learning means 21 uses the input unlabeled data _Du to learn the identification model parameter vector set Λ previously learned and the predetermined model integration parameter set Γ. A generation model parameter vector set Θ that maximizes one objective function is determined.

The model integrated learning means 22 maximizes the second objective function by learning the identification model parameter vector set Λ and the generated model parameter vector set Θ using the input labeled data D _l. Such a model integration parameter set Γ is determined.

  The convergence determination unit 23 causes the generation model learning unit 21 and the model integration learning unit 22 to alternately determine the generation model parameter vector set Θ and the model integration parameter set Γ, one of which is a predetermined convergence. When the condition is satisfied, the generation model parameter vector set Θ and the model integration parameter set Γ at that time are output.

<Learning support information storage means>
The learning support information storage unit 4 includes an identification model feature extraction template storage unit 41, a generated model feature extraction template storage unit 42, and an output label candidate storage unit 43. The identification model feature extraction template storage means 41 stores an identification model feature extraction template. The generation model feature extraction template storage means 42 stores the generation model feature extraction template. The output label candidate storage unit 43 stores output label candidates. Each storage means 41, 42, 43 is composed of, for example, a general hard disk or memory.

<Parameter set storage means>
The parameter set storage unit 5 includes an identification model parameter vector set storage unit 51, a generated model parameter vector set storage unit 52, and a model integration parameter set storage unit 53. The identification model parameter vector set storage means 51 stores the identification model parameter vector set Λ. The generation model parameter vector set storage means 52 stores the generation model parameter vector set Θ. The model integration parameter set storage means 53 stores the model integration parameter set Γ. Each storage means 51, 52, 53 is composed of a general hard disk or memory, for example.

[Configuration of Sequence Structure Predictor Learning Means]
FIG. 9 is a functional block diagram schematically showing the configuration of the sequence structure predictor learning unit shown in FIG.
≪Generation model learning means≫
The generation model learning unit 21 includes an output candidate graph generation unit 211, a feature extraction unit 212, an objective function calculation unit 213, an auxiliary function calculation unit 214, and a parameter update unit 215. Among these, the output candidate graph generation unit 211 and the feature extraction unit 212 are for performing preprocessing for learning. Further, the generation model learning means 21 uses the unlabeled data _Du shown in the above equation (2).

The output candidate graph generation unit 211 is the same as the output candidate graph generation unit 11 of the identification model learning unit 10 shown in FIG.
The feature extraction unit 212 extracts a feature from the output candidate graph by combining the estimated position label and the instance in the input series described in the generated model feature extraction template. A specific example of the generated model feature extraction template is shown in FIG. FIG. 10 is a diagram illustrating an example of a generated model feature extraction template. The basic format of the generation model feature extraction template is the same as that of the identification model feature extraction template, and the feature extraction method is also the same. The only difference is that the feature extraction template for the generated model needs to satisfy the condition that the features extracted as constraints of the hidden Markov model are independent from each other. Therefore, when this independent condition is satisfied, the same template as the identification model feature extraction template may be used. Moreover, it may be a completely different template, and can be designed freely according to the target task and prior knowledge.

The generation model feature extraction template 1000 illustrated in FIG. 10 extracts as input features a label at an estimated position and an input word corresponding to the preceding label. Therefore, the feature extraction unit 212 extracts the feature shown in FIG. 11A when estimating the label at the position “y ₄ ” shown in FIG. 10 when the generated model feature extraction template 1000 is used. . Specifically, when the label at the position of “y ₄ ” is estimated as “B-organization name” as the i-th output sequence shown in FIG. 10, the feature shown in FIG. 11B is extracted. Similarly, when the label at the position of “y ₄ ” is estimated as “B-person name” as the i-th output sequence shown in FIG. 10, the feature shown in FIG. 11C is extracted. Further, the feature extraction unit 212 assigns a symbol generation probability and a transition probability to each node and link of the output candidate graph using the output candidate graph and the generation model feature extraction template, and outputs an output candidate graph with probability.

