JP5295037B2 - Learning device using Conditional Random Fields or Global Conditional Log-linearModels, and parameter learning method and program in the learning device - Google Patents

Description

  The present invention relates to Conditional Random Fields and Global Conditional Log-linear Models in the field of machine learning, and more particularly to learning of parameters thereof.

  Conditional Random Fields (Conditional Random Fields: CRFs) and Global Conditional Log-linear Models (GCLMs) are learning machines (learning devices) for determining a symbol sequence that is predicted to be most appropriate from among symbol sequences. Applications range widely from natural language processing to labeling, morphological analysis, speech recognition, and error correction in machine translation. There are some learning machines that can be applied to similar applications, but CRFs and GCLMs have advantages such as guaranteeing convergence to an optimal solution based on the concave nature of the objective function, and the most accurate model is available. One of the learning machines to generate.

  Non-Patent Document 1 describes an example in which CRFs are applied to Japanese morphological analysis, and Non-Patent Document 2 describes an example in which GCLMs is applied to an error correction language model for speech recognition. All of them achieve high-accuracy model generation compared to alternative methods.

Taku Kudo, Atsushi Yamamoto, Yuji Matsumoto, "Japanese Morphological Analysis Using Conditional Random Fields", IPSJ Research Report 2004-NL-161, Vol.2004, No.47, pp.89-96 (2004) Roark B., Saraclar M. and Collins M. “Discriminative n-gram language modeling”, Computer Speech and Language, Vol.21, No.2, pp.373-392 (2007)

  By the way, when using a learning machine for determining a symbol sequence that is predicted to be most appropriate from symbol sequences, it is necessary to perform model learning (model parameter learning) in advance using learning data. The learning data is a collection of a large number of lists each consisting of a set of a plurality of symbol sequences and corresponding correct symbol sequences. However, in some cases, the symbol sequence weight (importance) of each symbol sequence may be given.

For example, in the error correction language model for speech recognition, the following reference describes that the word error rate of a symbol series in a list corresponding to a correct word string is given as a symbol series weight.
References: Akio Kobayashi and 5 others, “Speech recognition by discriminative scoring of word lattice”, Proc. Of the Acoustical Society of Japan, pp.233-234 (2007)
Since the symbol sequence weight is auxiliary information that aids artificially given model learning, learning using this leads to generation of a highly accurate model. Therefore, when a symbol sequence weight is given, it is necessary to select a learning machine having a framework that handles the symbol sequence weight.

  However, conventional Conditional Random Fields (CRFs) and Global Conditional Log-linear Models (GCLMs) are learning machines that do not have a framework for handling symbol sequence weights.

  In view of such a situation, an object of the present invention is to provide a learning apparatus using Conditional Random Fields or Global Conditional Log-linear Models that has a framework for handling symbol sequence weights, its parameter learning method, and a program There is to do.

According to the present invention, a learning apparatus using Conditional Random Fields or Global Conditional Log-linear Models includes a plurality of symbol sequence feature vectors, a corresponding correct symbol sequence feature vector, and a plurality of symbol sequence feature vectors. A list input unit for acquiring a set of lists including symbol sequence weights and symbol sequence weights of feature vectors of correct symbol sequences as learning data, a parameter initialization unit for initializing parameters of an objective function, a plurality of symbol sequences and correct answers for each feature vector of the symbol sequence, the parameters and the by the inner product of the feature vector to calculate the linear score, its linear score and a list in the processing unit to output calculated a weighted index score and a symbol sequence weights of the feature vector When, all of the index score calculated in a list in the processing unit An objective function calculation unit that calculates the objective function and its gradient using the feature vectors of the plurality of symbol sequences and the feature vector of the correct symbol sequence; a convergence determination unit that determines convergence of the objective function from the gradient; And a parameter updating unit for updating the parameter.

