JP5264649B2 - Information compression model parameter estimation apparatus, method and program - Google Patents

Description

  The present invention relates to an information compression model parameter estimation apparatus, method, and program used for model learning in a symbol sequence rearrangement problem.

  In speech recognition and machine translation, multiple provisional recognition results and translation results (word sequences) are output, and by finding a sequence with few errors (close to the correct answer), recognition and translation accuracy are improved. be able to. When a word sequence of individual correct candidates output by a speech recognizer or machine translator is called a symbol series, and a set of output correct candidates is called a list, the extraction of the correct symbol series from such a list is generally performed. A score is assigned to each symbol series, and the symbol series in the list is rearranged in the order of score. In other words, the word sequence with the highest score is usually the recognition / translation result, and even if it is not, it is possible to efficiently extract the result of the symbol sequence close to the correct answer by sequentially verifying the symbol sequence with the highest score. (See Non-Patent Documents 1, 2 and 6 for speech recognition, and Non-Patent Documents 3 and 4 for machine translation).

  When a target symbol series is extracted from a list of symbol series, a model obtained by learning in advance is generally used. Hereinafter, a method of finding a sequence close to the correct answer using a model prepared in advance will be described with reference to FIG.

First, a list consisting of a plurality of symbol sequences is read (S11). Each symbol series is generally represented by a feature (feature) vector, and the feature is an N-gram or co-occurrence of words, parts of speech, phonemes, etc., frequency of dependency obtained from the result of applying parsing or dependency analysis, boolean (A binary representation of presence / absence) is used. However, the form of the list is not necessarily limited to a column of feature vectors, and any form that can finally extract a feature vector may be used even if it is an expression form such as a network. Note that the symbol sequence can be expressed by a feature vector by the following method (see Non-Patent Document 3). For example, consider a method of representing a symbol sequence OOXX with a symbol set {◯, χ, △} by a feature vector. If a feature value is 1 when a symbol appears in the symbol series and 0 when it does not appear, 1 and △ do not appear in the symbol series XXX because O and X appear. So it becomes 0. The feature vector is a vector representation of such a feature as [1, 1, 0] ^T. When a natural language word string is handled as a symbol series, additional information such as a syntax analysis result of each symbol series and its score is added, and a feature vector may be created including the information.

Next, referring to the model obtained by learning, a score corresponding to the symbol series is given (S12). There are various ways to calculate the score. When the vector w is a model parameter obtained by learning in advance _, the score S _w (f _{i, j} ) of the symbol sequence f _{i, j} represented by the feature vector is, for example, S _w (f _{i, j} ) = w ^T · f _{i, j} can be calculated (i is an index of a list (i = 1, 2,..., N), j is an index of a symbol sequence in each list i (j = 1, 2,. .., n _i )), T is the transpose of the matrix).

Then, by rearranging the symbol series f _{i, j} according to the assigned score, the symbol series in the list can be arranged in the order from the closest to the correct answer (S13).

  Further, a method for estimating the model parameter w used for score calculation will be described below with reference to FIG.

  First, a plurality of lists consisting of a plurality of symbol series are read (S21). As the number of lists to be read increases, it can be expected that model parameters that function with high accuracy can be obtained for various data. Also, the correct symbol series of each list is read together. However, the same symbol series as the correct symbol series may or may not be included in each list.

  Next, the model parameter w is estimated by learning based on the read information (S22). The parameter is estimated so that the correct symbol series is given a higher score than the other symbol series. That is, the model parameter w may be determined so as to reduce the number ErrorCount of symbol series to which a score higher than the score given to the correct symbol series is given. For example, w that minimizes the expression (1) is obtained.

Here, I (x) is a symbol included 0 when the value of x is positive, the function which gives 1 in other cases, f _{i, 0} is correct symbol sequence, N is the number of lists, the n _i list i The number of series. Non-Patent Document 5 discloses a method for determining a model parameter w by the Global Conditional Log-linear Model (GCLM) method. In this case, w for minimizing the value of L in Equation (2) can be obtained. That's fine.

Here, ‖w‖ is a norm, and it is known that a robust estimation result can be obtained by using this norm. C is a hyperparameter, which is determined using a development set or the like. According to Equation (2), it is guaranteed that the estimation result of the model parameter w converges to a global optimum solution. Specifically, the model parameter w can be estimated by a known method such as L-BFGS.