The objective function calculation means 213 calculates an objective function (first objective function) G to which the output candidate graph with probability and the generation model parameter vector set Θ ^(t) are input. As a premise of the objective function G, based on R (y | x; Λ, Θ, Γ) of Equation (6) described above, a discriminant function g (y Consider | x; Λ, Θ, Γ). Since the denominator on the right side of the equation ( 6 ) is a normalized term, it does not contribute to the determination of y. Therefore, the discriminant function g (y | x; Λ, Θ, Γ) uses only the numerator on the right side of the equation ( 6 ). It can be defined as follows:

  The sum of the output values of the discrimination function g for all outputs y is defined by an objective function G (Θ | Γ) shown in Expression (14). Here, p (Θ) represents a prior probability distribution for the generation model parameter vector set Θ. Therefore, the objective function calculation means 213 calculates an objective function G (Θ | Γ) shown in Expression (14).

  The auxiliary function calculating unit 214 performs processing for obtaining a parameter vector set Θ that maximizes the objective function G (Θ | Γ) represented by the equation (14). That is, when Γ is known, Θ that maximizes G (Θ | Γ) in the vicinity of the initial value can be estimated by iterative calculation such as the EM algorithm. Specifically, the auxiliary function calculating unit 214 obtains a parameter vector set Θ ′ that maximizes the Q function (auxiliary function) shown in Expression (15) from the current parameter vector set Θ, and Θ ′ is equal to Θ. The Q function is repeatedly obtained while Θ is replaced with Θ ′ until it no longer increases.

Since the form of the Q function (Q (Θ ′, Θ; Γ)) shown in Equation (15) is the same as the hidden Markov model, the parameters can be efficiently obtained using the Baum-Welch algorithm used in the hidden Markov model. Can be updated. However, in the hidden Markov model, the peripheral probability is calculated using the conditional probability p (y | x; θ), but in this embodiment, R (y | x; Λ, Θ, Γ shown in Expression (6). ) Is used to calculate the marginal probability.
The parameter updating unit 215 updates the generated model parameter vector set Θ ^(t) when the convergence determining unit 23 determines that the convergence has not occurred.

In this product model learning unit 21, the objective function G of the formula (14) using the unlabeled data D _u | thereby maximizing (theta gamma) is for any unknown input sequence x, an output sequence y This means that the discriminant function g in between gives a large value, which contributes to increasing the reliability of discrimination. This is because if the situation is such that the discriminant function g gives a very small value to all output series y, the difference in the value of the discriminant function g between all the output series y is very large. This is because it is equivalent to the fact that it becomes smaller and becomes almost the same value, so it can be considered that the reliability of identification is low. Further, the sequence structure predictor learning unit 20 is a of being only utilized to maximize (increase) the sum of the unlabeled data D _u for all the output of the discriminant function g, the final the unlabeled data D _u It is not used directly for optimizing typical sequence structure predictors. This is because the correct output of the unlabeled data _Du is unknown, so that it cannot contribute to the optimization of the discriminator for the output sequence y.

≪Model integrated learning means≫
The model integrated learning means 22 performs parameter estimation (MAP estimation) by posterior probability maximization (MAP: Maximum A Posteriori) for the model integration parameter set Γ for an arbitrary generated model parameter vector set Θ. Then, a parameter set Γ for model integration that maximizes the objective function (second objective function) is estimated. The model integrated learning unit 22 includes an output candidate graph generation unit 221, a feature extraction unit 222, an objective function calculation unit 223, an identification model partial differential calculation unit 224, a generation model partial differential calculation unit 225, a parameter Update means 226. Among these, the output candidate graph generation unit 221 and the feature extraction unit 222 are the same as the output candidate graph generation unit 11 and the feature extraction unit 12 of the identification model learning unit 10 shown in FIG. Further, the model integrated learning means 22 uses the labeled data D _l shown in the above equation (1).

The objective function calculation means 223 calculates an objective function (second objective function) L ^SS-Hyb to which the model integration parameter set Γ ^(t) is input. The objective function L ^SS-Hyb (Γ | Θ) is defined by Expression (16). However, p (Γ) is a prior probability distribution of Γ.

Since the objective function L ^SS-Hyb (Γ | Θ) is a convex function with respect to the parameter set Γ on any fixed Θ, this optimization guarantees a global optimal solution. Therefore, if the gradient of the objective function L ^SS-Hyb (Γ | Θ) is calculated, a solution can be easily obtained by applying an optimization algorithm using a gradient such as L-BFGS.