  In the above configuration, the in-list processing unit preferably includes a linear score calculation unit that calculates the linear score, an exponent score calculation unit that calculates an exponent score from the linear score calculated by the linear score calculation unit, and the exponent score A weight multiplier for calculating the weighted exponent score by multiplying the exponent score calculated by the calculator by the symbol sequence weight.

  Further, the list processing unit includes a linear score calculation unit that calculates the linear score, and a weight addition that calculates a weighted linear score by adding the symbol series weight to the linear score calculated by the linear score calculation unit And an exponent score calculation unit that calculates the weighted exponent score from the weighted linear score calculated by the weight addition unit.

  A parameter learning method in a learning apparatus using Conditional Random Fields or Global Conditional Log-linear Models according to the present invention includes a feature vector of a plurality of symbol sequences, a feature vector of a correct symbol sequence corresponding thereto, and a symbol sequence of each symbol sequence A linear score is calculated by a list input process that takes a set of lists including weights as learning data, a parameter initialization process that initializes parameters of an objective function, and an inner product of the parameter and the feature vector, and the linear score And an index score weighted from the symbol series weight for each symbol series, and the objective function using all the exponent scores and the feature vectors calculated in the list processing process. And an objective function calculation process for calculating the inclination thereof, A convergence determination process for determining the convergence of the objective function from the slope, and a parameter update process for updating the parameter are included.

  In the above, the list processing process preferably includes a linear score calculation process for calculating the linear score, an exponent score calculation process for calculating an exponent score from the linear score calculated in the linear score calculation process, and an exponent score calculation The exponent score calculated in the process is multiplied by the symbol sequence weight to calculate the weighted exponent score.

  Further, the in-list processing step includes a linear score calculation step for calculating the linear score, and a weight addition for calculating a weighted linear score by adding the symbol sequence weight to the linear score calculated in the linear score calculation step. It is also possible to comprise a process and an exponent score calculation process for calculating the weighted exponent score from the weighted linear score calculated in the weight addition process.

  According to the present invention, in a learning device using Conditional Random Fields or Global Conditional Log-linear Models that guarantees convergence to an optimal solution, a learning device having a framework for handling symbol sequence weights and its parameter learning method can be realized. it can.

  Therefore, according to the present invention, learning accuracy can be improved and model performance can be improved as compared to learning without symbol sequence weight.

The block diagram which shows the function structural example of the existing learning apparatus which uses CRFs and GCLMs. The flowchart for demonstrating the procedure of the parameter learning method in the learning apparatus shown in FIG. The block diagram which shows the function structure of one Example of the learning apparatus using CRFs and GCLMs by this invention. The flowchart for demonstrating the procedure of the parameter learning method in the learning apparatus shown in FIG. The block diagram which shows the function structure of the other Example of the learning apparatus using CRFs and GCLMs by this invention. The flowchart for demonstrating the procedure of the parameter learning method in the learning apparatus shown in FIG.

  First, existing Conditional Random Fields (CRFs) and Global Conditional Log-linear Models (GCLMs) will be described.

Although the names differ depending on the conditions, these CRFs and GCLMs are learning machines in which learning is achieved by obtaining a parameter (parameter vector) w ^→ that minimizes the next objective function L with learning data given.

In learning, each symbol series is represented by a feature vector. f ^→ _{i, 0} is a feature vector of the correct symbol sequence of the i-th list (symbol sequence set), and f ^→ _{i, j} is a feature vector of the j-th symbol sequence belonging to the i-th list. The correct answer may or may not be included in the list. <W ^→ , f ^→ > represents an inner product of the parameter w ^→ and the feature vector f ^→ .

  When learning is performed using Expression (1), although the performance with respect to the learning data is high, a model with low performance with respect to different data may be generated due to over-learning. In order to prevent this, parameter estimation in CRFs and GCLMs is generally formulated as a minimization of an objective function with regularization, such as:

Although w ^→ which minimizes the objective function can be obtained by a method based on the quasi-Newton method, this explanation is out of the scope of the present invention, and is omitted here. In general, the function that has the form of equation (1), is known to be against w ^→ is concave. Therefore, convergence to a global optimal solution is guaranteed.