Of course, each symbol series f _{i, j} output from a speech recognizer or machine translator is usually assigned an importance e _{i, j} based on an arbitrary evaluation scale (for example, ranking in a list). Therefore, estimation accuracy can be improved by using this for parameter estimation. For example, in the case of the ExpLoss Boosting (ELBst) method disclosed in Non-Patent Document 3, w that minimizes the value of L in Equation (3) may be obtained.

In equation (3), there is an algorithm for efficiently estimating w, particularly when the feature value is binary of 0 and 1. Further, in the case of the Minimum Error Rate Training (MERT) method disclosed in Non-Patent Document 4, it is only necessary to obtain w that minimizes the value of L in Equation (4).

Here, α is a hyper parameter, which is determined using a development set or the like. According to Equation (4), the model parameter w can be estimated without using the correct symbol sequence. The estimation of the model parameter w can be specifically performed by a known method such as L-BFGS.

  The above model parameter estimation method is based on the learning method (batch type) that reads the entire learning data (= all lists) and performs overall optimization. There is also an online learning method that updates model parameters each time. However, in order to obtain a good parameter estimation result even in online learning, in general, all lists are recursively read a plurality of times. Since an ordinary computer takes time to input data, the entire computer is often read into a memory on the computer.

Z. Zhou, J. Gao, FKSoong, and H. Meng, "A Comparative Study of Discriminative Methods for Reranking LVCSR N-Best Hypotheses in Domain Adaptation and Generalization," Proceedings of ICASSP, 2006, Vol.1, p.141 -144 Akio Kobayashi, Shohei Sato, Kazuho Onoe, Shinichi Honma, Satoshi Imai, Toru Toki, “Speech Recognition by Discriminative Scoring of Word Lattice”, Proceedings of the Acoustical Society of Japan, September 2007, p.233-234 M. Collins and T. Koo, "Discriminative Reranking for Natural Language Parsing," Association for Computational Linguistics, 2005, Vol. 31, No. 1, p. 25-70 F.J.Och, "Minimum Error Rate Training in Statistical Machine Translation," Proceedings of the 41st Annual Meeting on Association for Computational Linguistics, 2003, p.160-167 B.Roark, M.Saraclar and M.Collins, "Discriminative n-gram language modeling," Computer Speech and Language, 2007, Vol.21, No.2, p.373-392 B. Roark, M. Saraclar and M. Collins, "Corrective Language Modeling For Large Vocabulary ASR with The Perceptron Algorithm," Association for Computational Linguistics, Proceedings of ICASSP, 2004, Vol.1, p.749-752

  To learn a model, it is necessary to prepare multiple lists, but a single list has many symbol sequences. Therefore, it is necessary to handle a huge number of symbol sequences as a whole. For example, when a word string is a symbol series as in speech recognition or machine translation, a large number of lists are required to generate a highly accurate model over a large amount of data. It is necessary to extract features. For example, in Non-Patent Document 5, learning is performed with approximately 280,000 lists each having 100 to 1000 symbol sequences. In this case, even if the memory area required to store the features (features) extracted from each symbol series is estimated to be as small as 100 bytes on average, a memory area of 1000 × 280,000 × 100 = 28 gigabytes is consumed. Since such a large work area (computer memory, etc.) is required, it is difficult to handle it with a general-purpose computer.

  The present invention provides an information compression model parameter estimation apparatus, method, and program capable of solving such a problem and performing model parameter estimation processing with estimation accuracy equivalent to that of a conventional computer using a general-purpose computer. Objective.

Information compression model parameter estimation apparatus according to an aspect of the present invention, respectively importance e _i, symbol sequence of n _i number expressed in feature vector _j is assigned f _i, consisting of _j, N number of list i ( i is the index of the list, n is the integer of 1 or more, j is the index of the symbol sequence for each i, n _i is the integer of 4 or more), the correct answer symbol sequence f _i for each list i which are respectively represented by a feature _{vector, 0} is an information compression type model parameter estimation apparatus that estimates the model parameter w, and includes a grouping unit, a merging unit, and a model parameter estimation unit.

For each list i, the grouping unit groups a plurality of symbol sequences f _{i, j} belonging to the list into a plurality of groups G _i (x) (x is a group index) by a predetermined method. Merging unit, the group G _i (x) belonging to a plurality of symbol sequences f _i, representatives from _j symbol sequence f _i, the _x, also the group G _i (x) belonging to a plurality of symbol sequences f _i, the _j The representative importance e _{i, x} is obtained from the corresponding importance e _{i, j} . The model parameter estimation unit estimates the model parameter w from the representative symbol series f _{i, x} , the correct symbol series f _{i, 0,} and the representative importance e _{i, x} .

