JP4537970B2 - Language model creation device, language model creation method, program thereof, and recording medium thereof - Google Patents

Description

  The present invention relates to a technique for purposely creating a language model.

  For example, in speech recognition and statistical machine translation, a language model is used as a linguistic restriction for improving recognition performance. When the usage (task) such as voice recognition is limited, it is generally said that the recognition accuracy can be increased by using a language model specially constructed for the usage.

  The statistical language model N-gram, which is a language model actively used in recent years, needs to learn a large amount of data in order to construct a high-performance model. When the usage is limited, it is generally difficult to collect a large amount of text data related to the usage. In order to solve this problem, an adaptation method of a language model has been proposed in which a model is adapted using a target text from a language model learned from a large amount of text data including non-use text.

In Non-Patent Document 1, a plurality of text databases are prepared, an optimum weight is found using a text for a specific purpose with respect to the language model constructed by each, and each language model is added by the weight, thereby obtaining a language. A method for adapting the model is disclosed.
Non-Patent Document 2 discloses a method of adapting a language model to a specific application from a language model constructed by each of a plurality of text databases by estimating maximum posterior probability instead of weight.
PR Clarkson and AJ Robinson, "Language model adaptation using mixture and an exponentially decaying cache", Proc. ICASSP'97, vol.2, pp.799-802, April, 1997. Masakazu Masatsugu, Yoshinori Kasaka, Kazuya Hisaki, Tatsuya Kawahara, "Task Adaptation of N-gram Language Model by Maximum A posteriori Probability Estimation", IEICE Transactions, Vol. J81-D-II, no. 11, pp. 2519-2525, November 1998

  The prior art has a problem in that text data suitable for the intended application is required even in a small amount. Optimum weights and maximum posterior probability estimation are focused on obtaining optimal performance as a language model for texts related to a given application. There was a problem that accuracy was not always improved in the field.

  In view of the above problems, an object of the present invention is to create a language model (adaptive language model) suitable for a target application without requiring text data suitable for the target application.

In order to solve the above problems, the present invention creates an adaptive language model as follows. That is, a cluster language model corresponding to each text data cluster is created from a plurality of text data clusters. A language model (hereinafter referred to as “partial selection synthesis cluster language model”) is synthesized from the combination of the remaining cluster language models excluding some cluster language models from all cluster language models. Then, each partial selection synthesis cluster language model is evaluated using the evaluation data and the evaluation acoustic model, and the evaluation result of each partial selection synthesis cluster language model is output. Furthermore, the cluster language model excluded in the synthesis of the partial selection synthesis cluster language model that gave the low evaluation result among these evaluation results is selected, and the cluster language model is created when there is only one selected cluster language model Is output as a language model. When there are a plurality of selected cluster language models, one language model is synthesized from them and output.

Further, in the above-described selection, the low evaluation results, as walk lower than the threshold value is an evaluation result is smaller than the threshold value, the language model or partial selective synthesis cluster language model excluded cluster language model in the synthesis of May be selected.

Further, the adaptive language model may be created as follows. That is, the input text data is divided into a plurality of text data clusters according to the classification criteria.
According to this, when using, for example, existing general-purpose mass text data as text data, a text data cluster can be created with an appropriate classification standard according to the application.

  Further, the computer can be operated as a language model creation device by a language model creation program that causes the computer to function as the language model creation device of the present invention. Then, it becomes possible to cause another computer to function as a language model creation device or to distribute the language model creation program by using a computer-readable program recording medium that records this language model creation program.

  According to the present invention, a synthesized cluster language model is synthesized from a plurality of combinations among cluster language models corresponding to a plurality of text data clusters, and evaluation data of each cluster language model and / or each synthesized cluster language model is obtained. Since a language model that gives a high evaluation result is selected, it is possible to create a language model (adaptive language model) suitable for the intended application without requiring text data suitable for a specific application.

  Further, since the cluster language model or the like is selected based on the recognition accuracy rather than the performance as the language model, an adaptive language model with improved recognition accuracy can be obtained more directly.