  FIG. 1 shows a basic functional configuration example of a learning device using existing CRFs and GCLMs. The learning device includes a list input unit 11, a condition input unit 12, and a parameter estimation unit 20, and includes a parameter estimation unit. 20 includes a parameter initialization unit 21, an in-list processing unit 22, an objective function calculation unit 22, a convergence determination unit 24, and a parameter update unit 25. In this example, the list processing unit 22 includes a linear score calculation unit 22a and an exponent score calculation unit 22b.

FIG. 2 shows a basic procedure of parameter learning in the learning apparatus shown in FIG. 1, and the processing and procedure of each part in parameter learning will be described below with reference to FIGS.
・ List input (step S1)
List input section 11 takes in the feature vectors of a plurality of symbol sequences, a set of list of the feature vector of the correct symbol sequence corresponding thereto as the learning data, all i, for j f ^→ _{i, 0} and f ^→ _{i , j} is entered.
-Parameter initialization (step S2)
The parameter initialization unit 21 initializes the parameter w ^→ of the objective function L shown in Expression (2).
-Linear score calculation (step S3)
The linear score calculation unit 22a of the in-list processing unit 22 calculates a linear score <w ^→ , f ^→ > by an inner product of the parameter w ^→ and the feature vector f ^→ .
-Index score calculation (step S4)
The exponent score calculation unit 22b calculates an exponent score exp (<w ^→ , f ^→ >) from the linear score <w ^→ , f ^→ >. In each i, index score exp of the correct symbol sequence _{^{(<w →, f → i}} , 0>) including the ^{{exp (<w →, f} → i, j>) | j = 0,1, ..., n i } Is calculated.
-Objective function and its slope calculation (step S5)
The objective function calculation unit 23 calculates the objective function L of equation (2) and its gradient using all the exponent scores and feature vectors calculated by the in-list processing unit 22. The slope is

It is represented by
・ Convergence determination (step S6)
The convergence determination unit 24 determines the convergence of the objective function L from the slope shown in Expression (3).
-Parameter update (step S7)
The parameter update unit 25 updates the parameter w ^→ when the convergence determination unit 24 determines that the convergence has not been completed.

Thereafter, steps S3 to S7 are repeatedly executed until the optimal solution is converged. The parameter learning is completed by convergence to the optimal solution, and the optimal parameter w ^→ is estimated. The convergence determination condition, the parameter update condition, the constant C in equation (2), and the like are input from the condition input unit 12 to the parameter estimation unit 20.

  Next, an embodiment of the present invention will be described with reference to the drawings based on the above-described learning device using CRFs and GCLMs and its parameter learning method. In addition, in each figure, the same code | symbol is attached | subjected to the part corresponding to FIG.1 and 2, and the detailed description is abbreviate | omitted.

  FIG. 3 shows the configuration of Embodiment 1 of the learning apparatus using CRFs and GCLMs according to the present invention, and FIG. 4 shows the procedure of parameter learning in the learning apparatus shown in FIG.

In this example, a weighted list is input to the list input unit 11 (step S1), that is, a symbol sequence weight is added to the input, and feature vectors f ^→ _{i, 0} , f ^→ _i, for all i, j _{. j} and symbol sequence weights s _{i, 0} , s _{i, j} are input.

The in-list processing unit 22 includes a linear score calculation unit 22a, an exponent score calculation unit 22b, and a weight multiplication unit 22c. The weight multiplication unit 22c calculates the exponent score exp (<w ^→ , f ^→) calculated by the exponent score calculation unit 22b. >) Is multiplied by a symbol sequence weight s (step S11), and a weighted exponent score s exp (<w ^→ , f ^→ >) is calculated. Thereby, in this example, it is possible to realize a learning apparatus in which symbol sequence weights are introduced into CRFs and GCLMs, and a parameter learning method in the learning apparatus.