The partial differential calculation means 224 for the identification model performs a calculation for partial differentiation of the objective function L ^SS-Hyb (Γ | Θ) shown in Expression (16) by the model integration parameter γ _i for the identification model (CRF in this embodiment). Is what you do. Specifically, the partial differential calculation means for identification model 224 calculates the right side of Expression (17). Since the first term and the second term on the right side of Equation (17) are constants during the optimization process, they need only be calculated once in advance. The calculation of the third term on the right side of Equation (17) will be described later for convenience of explanation.

The partial differential calculation means for generation model 225 performs a calculation for partial differentiation of the objective function L ^SS-Hyb (Γ | Θ) shown in Expression (16) with the model integration parameter γ _j for the generation model. Specifically, the generation model partial differential calculation means 225 calculates the right side of the equation (18).

Since the first term on the right side of Equation (18) becomes a constant during the optimization process, it may be calculated once in advance. Next, the calculation of the second term on the right side of Equation (18) will be described together with the calculation of the third term on the right side of Equation (17). Here, if the denominator on the right side of Equation (6) is expressed by N _R (x), Equation (6) can be expressed as Equation (19).

According to this equation (19), the cost of each position s is obtained by the sum of the values corresponding to each position s of the identification model and the generation model, and the conditional probability R (y | x shown in equation (19) ; Λ, Θ, Γ) is expressed as the sum of the cost at all positions and the overall ratio. Since the third term on the right side of Equation (17) and the second term on the right side of Equation (18) are expected values for the output values of each model, they can be efficiently calculated from Equation (19) using the forward-backward algorithm. . In other words, the estimation of Γ can be efficiently derived using the same forward-backward algorithm as conventionally used in the conditional random field.
The parameter updating unit 226 updates the model integration parameter set Γ ^(t) when the convergence determining unit 23 determines that the model has not converged.

In the present embodiment, the convergence condition determined by the convergence determination means 23 is determined to have converged when Δ shown in Expression (20) becomes a predetermined value or less. This is because the model integration parameter set Γ has a global optimal solution for the fixed generation model parameter vector set Θ. For example, | Θ ^(t) −Θ ^(t−1) |, | Γ ^(t) −Γ ^(t−1) |, etc. may be used instead of Δ shown in Expression (20). Absent.

  With the above configuration, when the convergence determination unit 23 determines that the convergence has occurred, the generation model parameter vector set Θ at that time is output to the generation model parameter vector set storage unit 52 and the model integration at that time is integrated. The parameter set Γ for use is output to the parameter set storage means 53 for model integration. An example of the output generation model parameter vector (column vector) is shown in FIG. An example of the model integration parameter set to be output is shown in FIG. 13 in a column vector format.

  The generation model learning means 21 and the model integrated learning means 22 are realized by the CPU developing and executing a predetermined program stored in the HDD or the like of the storage means on the RAM. .

[Operation of language analysis model creation device]
The operation of the language analysis model creation apparatus shown in FIG. 1 will be described with reference to FIG. 14 (refer to FIGS. 1 and 2 as appropriate), mainly focusing on the operation of the language analysis model learning apparatus 2. FIG. 14 is a flowchart showing the operation of the language analysis model creation device shown in FIG. The language analysis model learning device 2 of the language analysis model creation device 1 inputs the labeled data D _l , the unlabeled data D _u , and learning support information (step S1). As learning support information, “output label candidate set” automatically determined by the target problem, “identification model feature extraction template and generation model feature extraction template” manually determined by the target problem Enter.

Then, the language analysis model learning device 2 uses the identification model learning means 10 to perform supervised learning processing (conditional random field using the labeled data D _{l ′} , the output label candidate set, and the identification model feature extraction template. ) (Step S2), and outputs the identification model parameter vector set Λ obtained as a result of the supervised learning process to the identification model parameter vector set storage means 51 (step S3). Then, the language analysis model learning device 2 acquires and inputs the identification model parameter vector set Λ from the identification model parameter vector set storage unit 51 by the sequence structure predictor learning unit 20 (step S4), and generates the generated model. The parameter vector set Θ and the model integration parameter set Γ are initialized at t = 0 (step S5).