<< First Embodiment >>
A first embodiment of the present invention will be described with reference to the drawings.
<Language Model Creation Device of First Embodiment>
As illustrated in FIG. 1, the language model creation device (1) includes an input unit (11) to which a keyboard or the like can be connected, an output unit (12) to which a liquid crystal display or the like can be connected, and a CPU (Central Processing Unit; 14). [A cache memory or the like may be provided. RAM (Random Access Memory) (15), ROM (Read Only Memory) (16) and external storage device (17) which is a hard disk, and these input unit (11), output unit (12), A CPU (14), a RAM (15), a ROM (16), a bus (18) connected so as to exchange data between the external storage devices (17), and the like are provided. If necessary, the language model creation device (1) may be provided with a device (drive) that can read and write a storage medium such as a CD-ROM. A physical entity having such hardware resources includes a general-purpose computer.

  The external storage device (17) of the language model creation device (1) stores and stores a program for creating a language model, data necessary for processing the program, and the like. Further, data obtained by the processing of these programs is appropriately stored and stored in the RAM (15) or the like.

  The external storage device (17) stores and stores text data (111). It is assumed that the text data (111) is divided into a plurality (N) of data in advance. Each piece of data is called a “text data cluster”. That is, the text data (111) is composed of a text data cluster [1] (111-1), a text data cluster [2] (111-2),..., A text data cluster [N] (111-N). Is done. In addition, each text data cluster is not limited to what divided | segmented one text data, For example, several text data may be prepared and each may be regarded as another text data cluster, or a text data cluster May be a combination of a plurality of text data. Furthermore, if a plurality of text data clusters are prepared in such a case, text data clusters having the same contents can exist. However, the present invention is also acceptable in such a case. In other words, the plurality of text data clusters are preferably different from each other, but the present invention can be implemented even when text data clusters having the same contents can exist [this is the same in each embodiment, etc. It is. ].

  Further, in the external storage device (17), evaluation data (118) for evaluating the language model [the evaluation data is assumed to be composed of voice data for evaluation and voice transcription text. This assumes the case where a language model is used for speech recognition, and is not intended to exclude preparing evaluation data according to applications such as machine translation. Also, the evaluation voice data is preferably acoustically close to the scheduled voice recognition target voice. , And an acoustic model for evaluation (119) are stored and stored as data.

  In the external storage device (17), a program for creating a language model from a text data cluster (hereinafter, a language model created from the text data cluster is referred to as “cluster language model”), and a language from a plurality of cluster language models. A program for synthesizing a model (hereinafter referred to as “synthesis cluster language model”), a program for evaluating a language model, and a program for selecting a language model from an evaluation result are stored and stored.

  In the language model creation device (1), each program stored in the external storage device (17) and data necessary for the processing of each program are read into the RAM (15) as necessary, and the CPU (14). Interpreted and processed. As a result, the creation of the language model is realized by the CPU (14) realizing predetermined functions (cluster language model creation unit, model synthesis unit, model evaluation unit, selection unit).

<Outline of First Embodiment>
In the first embodiment, a cluster language model is created for each text data cluster, and a synthesized cluster language model is synthesized from a plurality of combinations among the respective cluster language models. Each cluster language model and each synthesized cluster language model are evaluated using the evaluation data (118) and the acoustic model for evaluation (119), and a language model that achieves the highest recognition performance as an example of a predetermined evaluation result Is identified as an adaptive language model. The predetermined evaluation result is not limited to the highest recognition performance, and a good evaluation that the recognition rate is improved by a relative evaluation with respect to a threshold value or the like may be used as the predetermined evaluation result.

<Language Model Creation Processing of First Embodiment>
Next, the flow of language model creation processing in the language model creation device (1) will be described descriptively with reference to FIGS.