  FIG. 5 shows the configuration of Embodiment 2 of the learning apparatus using CRFs and GCLMs according to the present invention, and FIG. 6 shows the parameter learning procedure in the learning apparatus shown in FIG.

  As in the first embodiment, a weighted list is input to the list input unit 11 (step S1).

In this example, the list processing unit 22 includes a linear score calculation unit 22a, a weight addition unit 22d, and an exponent score calculation unit 22b. The weight addition unit 22d adds the symbol sequence weight s to the linear score <w ^→ , f ^→ > calculated by the linear score calculation unit 22a (step S21), and the weighted linear score <w ^→ , f ^→ >. + S is calculated. The weighted linear score <w ^→ , f ^→ > + s is passed to the exponent score calculation unit 22b, and the exponent score calculation unit 22b calculates the weighted exponent score exp (<w ^→ , f ^→ > + s). As a result, in this example as well as in the first embodiment, it is possible to realize a learning device in which symbol sequence weights are introduced into CRFs and GCLMs and a parameter learning method in the learning device.

  In the above description, the list (weighted list) input to the list input unit 11 includes a network representation.

Hereinafter, it will be described that the introduction (multiplication, addition) of the symbol sequence weight s in the first and second embodiments is equivalent and that the objective function L in the present invention is also concave.
[Introduction of two equivalent symbol sequence weights]
The objective function L in the first embodiment is as follows.

  On the other hand, the objective function L in the second embodiment is as follows.

  Here, using the fact that x = exp (log (x)) and exp (x) exp (y) = exp (x + y), Expression (4) can be converted into the following expression.

Equations (5) and (6) show that s has only been converted to log (s). Since the logarithm is a monotonically increasing function, the two are common in that they are functions designed to increase the influence of a symbol sequence with a large weight, although the difference in scaling is, the symbol sequence weight s plays a role Equivalent work from the point of view.
[Concave objective function]
In order to show that equation (4) is concave, we consider using equation (6), which is an equational transformation. Now, c and log (s _{i, j} ) are added as new elements to w ^→ and f ^→ _{i, j} , respectively. The added ones are written as W ^→ and F ^→ _{i, j} , respectively. At this time, the equation (6) is

It becomes. In this equation, c = 1 is implicitly requested. However, if c is also an estimation target, it is known that the function of this shape is concave as described above. However, W ^→ is concave. Learning in CRFs and GCLMs is a minimization of the objective function parameter w 1 ^→ . Therefore, the interest is whether equation (7) is concave with respect to w ^→ . However, L being concave with respect to W ^→ means concave with respect to any W ^→ elements w _k and c. In other words, it is kept concave with respect to w ^→ .
It can be seen that equation (5) is also concave with respect to the parameter w ^→ for the same reason.

  The learning device and the parameter learning method in the learning device according to the present invention described above can be realized by a computer and a parameter learning program installed in the computer.

《How to use the model obtained by learning》
When obtaining the optimum symbol sequence from the list, the symbol sequence having the largest inner product <w ^→ , f ^→ > of the parameter w ^→ obtained by learning and the feature vector f ^→ of each symbol series is selected.
<Verification>
The effect of the present invention was verified using a Japanese spoken corpus (CSJ). CSJ is a database consisting of speech data and transcripts. Two evaluation sets were prepared for learning as shown in the table below.

  The lecture was divided into utterance units, and a 5000-best list was created by the speech recognition system. That is, the number of lists matches the number of utterances. The symbol series is a speech recognition result, and there are a maximum of 5000 symbol series in each list. Uni-, bi-, tri-gram boolean and speech recognition score were used for the features. For the symbol sequence weight (importance), the rank in the list of each symbol sequence (ascending order of word error rate) was used. The word error rate in the table is calculated for the recognition result having the largest recognition score in the 5000-best list output by the speech recognition system.

By rearranging the symbol sequences in descending order of <w ^→ , f ^→ >, the symbol sequence having the highest score is finally set as a new speech recognition result, and the word error rates are compared. That is, the purpose of model learning is to give a high score to a symbol series having a low word error rate. The results were as follows.