Next, the language analysis model learning device 2 uses the generation model learning unit 21 of the sequence structure predictor learning unit 20 to generate the unlabeled data D _u , the output label candidate set, and the generation model feature extraction template. Model parameter vector set estimation processing is executed (step S6). Although this process will be described in detail later, Θ is estimated under fixed Λ and Γ. Then, the language analysis model learning device 2 uses the model integrated learning unit 22 of the sequence structure predictor learning unit 20 to execute the model integration parameter set estimation process using the labeled data D _L and the output label candidate set ( Step S7). Although this process will be described later in detail, Γ is estimated under fixed Λ and Θ.

  Then, the language analysis model learning device 2 determines whether or not the model integration parameter set Γ has converged by the convergence determination unit 23 (step S8). If not converged (step S8: No), the language analysis model learning device 2 adds “1” to the current value of t by the sequence structure predictor learning unit 20 (step S9), and returns to step S6. On the other hand, when it converges (step S8: Yes), the language analysis model learning device 2 outputs the current generation model parameter vector set Θ and the model integration parameter set Γ by the sequence structure predictor learning unit 20 (step S8). S10). Then, the parameter integration device 3 of the language analysis model creation device 1 generates the generated model parameter vector set Θ, the model integration parameter set Γ, and the identification model parameter vector set Λ output from the language analysis model learning device 2. Integration into a single parameter set (step S11).

<Supervised learning process>
FIG. 15 is a flowchart showing the supervised learning process shown in FIG. The identification model learning means 10 of the language analysis model learning device 2 generates an output candidate graph from the input labeled data D1 _′ by the output candidate graph generation means 11 (step S21). Then, the identification model learning unit 10 extracts a feature from the output candidate graph by the combination of the label of the position to be estimated and the instance in the input sequence described in the feature extraction template for the identification model by the feature extraction unit 12 ( Step S22). Then, the discrimination model learning means 10 inputs the initial value of the discrimination model parameter vector λ to the parameter learning means 13 (step S23). Next, the parameter learning means 13 of the discrimination model learning means 10 calculates the objective function L ^CRF (λ) represented by the above-described equation (11) by the objective function calculation means 131 (step S24). Then, the parameter learning means 13 calculates the objective function gradient ∇L ^CRF (λ) based on the above equation (12) by the objective function gradient calculating means 132 (step S25), and the convergence determining means 133 calculates the objective function gradient ∇L ^CRF (λ). It is determined whether or not the value of the function gradient ∇L ^CRF (λ) has converged (step S26). When it has converged (step S26: Yes), the discrimination model learning means 10 outputs the discrimination model parameter vector λ at that time to the discrimination model parameter vector set storage means 51 (step S27). On the other hand, when it has not converged (step S26: No), the parameter learning unit 13 updates the parameter vector λ by the parameter update unit 134 (step S28).

<Generation model parameter vector set estimation processing>
FIG. 16 is a flowchart showing the generation model parameter vector set estimation processing shown in FIG. Generating model learning unit 21 of the language analysis model learning device 2, the output candidate graph generation unit 211 generates an output candidate graph from unlabeled data D _u which is input (step S41). Then, the generation model learning unit 21 extracts a feature from the output candidate graph by a combination of the label of the position to be estimated and the instance in the input sequence described in the generation model feature extraction template by the feature extraction unit 212 ( Step S42). Then, the generation model learning unit 21 inputs the generation model parameter vector set Θ ^(t) to the objective function calculation unit 213 (step S43). Next, the generation model learning unit 21 calculates the objective function G (Θ | Γ) shown in the above equation (14) by the objective function calculation unit 213 (step S44), and the auxiliary function calculation unit 214 calculates the equation ( The Q function (Q (Θ ′, Θ; Γ)) shown in 15) is calculated (step S45). Then, the generation model learning unit 21 outputs the generation model parameter vector set Θ ^{(t + 1)} to the convergence determination unit 23 as a result of the processing (step S46).