The information compression type model parameter estimation device according to another aspect of the present invention has an importance e _{i, j} Is assigned and represented by a feature vector _i Symbol series f _{i, j} N lists i (where i is an index of the list, N is an integer of 1 or more, j is an index of a symbol sequence in each i, n _i Is an information compression type model parameter estimation device that estimates a model parameter w, and includes a grouping unit, a merging unit, and a model parameter estimation unit.
For each list i, the grouping unit includes a plurality of symbol sequences f belonging to the list. _{i, j} A plurality of groups G by a predetermined method _i Group into (x) (x is the index of the group). The merging part is group G _i A plurality of symbol sequences f belonging to (x) _{i, j} To representative symbol series f _{i, x} And Group G above _i A plurality of symbol sequences f belonging to (x) _{i, j} Multiple importance e corresponding to _{i, j} To representative importance e _{i, x} For each. The model parameter estimator receives the representative symbol sequence f _{i, x} And the above-mentioned representative importance e _{i, x} From these, the model parameter w is estimated.

According to another aspect of the present invention, there is provided an information compression type model parameter estimation apparatus in which n is represented by a feature vector. _i Symbol series f _{i, j} N lists i (where i is an index of the list, N is an integer of 1 or more, j is an index of a symbol sequence in each i, n _i Is an integer greater than or equal to 4) and the correct symbol sequence f of each list i represented by a feature vector. _{i, 0} Is an information compression model parameter estimation device that estimates the model parameter w, and includes a grouping unit, a merging unit, and a model parameter estimation unit.
For each list i, the grouping unit includes a plurality of symbol sequences f belonging to the list. _{i, j} A plurality of groups G by a predetermined method _i Group into (x) (x is the index of the group). The merging part is group G _i A plurality of symbol sequences f belonging to (x) _{i, j} To representative symbol series f _{i, x} Ask for. The model parameter estimator receives the representative symbol sequence f _{i, x} And the above correct symbol sequence f _{i, 0} From these, the model parameter w is estimated.

  According to the information compression model parameter estimation apparatus, method, and program of the present invention, since it is possible to compress symbol sequence information used for learning while ensuring the same estimation accuracy as in the past, model parameters can be estimated by a general-purpose computer. Processing can be performed.

The figure which shows the function structural example of the information compression type model parameter estimation apparatus 100. The figure which shows the example of a processing flow of the information compression type model parameter estimation apparatus. The figure which shows the content of each set for learning, development, and evaluation used for verification. The figure which shows the memory size required for data holding. The figure which shows the comparison verification result of the word error rate of this invention at the time of using ELBst method for estimation of a model parameter, and a prior art. The figure which shows the comparison verification result of the word error rate of this invention at the time of using GCLM method for estimation of a model parameter, and a prior art. The figure which shows the comparison verification result of the word error rate of this invention at the time of using MERT method for estimation of a model parameter, and a prior art. The figure which shows the example of the rearrangement process flow of a symbol series. The figure which shows the example of the processing flow of model learning.

FIG. 1 shows a functional configuration example of the information compression model parameter estimation apparatus 100 of the present invention, and FIG. 2 shows a processing flow example thereof. The information compression model parameter estimation apparatus 100 includes one or more lists i (i is an index of a list (i), each of which is composed of a plurality of symbol sequences f _{i, j} each assigned an importance e _{i, j} and expressed by a feature vector. = 1, 2,..., N), j is the index of the symbol sequence in each i (j = 1, 2,..., N _i )) and the correct answer for each list i represented by a feature vector. The apparatus receives a symbol sequence f _{i, 0} and estimates and outputs a model parameter w, and includes a grouping unit 101, a merging unit 102, and a model parameter estimation unit 103.

The grouping unit 101 groups a plurality of symbol sequences f _{i, j} belonging to the list into a plurality of groups G _i (x) (x is a group index) by a predetermined method (S1). The method of grouping is arbitrary, for example, by using a general method such as K-means, groups of features in the feature vector space that are close to each other are grouped, or those having a close importance value are used. Grouping is possible. Further, when the degree of importance is an error rate, an operation may be performed such as dividing into a group that is close to the correct symbol sequence and another group, and further allowing the correct symbol sequence to belong to a group that is close to the correct answer.