  The model performance was improved by the effect of weighted learning, and a lower word error rate could be realized.

Claims (7)

A learning device that uses Conditional Random Fields or Global Conditional Log-linear Models,
A set of lists comprising a feature vector of a plurality of symbol sequences, a feature vector of a corresponding correct symbol sequence, a symbol sequence weight of the feature vector of the plurality of symbol sequences, and a symbol sequence weight of the feature vector of the correct symbol sequence A list input unit that captures as learning data,
A parameter initialization unit for initializing parameters of the objective function;
For each feature vector of said plurality of symbol sequences and the correct symbol sequence, the parameters and the by the inner product of the feature vector to calculate the linear scores, exponent score weighting from the symbol sequence weight of the linear score and the feature vector and the list in the processing unit that de San a,
And the objective function calculation section that calculates the objective function and its gradient by using the feature vector of the feature vector and the correct symbol sequence with all index score calculated by the list in the processing unit of the plurality of symbol sequences,
A convergence determination unit for determining convergence of the objective function from the slope;
A learning apparatus comprising: a parameter updating unit that updates the parameter. In the learning apparatus using Conditional Random Fields or Global Conditional Log-linear Models according to claim 1,
The in-list processing unit is calculated by a linear score calculation unit that calculates the linear score, an exponent score calculation unit that calculates an exponent score from the linear score calculated by the linear score calculation unit, and an exponent score calculation unit. And a weight multiplication unit for calculating the weighted exponent score by multiplying the exponent score by the symbol series weight. In the learning apparatus using Conditional Random Fields or Global Conditional Log-linear Models according to claim 1,
The list in the processing unit, the weight for calculating the linear score calculator for calculating the linear score, the linear scores weighted by adding each of the symbol sequences weight linear score calculated in its linear score calculator A learning apparatus comprising: an addition unit; and an exponent score calculation unit that calculates the weighted exponent score from the weighted linear score calculated by the weight addition unit. A parameter learning method in a learning apparatus using Conditional Random Fields or Global Conditional Log-linear Models,
A set of lists comprising a feature vector of a plurality of symbol sequences, a feature vector of a corresponding correct symbol sequence, a symbol sequence weight of the feature vector of the plurality of symbol sequences, and a symbol sequence weight of the feature vector of the correct symbol sequence List input process to import as learning data,
A parameter initialization process for initializing the parameters of the objective function;
For each feature vector of said plurality of symbol sequences and the correct symbol sequence, the parameters and the by the inner product of the feature vector to calculate the linear scores, exponent score weighting from the symbol sequence weight of the linear score and the feature vector and the list in the process of de San a,
And the objective function calculation step of calculating the objective function and its gradient by using the feature vector of the feature vector and the correct symbol sequence with all index score calculated by the list within the process the plurality of symbol sequences,
A convergence determination process for determining convergence of the objective function from the slope;
A parameter updating process for updating the parameter. In the parameter learning method in the learning apparatus using Conditional Random Fields or Global Conditional Log-linear Models according to claim 4,
The in-list processing process is calculated in a linear score calculation process for calculating the linear score, an exponent score calculation process for calculating an exponent score from the linear score calculated in the linear score calculation process, and an exponent score calculation process. And a weight multiplication process for calculating the weighted exponent score by multiplying the symbol score by the symbol sequence weight. In the parameter learning method in the learning apparatus using Conditional Random Fields or Global Conditional Log-linear Models according to claim 4,
The in-list processing step includes a linear score calculation step for calculating the linear score, and a weight addition step for calculating a weighted linear score by adding the symbol sequence weight to the linear score calculated in the linear score calculation step. And an exponent score calculation step of calculating the weighted exponent score from the weighted linear score calculated in the weight addition step.   A parameter learning program for causing a computer to function as a learning device using the Conditional Random Fields or Global Conditional Log-linear Models according to any one of claims 1 to 3.

Images