<Model set parameter set estimation process>
FIG. 17 is a flowchart showing the model integration parameter set estimation processing shown in FIG. Model integrated learning unit 22 of the language analysis model learning device 2, the output candidate graph generation unit 221 generates an output candidate graph from there Labels input data D _l (step S61). Then, the model integrated learning means 22 uses the feature extraction means 222 to extract features from the output candidate graph by combining the estimated position label and the instance in the input sequence described in the identification model feature extraction template ( Step S62). Then, the model integration learning unit 22 inputs the model integration parameter set Γ ^(t) to the objective function calculation unit 223 (step S63). Next, the model integrated learning means 22 uses the objective function calculation means 223 to calculate the objective function L ^SS-Hyb (Γ | Θ) shown in the above equation (16) (step S64), and the partial differential calculation for the discrimination model The means 224 calculates the partial differentiation based on the parameter γ _i of the identification model based on the above equation (17) (step S65), and the generation model partial differentiation calculation means 225 based on the above equation (18). Then, the partial differentiation by the parameter γ _j of the generation model is calculated (step S66). Then, the model integrated learning unit 22 outputs the model integration parameter set Γ ^{(t + 1)} to the convergence determination unit 23 as a result of the processing (step S67).

  The language analysis model learning device 2 can also be realized by executing a language analysis model learning program that causes a general computer to execute each step described above. This program can be distributed via a communication line, or can be written on a recording medium such as a CD-ROM for distribution.

According to the language analysis model learning device 2 of the present embodiment, to determine the parameter vector set Θ for generating model utilizing unlabeled data D _u, with parameter vector set Θ and the label determined generated model data D _l preparative by determining the model integration parameter set Γ using the unlabeled data D _u captured by generation approach may be learned by the identification approach. Therefore, it is possible to learn the structure predictor using the unlabeled data _Du that is relatively easy to acquire. As a result, if the amount of the labeled data D _l is the same, it is possible to learn a structure predictor with better performance than the conventional conditional random field learning method. Furthermore, when the label has the data D _l of a domain does not exist, the structure predictor learned using a label with data D _l of the task-specific domain, a learning by using the unlabeled data D _u target domain Can be done. That is, conventionally, the domain prediction is difficult when the label has the data D _l is not present, only the cost of acquiring the unlabeled data D _u, can create a high-performance structural predictor It becomes.

  As mentioned above, although embodiment of this invention was described, this invention is not limited to this, It can implement in the range which does not change the meaning. For example, in the present embodiment, the labeled data used by the identification model learning unit 10 and the labeled data used by the model integrated learning unit 22 have been described as different, but the present invention is not limited thereto. It may be the same. However, since it has been experimentally known that it is often possible to construct a language analysis model having high compatibility when using data with different labels in each processing unit, the identification model learning unit 10 and the model integrated learning unit 22 It is preferable that the data with labels used in and are different.

In the present embodiment, the feature extraction template for the identification model and the feature extraction template for the generation model have been described as being manually determined depending on the target problem. However, the feature extraction template for the generation model is created in advance so as to deal with a plurality of problems. A plurality of templates may be prepared, and a template may be automatically selected by inputting a target problem.
In the present embodiment, the language analysis model learning device 2 has been described as having the best mode configuration including the identification model learning means 10, but is configured so that a previously learned identification model parameter vector set Λ can be used. If so, the identification model learning means 10 may not be provided.
In the present embodiment, the parameter integration device 3 has been described as being provided separately from the language analysis model learning device 2, but may be configured to be included in the language analysis model learning device 2.