First, the cluster language model creation unit (113) performs text data cluster [1] (111-1), text data cluster [2] (111-2),..., Text data cluster [N] (111-N ) Is read from the RAM (15), and a set (114) of cluster language models corresponding to each text data cluster is created using each text data cluster (step S1).
That is, the cluster language model creation unit (113) creates the cluster language model [1] (114-1) from the text data cluster [1] (111-1), and similarly the text data cluster [2] (111). -2) creates a cluster language model [2] (114-2), and so on, and creates a cluster language model [N] (114-N) from a text data cluster [N] (111-N). is there. N cluster language models [1] (114-1), cluster language models [2] (114-2),..., And cluster language models [N] (114-N) are appropriately stored in the RAM (15) or the like. Stored in memory.
Note that a method for creating a language model (cluster language model) from text data (cluster text data) is based on a publicly known method, and therefore a description thereof is omitted. The language model can be used as long as it is a model that can automatically learn from text data such as N-gram.

  Next, the model synthesis unit (115) generates a synthesized cluster language model from a plurality of combinations among the cluster language models (114-1) (114-2) (114-N) read from the RAM (15). (116) is synthesized (step S2). In the first embodiment, the “plurality of combinations” means all exhaustive combinations that can be considered as a plurality of combinations.

More specifically, the model synthesis unit (115) calculates Σ _{j = 2} ^N _N C from N cluster language models (114-1) (114-2) (114-N). That is, _j synthetic cluster language models are created. Here, _N C _j represents the number of combinations in which j is selected from N without allowing duplication. For the symbol M, M = Σ _{j = 2} ^N _N C _j = _{NC 2} + _{NC 3} +... + _{NC N}
That is, the model synthesis unit (115) synthesizes the synthesized cluster language model [1 + 2] (116-2) from, for example, the cluster language model [1] (114-1) and the cluster language model [2] (114-2). From the cluster language model [1] (114-1), the cluster language model [2] (114-2), and the cluster language model [3] (114-3), the synthesized cluster language model [1 + 2 + 3] (116-3) ), And synthesize the synthesized cluster language model [5 + 9] from the cluster language model [5] and the cluster language model [9] (114-3), and so on, from all the cluster language models. Synthesizes the synthesized cluster language model [N + (N-1) +... +1] (116-M). Note that a method for synthesizing one language model (synthetic cluster language model) from a plurality of language models (cluster language model) is based on a known method, so the description thereof is omitted [for example, refer to the reference literature. ].
(Reference) Japanese Patent Application Laid-Open No. 2004-53745

  M synthetic cluster language models (116-1), (116-2),... (116-M) are appropriately stored and stored in the RAM (15) or the like. When the processing in step S2 is completed, the RAM (15) has N cluster language models (114-1) (114-2) (114-N) and M synthetic cluster language models. A total of (N + M) language models (116-1), (116-2),... (116-M) are stored and stored.

  In the first embodiment, “plural combinations” refers to all exhaustive combinations that can be considered as a plurality of combinations. However, when the number of text data clusters is large, the number of exhaustive combinations becomes very large, so that the synthesized cluster language model may be synthesized with arbitrary combinations. That is, fewer than M synthetic cluster language models may be synthesized in any combination.

Subsequently, the model evaluation unit (117) includes (N + M) language models (114-1) (114-2) (114-N) (116-1) (116-2). (116-M), the evaluation data (118) and the evaluation acoustic model (119) are read from the RAM (15), and each language model and the evaluation acoustic model (119) are used for the evaluation data (118). A recognition rate (150) is obtained and calculated as an evaluation result of each language model (step S3).
Since the difference in recognition rate as an evaluation result is evaluated using the same evaluation data (118) and acoustic model (119) for evaluation, each language model (114-1) (114-2). Based on the difference of (114-N) (116-1) (116-2) (116-M). The calculation of the recognition rate of the evaluation voice data is achieved by a known method.