Merging unit 102, the group G _i (x) the group G _i (x) belonging to a plurality of symbol sequences f _i for each _{representative} from _j symbol sequence f _i, the _x, also belonging to the group G _i (x) The representative importance e _{i, x} is obtained from the plurality of importance e _{i, j} corresponding to each symbol series (S2). Specifically, for example, the representative symbol series f _{i, x} is obtained by the merge function F in Expression (5), and the representative importance e _{i, x} is obtained by the merge function E in Expression (6). In equations (5) and (6), (f _{i, j} , e _{i, j} ) represents a set of symbol sequences and corresponding importance levels.

As an example of merging into the representative symbol series f _{i, x} by the merge function F, for example, methods shown in equations (7) and (8) can be given.

Expression (7) is a method for obtaining the representative symbol series as a centroid of the symbol series belonging to the group. Equation (8) is a method for obtaining a representative symbol sequence as a weighted internal point of a symbol sequence belonging to a group, and even when a model parameter estimation method that cannot take into account importance is adopted, There is an advantage that importance can be considered. In addition to these merging, the storage area can be further reduced by quantizing each element of the feature vector of the symbol series and truncating the decimal digits.
As a method of merging the representative importance e _{i, x} with the merge function E, for example, a method using an average value of importance shown in Expression (9) can be cited.

The model parameter estimation unit 103 calculates and outputs a model parameter w from the representative symbol series f _{i, x} , the correct symbol series f _{i, 0,} and the representative importance e _{i, x} (S3). For example, equation (3) based on the ELBst method disclosed in Non-Patent Document 3 may be transformed into equation (10) to obtain w that minimizes the value of L in equation (10).

In the equation (10), there is an algorithm for efficiently estimating w particularly when the feature value is binary of 0 and 1. Further, in the case of the MERT method disclosed in Non-Patent Document 4, equation (4) may be transformed into equation (11) to obtain w that minimizes the value of L in equation (11). .

Here, α is a hyper parameter, which is determined using a development set or the like. According to Equation (11), the model parameter w can be estimated without using the correct symbol sequence. The estimation of the model parameter w can be specifically performed by a known method such as L-BFGS. Further, in the case of the GCLM method disclosed in Non-Patent Document 5, the formula (2)
May be transformed as shown in Equation (12) to obtain w that minimizes the value of L in Equation (12).

Here, ‖w‖ is a norm, and it is known that a robust estimation result can be obtained by using this norm. C is a hyperparameter, which is determined using a development set or the like. According to Expression (12), it is guaranteed that the estimation result of the model parameter w converges to the optimal solution. Specifically, the model parameter w can be estimated by a known method such as L-BFGS.

<Verification of effects>
A Japanese spoken corpus (CSJ) is used to verify the effect of the present invention. CSJ is a database consisting of speech data and transcripts. In the verification, two sets for evaluation and for learning and for development shown in FIG. 3 were prepared.

  The lecture was divided into utterance units, and a 5000-best list was created with a 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. For the feature, uni, bi-, tri-gram boolean and speech recognition score were used. For the importance, the rank in the list of each symbol series (ascending order of word error rate) was used. Note that the word error rate shown in FIG. 3 is calculated for the recognition result having the largest recognition score in the 5000-best list output by the speech recognition system. Perplexity is an index representing the proximity of data, and is calculated by a language model in the speech recognition system. It can be seen from the size of the perplexity that the evaluation B includes many different properties from the other sets.

  When model parameter w is estimated using all symbol sequences and merged as in the present invention (merge using equation (8)), symbol sequences are rearranged using these, respectively, Finally, the symbol sequences with the highest scores were used as speech recognition results, and the word error rates were compared. Note that FIG. 4 shows the memory size required for data retention in this verification. When all data is used, a storage area of several tens of gigabytes is required, whereas when symbol sequences are merged using equation (8) In other words, the consumption of the storage area is reduced to a mega-order that can be operated by a general-purpose computer. FIG. 5 compares the word error rates in the speech recognition results obtained using the model parameters learned by Equation (10) based on the ELBst method. It can be seen that the error rate is comparable when learning is performed using all the data and when learning is performed using the compressed data merged in Equation (8). Note that, when learning is performed using all data, that is, when a plurality of symbol sequences are used for one correct answer, there is a possibility that the parameter estimation value is strongly influenced by the correct symbol sequence according to the ELBst method. On the other hand, it is considered that the influence is deleted without Expression (8) + importance merging, and as a result, a model with higher accuracy than learning using all data is generated. In this verification, when importance level merging is performed, the variation in importance level between lists increases, and the accuracy decreases. Of course, the B set for evaluation has the same accuracy as when all data is used. This is because there is an expressive power that includes importance merging and no importance merging, so importance merging may work effectively depending on the importance design method and the application target of the present invention. I think that represents. FIG. 6 compares the word error rates in the speech recognition results obtained using the model parameters learned by Equation (12) based on the GCLM method. The GCLM method has no framework for handling importance. Nevertheless, in learning using all data, a model having performance equivalent to or better than the ELBst method is generated. One of the reasons may be convergence to a global optimal solution. When symbol sequences are merged, importance is expressed in the feature vector space, so that a more accurate model is generated. FIG. 7 compares the word error rates in the speech recognition results obtained using the model parameters learned by Equation (11) based on the MERT method. Comparing the case of using all data and the case of merging with equation (8), although the accuracy is greatly reduced in the evaluation B set, it is the same in the evaluation A set having characteristics similar to the learning / development set Performance is obtained.