It is a block diagram which shows typically the outline | summary of the language analysis model production apparatus which concerns on embodiment of this invention. It is a functional block diagram which shows typically the structure of the language analysis model learning apparatus shown in FIG. It is a figure which shows the example of the information input into the linguistic analysis model learning apparatus shown in FIG. 2, Comprising: (a) has shown data with a label, (b) has each shown the output label candidate set. It is a figure which shows typically an example of the output candidate graph produced | generated with the language analysis model learning apparatus shown in FIG. It is a figure which shows the example of the feature extraction template for identification models. It is a figure which shows the example of the feature extracted with the feature extraction template for identification models shown in FIG. FIG. 6 is an explanatory diagram of a feature vector created using the identification model feature extraction template shown in FIG. 5 for the nodes indicated by circles in FIG. 4. It is a figure which shows an example of the parameter vector for identification models output from the identification model learning means shown in FIG. It is a functional block diagram which shows typically the structure of the sequence structure predictor learning means shown in FIG. It is a figure which shows the example of the feature extraction template for generation | occurrence | production models. It is a figure which shows the example of the feature extracted by the feature extraction template for generation models shown in FIG. It is a figure which shows an example of the parameter vector for production | generation models output from the sequence structure predictor learning means shown in FIG. It is a figure which shows an example of the parameter set for model integration output from the sequence structure predictor learning means shown in FIG. It is a flowchart which shows operation | movement of the language analysis model creation apparatus shown in FIG. It is a flowchart which shows the supervised learning process shown in FIG. It is a flowchart which shows the parameter vector set estimation process for generation | occurrence | production models shown in FIG. It is a flowchart which shows the parameter set estimation process for model integration shown in FIG. It is a figure which shows typically the example of the input series and output series which are handled with a sequence structure prediction problem.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Language analysis model creation apparatus 2 Language analysis model learning apparatus 10 Discrimination model learning means 11 Output candidate graph generation means 12 Feature extraction means 13 Parameter learning means 131 Objective function calculation means 132 Objective function gradient calculation means 133 Convergence determination means 134 Parameter update means 134 20 Sequence structure predictor learning means 21 Generation model learning means 211 Output candidate graph generation means 212 Feature extraction means 213 Objective function calculation means 214 Auxiliary function calculation means 215 Parameter update means 22 Model integrated learning means 221 Output candidate graph generation means 222 Feature extraction Means 223 Objective function calculation means 224 Identification model partial differential calculation means 225 Generation model partial differential calculation means 226 Parameter update means 23 Convergence determination means 3 Parameter integration device (parameter integration means)
4 learning support information storage means 41 identification model feature extraction template storage means 42 generation model feature extraction template storage means 43 output label candidate storage means 5 parameter set storage means 51 identification model parameter vector set storage means 52 generation model parameter vector Set storage means 53 Parameter set storage means for model integration 6 Language analysis model storage means 7 Sequence structure prediction device

Claims (12)

Based on the identification model and the generation model, the character string or symbol string based on the identification model and the generation model with the labeled data indicating the data with the label attached to the character string or symbol string and the unlabeled data indicating the character string or symbol string as input data A language analysis model learning device for learning a language analysis model used for estimating a label to be assigned to
The identification model is a model for estimating the label to be given using a conditional probability indicating a probability that a label candidate predetermined in the condition of an input character string or symbol string appears,
The generation model is a model for estimating the label to be assigned using a joint probability indicating a probability that an input character string or symbol string and the predetermined label candidate are generated simultaneously;
Using the unlabeled data the input, advance and parameter vector set for learning identification model, using the model integration parameter set predetermined for Generating Models that maximize the first objective function A generation model learning means for determining a parameter vector set;
Using the input labeled data , the second objective function is obtained using the previously learned identification model parameter vector set and the generated model parameter vector set determined by the generated model learning means. and a model integrated learning means for determining a pre-SL model integration parameter set you maximized,
Alternately executing the process of determining the generated model parameter vector set by the generated model learning means and the process of determining the model integration parameter set by the model integrated learning means,
One of said generated model learning means and the model integration learning means the model integration parameter set and the parameter vector set for the product model determined alternately, it is determined whether or not a predetermined convergence condition is satisfied, the Convergence determining means for outputting the generated model parameter vector set and the model integration parameter set at that time when it is determined that the convergence condition is satisfied ,
The first objective function uses the identification model parameter vector set, the generated model parameter vector set, and the model integration parameter set to identify all outputs when unlabeled data is given. A function that calculates the sum of output values of a function
The second objective function is a function that calculates a degree to which labeled data can be correctly identified using the identification model parameter vector set, the generated model parameter vector set, and the model integration parameter set. ,
A language analysis model learning device characterized by that. The generation model learning means includes:
Objective function calculating means for calculating an objective function G (Θ | Γ) shown in the following equation (14) as the first objective function;
Auxiliary function calculation means for performing processing for obtaining a parameter vector set Θ that maximizes an objective function G (Θ | Γ) represented by the following equation (14) under fixed Λ and Γ;
A parameter updating unit that updates the obtained generation model parameter vector set Θ when the convergence determining unit determines that the convergence condition is not satisfied, and
The model integrated learning means includes:
The objective function L shown in the following equation (16) as the second objective function ^SS-Hyb An objective function calculating means for calculating (Γ | Θ) under fixed Λ and Θ,
The objective function L shown in the following equation (16) ^SS-Hyb (Γ | Θ) is the model integration parameter γ for the discrimination model _i Partial differentiation calculation means for identification model for performing partial differentiation at
The objective function L shown in the following equation (16) ^SS-Hyb Model integration parameter γ for generating model (Γ | Θ) _j A partial differential calculation means for a generation model that performs a partial differential calculation at
Parameter updating means for updating the obtained model integration parameter set Γ when the convergence determining means determines that the convergence condition is not satisfied.
The language analysis model learning device according to claim 1.