A specific functional configuration of the model evaluation unit (117) will be described with reference to FIG.
The model evaluation unit (117) includes a feature amount calculation unit (1171), a solution search unit (1172), and a text matching unit (1173).
The feature amount calculation unit (1171) calculates the feature amount of the audio data (1181) constituting the evaluation data (118). This feature amount is, for example, multidimensional MFCC (Mel Frequency Cepstrum Coefficient).
Next, the solution search unit (1172) includes the feature amount of the speech data, the acoustic model for evaluation (119), each language model (114-1) (114-2) (114 -N) (116-1). ) (116-2) (116-M) is used to output a recognition result text for the voice data. In this process, the recognition result text corresponding to each language model (114-1) (114-2) (114-N) (116-1) (116-2) (116-M). Is obtained.
Subsequently, the text matching unit (1173) uses the speech transcription text (1182) and each recognition result text constituting the evaluation data (118) to recognize the recognition result text for the speech transcription text (1182). (150) is calculated. This recognition rate (150) is the evaluation result of each language model.

Subsequently, the selection unit (120) reads the language models (114-1) (114-2) (114-N) (116-1) (116-2) (116-M). Based on the evaluation result, the highest evaluation result among the language models (114-1) (114-2) (114-N) (116-1) (116-2) (116-M) Is selected and output as an adaptive language model (123) (step S4). This adaptive language model (123) is a language model adapted for the intended use.
In the first embodiment, the language model giving the highest evaluation result is selected. However, a threshold value is set in advance, and the language model giving an evaluation result equal to or higher than this threshold value (or larger than the threshold value) is selected. You may do it. In this case, a plurality of language models can be selected, but these become acoustic model candidates adapted to the intended use. In other words, it is possible to output a single language model by setting this threshold strictly, or set a low threshold if it is not desired to output only one language model. Thus, a plurality of language models can be output as candidates for the adaptive language model. This threshold value is stored and stored in advance in the external storage device (17).

  Note that the influence of the text data amount can be reduced by equalizing the data amount of each text data cluster. In addition, by changing the evaluation data according to the intended use, a language model adapted to the use can be created.

<< Second Embodiment >>
A second embodiment of the present invention will be described with reference to the drawings.
<Outline of Second Embodiment>
In the second embodiment, a cluster language model is created for each text data cluster, and a synthesized cluster language model is synthesized from a plurality of combinations of the respective cluster language models. Each cluster language model and each synthesized cluster language model are evaluated using the evaluation data (118) and the evaluation acoustic model (119). Furthermore, one language model is synthesized from the language models that give a predetermined evaluation result [hereinafter referred to as “integrated processing”. ], The language model output by this integration processing is set as an adaptive language model. For the predetermined evaluation result, see the description of the first embodiment.

<Language Model Creation Device of Second Embodiment>
The language model creation apparatus according to the second embodiment has the same hardware configuration as that of the language model creation apparatus according to the first embodiment, and only different parts from the first embodiment will be described.
In the second embodiment, in addition to the program of the first embodiment, the external storage device (17) also stores a program for performing an integration process for synthesizing one language model from a language model given a predetermined evaluation result. Has been.

  In the language model creation device (1) of the second embodiment, each program stored in the external storage device (17) and data necessary for processing each program are read into the RAM (15) as necessary, Interpretation is executed and processed by the CPU (14). As a result, the creation of the language model is realized by the CPU (14) realizing predetermined functions (cluster language model creation unit, model synthesis unit, model evaluation unit, selection unit, model integration unit).

<Language Model Creation Processing of Second Embodiment>
Next, with reference to FIG. 5 and FIG. 6, the flow of the language model creation process in the second embodiment will be described descriptively. Here, parts different from the flow of the language model creation process in the first embodiment will be described.

In the language model creation process in the second embodiment, the following process is performed following the process in step S3 in the first embodiment.
That is, the selection unit (120a) includes (N + M) language models (114-1) (114-2) (114-N) (116-1) (116-2) (116). Based on the evaluation result of -M), a language model giving an evaluation result equal to or higher than the threshold (or larger than the threshold) is selected (step S4a). This threshold value is stored in advance in the external storage device (17), and is read by the selection unit (120a) in the process of step S4a.
For example, when the language model that gave the evaluation result equal to or higher than the threshold in the process of step S3 is the cluster language model [1] and the synthetic cluster language model [5 + 9], the selection unit (120a) selects the cluster language model [1] and Select the synthetic cluster language model [5 + 9].
Note that when one language model is selected, which corresponds to the case of the first embodiment, it is assumed that a plurality of language models are selected in the second embodiment.