  As described above, according to the information compression type model parameter estimation apparatus and method of the present invention, it is possible to compress symbol sequence information used for learning while securing learning accuracy to the same level as in the past. Model parameter estimation processing can be performed.

  When each of the above devices is realized by a computer, processing contents of functions that each device should have are described by a program. The processing functions are realized on the computer by executing the program on the computer. In this case, at least a part of the processing content may be realized by hardware. Further, the various processes described above are not only executed in time series according to the description, but may be executed in parallel or individually according to the processing capability of the apparatus that executes the processes or as necessary. In addition, it can change suitably in the range which does not deviate from the meaning of this invention.

Claims (11)

N lists i (i is an index of the list, N is an integer greater than or equal to 1, and j is an integer greater than or equal to 1 _), each of which is made up of n _i symbol sequences f _{i, j} assigned importance levels e _{i, j} and represented by feature vectors. Symbol sequence index in each i, n _i is an integer of 4 or more) and correct symbol sequence f _{i, 0 of} each list i expressed by a feature vector, respectively _, and an information compression type for estimating a model parameter w A model parameter estimation device,
A grouping unit that groups a plurality of symbol sequences f _{i, j} belonging to the list into a plurality of groups G _i (x) (x is an index of a group) for each list i;
The group G _i (x) belonging to a plurality of symbol sequences f _i, representatives from _j symbol sequence f _{i, x,} and also the group G _i (x) belonging to a plurality of symbol sequences f _i, of the plurality corresponding to _j A merging unit for obtaining representative importance e _{i, x} from importance e _{i, j} , respectively;
A model parameter estimator for estimating a model parameter w from the representative symbol sequence f _{i, x} , the correct symbol sequence f _{i, 0,} and the representative importance e _{i, x} ;
An information compression model parameter estimation device comprising: N lists i (i is an index of the list, N is an integer greater than or equal to 1, and j is an integer greater than or equal to 1 _), each of which is made up of n _i symbol sequences f _{i, j} each assigned an importance e _{i, j} and represented by a feature vector An index of a symbol series in each i, where n _i is an integer of 4 or more), and is an information compression model parameter estimation device that estimates a model parameter w,
A grouping unit that groups a plurality of symbol sequences f _{i, j} belonging to the list into a plurality of groups G _i (x) (x is an index of a group) for each list i;
The group G _i (x) belonging to a plurality of symbol sequences f _i, representatives from _j symbol sequence f _{i, x,} and also the group G _i (x) belonging to a plurality of symbol sequences f _i, of the plurality corresponding to _j A merging unit for obtaining representative importance e _{i, x} from importance e _{i, j} , respectively;
A model parameter estimation unit for estimating a model parameter w from the representative symbol series f _{i, x} and the representative importance e _{i, x} ;
An information compression model parameter estimation device comprising: Each of N lists i (i is an index of the list, N is an integer greater than or equal to 1, j is an index of the symbol sequence in each i, and n is composed of n _i symbol sequences f _{i, j} each represented by a feature vector _i is an integer of 4 or more) and a correct symbol sequence f _{i, 0 of} each list i each represented by a feature vector, and is an information compression type model parameter estimation device for estimating a model parameter w,
A grouping unit that groups a plurality of symbol sequences f _{i, j} belonging to the list into a plurality of groups G _i (x) (x is an index of a group) for each list i;
A merging unit for obtaining a representative symbol sequence f _{i, x} from a plurality of symbol sequences f _{i, j} belonging to the group G _i (x);
A model parameter estimation unit for estimating a model parameter w from the representative symbol sequence f _{i, x} and the correct symbol sequence f _{i, 0} ;
An information compression model parameter estimation device comprising: In the information compression model parameter estimation device according to claim 1 or 2,
The information compression type model parameter estimation apparatus, wherein the grouping unit performs grouping based on a distance in a feature vector space or an importance value. In the information compression model parameter estimation device according to claim 1 or 2,