here,
Labeled data is labeled sample = (x ⁿ , Y ⁿ ) N sets of
Unlabeled data is unlabeled sample = (x ^m ) Shows M sets,
Λ represents an identification model parameter vector set represented by the following equation (3):
Θ represents a generation model parameter vector set represented by the following equation (4):
Γ represents a parameter set for model integration represented by the following formula (5),
p _i ^D Indicates the joint probability estimated from the discrimination model,
p _j ^G Indicates the joint probability estimated from the generated model,
p (Θ) represents the prior probability distribution for Θ,
p (Γ) represents a prior probability distribution for Γ.

An auxiliary function for maximizing the first objective function and the second objective function are:
The joint probability of the input sequence and the output sequence estimated from the identification model parameter vector set using the labeled data, and the model integration parameter set obtained in advance for the identification model parameter vector set The probability value calculated based on the identification model integration probability value indicating the result of integrating over the identification model parameter vector set to be integrated, and
The joint probability of the input sequence and the output sequence estimated from the generation model parameter vector set using the unlabeled data, and the model integration parameter set obtained in advance for the generation model parameter vector set The product of the probability value calculated based on the generation model integration probability value indicating the result of integrating the probability value calculated over the generation model parameter vector set to be integrated,
I viewed including as a parameter set indicating a posteriori probability of the label to be assigned to a character string or symbol string is the input,
The auxiliary function calculating means includes:
Using the Q function shown in the following equation (15) as the auxiliary function, a parameter vector set Θ ′ that maximizes the Q function is obtained from the current parameter vector set Θ, and Θ ′ increases with respect to Θ. by obtaining the repetition Q function while replacing the theta until no in theta ', the objective function G | claim 2, characterized in <br/> obtaining parameters vector set theta to maximize (theta gamma) Language analysis model learning device.