Next, the model integration unit (121) performs an integration process on the selected plurality of language models and outputs an adaptive language model (123) (step S5).
The integration process of the model integration unit (121) is the same as the model synthesis process of the model synthesis unit (115) [see the above-mentioned reference. ]. However, although M synthesized cluster language models are created in the model synthesis process of the model synthesis unit (115), one language model is created in the integration process of the model integration unit (121).
In other words, the case where the cluster language model [1] and the synthesized cluster language model [5 + 9] are selected will be described as an example. The model integration unit (121) determines the cluster language model [1] and the synthesized cluster language model [5 + 9]. Then, model synthesis is performed for one language model. This language model is the adaptive language model (123).

<< Third Embodiment >>
A third embodiment of the present invention will be described with reference to the drawings.
<Outline of Third Embodiment>
In the third embodiment, a cluster language model is created for each text data cluster, and a language model (partial selection synthesis cluster language model from the remaining combinations excluding some of all the cluster language models. Abbreviated as “composite cluster language model”). Each synthetic cluster language model is evaluated using the evaluation data (118) and the evaluation acoustic model (119). Further, the cluster language model excluded in the synthesis of the synthesized cluster language model giving the predetermined evaluation result is selected, and one language model is synthesized from the selected cluster language model [integration process]. The output language model is set as an adaptive language model. Here, the predetermined evaluation result is an evaluation result that is equal to or less than the threshold value or smaller than the threshold value.

<Language Model Creation Device of Third Embodiment>
The language model creation apparatus according to the third embodiment has the same hardware configuration as that of the language model creation apparatus according to the second embodiment, and only different parts from the second embodiment will be described.
In the third embodiment, the program for realizing the selection unit (120) in the second embodiment is a program for selecting a cluster language model excluded in the synthesis of the synthetic cluster language model that gave a predetermined evaluation result. And
In the third embodiment, the program for realizing the model integration unit (121) in the second embodiment is a program for creating an adaptive language model from the selected cluster language model.

  In the language model creation device (1) of the third embodiment, each program stored in the external storage device (17) and data necessary for the processing of each program are read into the RAM (15) as necessary. Interpretation is executed and processed by the CPU (14). As a result, the creation of the acoustic model is realized by the CPU (14) realizing predetermined functions (cluster language model creation unit, model synthesis unit, model evaluation unit, selection unit, model integration unit).

<Language Model Creation Processing of Third Embodiment>
Next, with reference to FIG. 7 and FIG. 8, the flow of the language model creation process in the third embodiment will be described descriptively. Here, parts different from the flow of language model creation processing in the second embodiment will be described.

In the language model creation process in the third embodiment, the following process is performed following the process in step S1 in the second embodiment.
That is, the model synthesizing unit (115b) selects each of the cluster language models (114-1), (114-2),... (114-N) read from the RAM (15) from each cluster language model. The synthesized cluster language model (116b) is synthesized from the remaining cluster language model combinations excluding the one cluster language model (step S2b).
That is, the model synthesis unit (115b) removes the cluster language model [1] (114-1) from all the cluster language models (114-1) (114-2) (114-N). The synthesized cluster language model [1 none] (116b-1) is synthesized from one cluster language model (114-2) (114-3) (114-N), and all cluster language models (114- 1) N-1 cluster language models (114-1) (114-3) obtained by excluding the cluster language model [2] (114-2) from (114-2) (114-N) A synthesized cluster language model [2 none] (116b-2) is synthesized from (114-N), and all cluster language models (114-1) (114-2) (114-N) are synthesized. ) To cluster language model [N From the N-1 cluster language models (114-1) (114-2) (114- (N-1)) excluding (114-N), the synthesized cluster language model [N None] (116b- N) is synthesized.
The N synthetic cluster language models (116b-1)... (116b-N) are appropriately stored and stored in the RAM (15) or the like.