The merging unit obtains a representative symbol sequence f _{i, x} as a centroid or a weighted inner dividing point of a plurality of symbol sequences f _{i, j} belonging to the group G _i (x), and sets the representative importance e _{i, x} to the group An information compression model parameter estimation apparatus characterized by obtaining an average value of a plurality of importance levels e _{i, j} corresponding to a plurality of symbol sequences f _{i, j} belonging to G _i (x). N or more lists i (i is an index of the list, N is an integer of 1 or more, and j is composed of n _i symbol sequences f _{i, j} each assigned an importance e _{i, j} and represented by a feature vector. Is a symbol sequence index for each i, n _i is an integer greater than or equal to 4), and correct symbol sequences f _{i, 0 of} each list i each represented by a feature vector, and information compression for estimating a model parameter w Type model parameter estimation method,
A grouping step for grouping a plurality of symbol sequences f _{i, j} belonging to the list into a plurality of groups G _i (x) (x is an index of the group) for each list i;
The group G _i (x) belonging to a plurality of symbol sequences f _i, representatives from _j symbol sequence f _{i, x,} and also the group G _i (x) belonging to a plurality of symbol sequences f _i, of the plurality corresponding to _j A merging step for obtaining the representative importance e _{i, x} from the importance e _{i, j} , respectively;
A model parameter estimation step for estimating a model parameter w from the representative symbol sequence f _{i, x} , the correct symbol sequence f _{i, 0} and the representative importance e _{i, x} ;
The information compression type model parameter estimation method which performs. N lists i (i is an index of the list, N is an integer greater than or equal to 1, and j is an integer greater than or equal to 1 _), each of which is made up of n _i symbol sequences f _{i, j} assigned importance levels e _{i, j} and represented by feature vectors. An index of a symbol sequence in each i, where n _i is an integer of 4 or more), and is an information compression model parameter estimation method for estimating a model parameter w,
A grouping step for grouping a plurality of symbol sequences f _{i, j} belonging to the list into a plurality of groups G _i (x) (x is an index of the group) for each list i;
The group G _i (x) belonging to a plurality of symbol sequences f _i, representatives from _j symbol sequence f _{i, x,} and also the group G _i (x) belonging to a plurality of symbol sequences f _i, of the plurality corresponding to _j A merging step for obtaining the representative importance e _{i, x} from the importance e _{i, j} , respectively;
A model parameter estimation step for estimating a model parameter w from the representative symbol series f _{i, x} and the representative importance e _{i, x} ;
The information compression type model parameter estimation method which performs. Each of N lists i (i is an index of the list, N is an integer greater than or equal to 1, j is an index of the symbol sequence in each i, and n is composed of n _i symbol sequences f _{i, j} each represented by a feature vector _i is an integer greater than or _equal to 4) and a correct symbol sequence f _{i, 0 of} each list i each represented by a feature vector, and is an information compression model parameter estimation method for estimating a model parameter w,
A grouping step for grouping a plurality of symbol sequences f _{i, j} belonging to the list into a plurality of groups G _i (x) (x is an index of the group) for each list i;
A merging step for obtaining a representative symbol sequence f _{i, x} from a plurality of symbol sequences f _{i, j} belonging to the group G _i (x);
A model parameter estimation step for estimating a model parameter w from the representative symbol sequence f _{i, x} and the correct symbol sequence f _{i, 0} ;
The information compression type model parameter estimation method which performs. In the information compression type model parameter estimation method according to claim 6 or 7 ,
In the information compression type model parameter estimation method, the grouping step includes grouping based on a distance in a feature vector space or a value of importance. In the information compression type model parameter estimation method according to claim 6 or 7 ,
In the merging step, the representative symbol sequence f _{i, x} is obtained as a centroid or a weighted inner dividing point of a plurality of symbol sequences f _{i, j} belonging to the group G _i (x), and the representative importance e _{i, x} is determined as a group. An information compression model parameter estimation method, characterized in that it is obtained as an average value of a plurality of importance levels e _{i, j} corresponding to a plurality of symbol sequences f _{i, j} belonging to G _i (x). Program for causing a computer to function as an apparatus according to any one of claims 1 to 5.

Images