The apparatus further comprises parameter integration means for integrating the output generation model parameter vector set and the model integration parameter set, and the previously learned identification model parameter vector set into a parameter set indicating the posterior probability. The language analysis model learning device according to claim 3 . Any of claims 1 to 4, characterized by further comprising an identification model learning unit for creating the parameter vector set identifying model by learning the label has data the input by using the identification model The language analysis model learning device according to one item. Based on the identification model and the generation model, the character string or symbol string based on the identification model and the generation model with the labeled data indicating the data with the label attached to the character string or symbol string and the unlabeled data indicating the character string or symbol string A language analysis model learning method of a language analysis model learning device for learning a language analysis model used for estimating a label to be given to
The identification model is a model for estimating the label to be given using a conditional probability indicating a probability that a label candidate predetermined in the condition of an input character string or symbol string appears,
The generation model is a model for estimating the label to be assigned using a joint probability indicating a probability that an input character string or symbol string and the predetermined label candidate are generated simultaneously;
Maximum by generation model learning unit, using the unlabeled data the input, and parameter vector set for pre-learning identification model, using the model integration parameter set predetermined first objective function determining a parameter vector set for generate models that turn into,
The model integrated learning unit, using a located labels that are the input data, the use and parameter vector set in advance for learning identification model, and a parameter vector set for generating model determined by the generation model learning means and determining the pre-SL model integration parameter set that maximize the second objective function,
Alternately
The convergence judging means, one of said model integration parameter set and the generated model parameter vector set for learning means and said generating models model integrated learning means has determined alternately whether a predetermined convergence condition is satisfied It determines the when it is determined that the convergence condition is satisfied, saw including a step of outputting said model integration parameter set and the parameter vector set for generating models that point,
The first objective function uses the identification model parameter vector set, the generated model parameter vector set, and the model integration parameter set to identify all outputs when unlabeled data is given. A function that calculates the sum of output values of a function
The second objective function is a function that calculates a degree to which labeled data can be correctly identified using the identification model parameter vector set, the generated model parameter vector set, and the model integration parameter set. ,
A language analysis model learning method characterized by this. Determining the generation model parameter vector set;
Calculating an objective function G (Θ | Γ) shown in the following equation (14) as the first objective function;
By calculating a predetermined auxiliary function, a process for obtaining a parameter vector set Θ that maximizes the objective function G (Θ | Γ) expressed by the following equation (14) under fixed Λ and Γ Steps to perform,
Updating the obtained generation model parameter vector set Θ when the convergence determining means determines that the convergence condition is not satisfied, and
The step of determining the parameter set for model integration includes:
The objective function L shown in the following equation (16) as the second objective function ^SS-Hyb Calculating (Γ | Θ) under fixed Λ and Θ;
The objective function L shown in the following equation (16) ^SS-Hyb (Γ | Θ) is the model integration parameter γ for the discrimination model _i Performing a partial differentiation with
The objective function L shown in the following equation (16) ^SS-Hyb Model integration parameter γ for generating model (Γ | Θ) _j Performing a partial differentiation with
Updating the obtained model integration parameter set Γ when it is determined by the convergence determination means that the convergence condition is not satisfied.
The language analysis model learning method according to claim 6.

here,
Labeled data is labeled sample = (x ⁿ , Y ⁿ ) N sets of
Unlabeled data is unlabeled sample = (x ^m ) Shows M sets,
Λ represents an identification model parameter vector set represented by the following equation (3):
Θ represents a generation model parameter vector set represented by the following equation (4):
Γ represents a parameter set for model integration represented by the following formula (5),
p _i ^D Indicates the joint probability estimated from the discrimination model,
p _j ^G Indicates the joint probability estimated from the generated model,
p (Θ) represents the prior probability distribution for Θ,
p (Γ) represents a prior probability distribution for Γ.

An auxiliary function for maximizing the first objective function and the second objective function are:
The joint probability of the input sequence and the output sequence estimated from the identification model parameter vector set using the labeled data, and the model integration parameter set obtained in advance for the identification model parameter vector set The probability value calculated based on the identification model integration probability value indicating the result of integrating over the identification model parameter vector set to be integrated, and
The joint probability of the input sequence and the output sequence estimated from the generation model parameter vector set using the unlabeled data, and the model integration parameter set obtained in advance for the generation model parameter vector set The product of the probability value calculated based on the generation model integration probability value indicating the result of integrating the probability value calculated over the generation model parameter vector set to be integrated,
I viewed including as a parameter set indicating a posteriori probability of the label to be assigned to a character string or symbol string is the input,
The step of obtaining the parameter vector set Θ is:
Using the Q function shown in the following equation (15) as the auxiliary function, a parameter vector set Θ ′ that maximizes the Q function is obtained from the current parameter vector set Θ, and Θ ′ increases with respect to Θ. by obtaining the repetition Q function while replacing the theta until no in theta ', the objective function G | claim 7, wherein <br/> obtaining parameters vector set theta to maximize (theta gamma) Language analysis model learning method.

A step of integrating the output generation model parameter vector set and the model integration parameter set, and the previously learned identification model parameter vector set into a parameter set indicating the posterior probability by parameter integration means; The language analysis model learning method according to claim 8 , comprising: By identifying the model learning unit, the claims 6 to 9, characterized by the step of creating a parameter vector set for the identification model by learning the label has data the input by using the identification model The language analysis model learning method according to any one of the above. A language analysis model learning program for causing a computer to execute the language analysis model learning method according to any one of claims 6 to 10 . A computer-readable recording medium in which the language analysis model learning program according to claim 11 is recorded.

Images