Subsequent to step S2b, step S3b is executed. Specifically, the model evaluation unit (117b) receives each of the N synthesized cluster language models (116b-1) (116b-N), the evaluation data (118), and the evaluation acoustic model (119). Reading from the RAM (15), the recognition rate for the evaluation data (118) is obtained using each synthetic cluster language model and the acoustic model for evaluation (119), and this is calculated as the evaluation result of each synthetic cluster language model ( Step S3b). The functional configuration of the model evaluation unit (117b) is the same as that of the model evaluation unit (117) of the first embodiment.
Since the difference in recognition rate as an evaluation result is evaluated using the same evaluation data (118) and evaluation acoustic model (119), each synthetic cluster language model (116b-1) (116b). -Based on the difference of N).

Subsequently, the selection unit (120b) was excluded in the synthesis of the synthesized cluster language model that gave a predetermined evaluation result among the evaluation results of each synthesized cluster language model (116b-1) (116b-N). A cluster language model is selected (step S4b).
Here, the “predetermined evaluation result” means that each evaluation result of the composite cluster language model [no j] (116b-j) [j = 1, 2,..., N] is equal to or less than a predetermined threshold (or It is the evaluation result that it is smaller than the threshold value. In other words, the cluster language model excluded in the synthesis of the synthesized cluster language model that gave the evaluation result below the threshold is selected.
For example, the recognition rate is set to 70%, the evaluation result of the synthetic cluster language model [ 5 None] is 60%, the evaluation result of the synthetic cluster language model [9 None] is 63% in the processing of Step S3b, and each other synthetic cluster language If the evaluation result of the model is larger than 70%, the evaluation results of the synthetic cluster language model [5 none] and the synthetic cluster language model [9 none] are below the threshold value 70%. Therefore, the selection unit (120b) selects the cluster language model [5] and the cluster language model [9] that are excluded by the synthesis of the synthetic cluster language model [5 none] and the synthetic cluster language model [9 none].

Following step S4b, step S5b is executed. The model integration unit (121b) performs an integration process on the selected cluster language model and outputs an adaptive language model (123) (step S5). If the selected cluster language model is one, the model integration unit (121b) is not necessary, and the cluster language model selected by the selection unit (120b) may be output as an adaptive language model. It can be set as the structure similar to having demonstrated by the process of step S4.
The integration process of the model integration unit (121b) is the same as the model synthesis process of the model synthesis unit (115) [see the above-mentioned references. ]. However, although M synthesized cluster language models are created in the model synthesis process of the model synthesis unit (115), one language model is created in the integration process of the model integration unit (121b).
That is, the case where the cluster language model [5] and the cluster language model [9] are selected will be described as an example. The model integration unit (121b) performs the processing on the cluster language model [5] and the cluster language model [9]. Perform model synthesis and output one language model. This language model is the adaptive language model (123).

The configuration of the third embodiment is that the recognition accuracy of a certain synthetic cluster language model k is low. That is, the recognition accuracy deteriorates by excluding the cluster language model k. In other words, the cluster language model k is recognized. This is based on greatly contributing to accuracy improvement. The third embodiment is not a method for guaranteeing that the language model having the highest evaluation criterion is selected from all the combinations, but the number of synthetic cluster language models to be evaluated is the number of text data clusters. since the same der Ru significantly reduces the number of evaluations, it can be said that an effective way to perform adaptive language model creating a realistic computational complexity.

<< 4th Embodiment >>
A fourth embodiment of the present invention will be described with reference to the drawings.
<Outline of Fourth Embodiment>
The fourth embodiment includes a data classification process for classifying text data having an enormous amount of data into a plurality of text data clusters according to an appropriate classification standard. This data classification process can be applied in combination with the first embodiment, the second embodiment, and the third embodiment. The fourth embodiment will be described as a combination with the second embodiment.

<Language Model Creation Device of Fourth Embodiment>
The language model creation apparatus according to the fourth embodiment has the same hardware configuration as that of the language model creation apparatus according to the second embodiment, and only different parts from the second embodiment will be described.
In the fourth embodiment, in addition to the program of the second embodiment, a program for classifying text data is also stored in the external storage device (17).
In the second embodiment, N text data clusters are stored and stored in the external storage device (17). However, in the fourth embodiment, the text data (111a before division) is stored in the external storage device (17). ) Is stored and stored.

  In the language model creation device (1) of the fourth embodiment, each program stored in the external storage device (17) and data necessary for the processing of each program are read into the RAM (15) as necessary. Interpretation is executed and processed by the CPU (14). As a result, creation of a language model is realized by the CPU (14) realizing predetermined functions (data classification unit, cluster language model creation unit, model synthesis unit, model evaluation unit, selection unit, model integration unit). .

<Language Model Creation Processing of Fourth Embodiment>
Next, with reference to FIG. 9 and FIG. 10, the flow of the language model creation process in the fourth embodiment will be described descriptively. Here, parts different from the flow of language model creation processing in the second embodiment will be described.

The data classification unit (110) reads large-scale existing text data (111a) obtained from newspaper articles, the Web, etc., and text data (111a) according to a preset classification standard (for example, topic). Are divided into N text data clusters [1] (111-1), text data clusters [2] (111-2),..., Text data clusters [N] (111-N) and output ( Step S1p). As a classification method, a general clustering method such as a K-means method can be used.
The N text data clusters are appropriately stored and stored in the RAM (15) or the like. In FIG. 9, text data (111b) indicates a set of text data clusters after division.

  Subsequent to step S1p, the processing after step S1 of the second embodiment is performed.

Although various embodiments have been described, all of them select a cluster language model or a synthetic cluster language model by evaluation using evaluation data, and create an adaptive language model. By using the data, it is possible to create an optimal language model specialized for the application.
In addition to selecting a cluster language model or a synthesized cluster language model according to a threshold for a difference in recognition performance, elements such as a text data amount restriction may be added in addition to the threshold. Furthermore, since the magnitude of the recognition performance difference is considered to be a contribution to the recognition performance improvement, for example, the adaptive language model is synthesized by weighting the selected synthesized cluster language model according to the recognition performance difference. In this case, the performance of the adaptive language model can be improved.

  In addition to the above embodiments, the language model creation apparatus and method according to the present invention are not limited to the above-described embodiments, and can be appropriately changed without departing from the spirit of the present invention. In addition, the processing described in the language model creation device / method is not only executed in time series in the order described, but also executed in parallel or individually as required by the processing capability of the device that executes the processing. It may be.

  When the processing function in the language model creation device is realized by a computer, the processing contents of the function that the language model creation device should have are described by a program. Then, by executing this program on a computer, the processing function in the language model creating apparatus is realized on the computer.

  The program describing the processing contents can be recorded on a computer-readable recording medium. As the computer-readable recording medium, for example, any recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, and a semiconductor memory may be used. Specifically, for example, as a magnetic recording device, a hard disk device, a flexible disk, a magnetic tape or the like, and as an optical disk, a DVD (Digital Versatile Disc), a DVD-RAM (Random Access Memory), a CD-ROM (Compact Disc Read Only). Memory), CD-R (Recordable) / RW (ReWritable), etc., magneto-optical recording medium, MO (Magneto-Optical disc), etc., semiconductor memory, EEP-ROM (Electronically Erasable and Programmable-Read Only Memory), etc. Can be used.

  The program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Furthermore, the program may be distributed by storing the program in a storage device of the server computer and transferring the program from the server computer to another computer via a network.

  A computer that executes such a program first stores, for example, a program recorded on a portable recording medium or a program transferred from a server computer in its own storage device. When executing the process, the computer reads a program stored in its own recording medium and executes a process according to the read program. As another execution form of the program, the computer may directly read the program from the portable recording medium and execute processing according to the program, and the program is transferred from the server computer to the computer. Each time, the processing according to the received program may be executed sequentially. Also, the program is not transferred from the server computer to the computer, and the above-described processing is executed by a so-called ASP (Application Service Provider) type service that realizes the processing function only by the execution instruction and result acquisition. It is good. Note that the program in this embodiment includes information that is used for processing by an electronic computer and that conforms to the program (data that is not a direct command to the computer but has a property that defines the processing of the computer).

  In this embodiment, the language model creation apparatus is configured by executing a predetermined program on a computer. However, at least a part of the processing contents may be realized by hardware.

The effect of the present invention was evaluated in continuous speech recognition of simulated dialogue data in accordance with the third embodiment.
As an existing text database, large-scale text data consisting of about 100 million words collected from the Web or the like was used. Using the K-means method in which the appearance probability of each 5000 word registered in the dictionary in the text data is an input vector and the Euclidean distance between the input vectors is a distance scale, the data is classified into 50 database clusters. A cluster language model (Trigram) was constructed for each database cluster, and a synthetic cluster language model (Trigram) was constructed based on the cluster language model (Trigram). The acoustic model uses a gender-dependent hidden Markov model, and the speech data used for evaluation is about one hour. In the selection unit, the threshold value is set to 58.7, and the number of cluster language models selected from the synthesized cluster language models having a recognition rate lower than that is three. These three cluster language models were synthesized by the model integration unit to obtain an adaptive language model (Trigram).

  As a language model to be compared that is not adapted to the application, a language model (Trigram) is constructed with all the text data clusters described above, and the recognition accuracy when the speech for evaluation was recognized was 59%. On the other hand, the recognition accuracy when the same evaluation speech was recognized by the adaptive language model (Trigram) adapted for use by classifying and selecting text data was 65%. That is, the recognition accuracy can be improved by 6% by applying the language model according to the present invention.

  The present invention is useful for creating a language model used for speech recognition--for example, character input based on speech recognition or speech recognition of a dialogue system.

The figure which shows the hardware structural example of a language model creation apparatus. The block diagram which shows the function structural example of the language model creation apparatus concerning 1st Embodiment. The figure which shows the processing flow of the language model creation process concerning 1st Embodiment. The block diagram which shows the function structural example of a model evaluation part. The block diagram which shows the function structural example of the language model creation apparatus concerning 2nd Embodiment. The figure which shows the processing flow of the language model creation process concerning 2nd Embodiment. The block diagram which shows the function structural example of the language model production apparatus concerning 3rd Embodiment. The figure which shows the processing flow of the language model creation process concerning 3rd Embodiment. The block diagram which shows the function structural example of the language model creation apparatus concerning 4th Embodiment. The figure which shows the processing flow of the language model creation process concerning 4th Embodiment.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Language model creation apparatus 110 Data classification | category part 111, 111a, 111b Text data 111-1 ... 111-N Text data cluster 113 Cluster language model creation part 114-1 ... 114-N Cluster language model 115,115b model Synthesis unit 116 Synthesis cluster language model 116-b Partial selection synthesis cluster language model 117, 117b Model evaluation unit 118 Evaluation data 119 Evaluation acoustic model 120, 120a, 120b Selection unit 121, 121b Model integration unit 123 Adaptive language model

Claims (6)

Storage means for storing a plurality of text data clusters, evaluation data that is data used for evaluation of the language model, and an acoustic model for evaluation that is an acoustic model used for evaluation of the language model;
Cluster language model creation means for creating a language model corresponding to each text data cluster (hereinafter referred to as “cluster language model”) from each text data cluster,
Model synthesis means for synthesizing each language model (hereinafter referred to as “partial selection synthesis cluster language model”) from a combination of the remaining cluster language models excluding some cluster language models from all cluster language models,
Model evaluation means for evaluating each partial selection synthesis cluster language model using the evaluation data and the evaluation acoustic model, and outputting an evaluation result of each partial selection synthesis cluster language model;
Of the above evaluation results, the cluster language model excluded in the synthesis of the partially selected synthesis cluster language model that gave the low evaluation result was selected, and the cluster language model was created when there was one selected cluster language model Selection means for outputting as a language model ;
A language model creation device comprising model integration means for synthesizing and outputting one language model from a plurality of the selected cluster language models . The storage means also stores a threshold value,
The lower evaluation result the selecting means is a selection criterion, walk below the upper Symbol threshold is the evaluation result is smaller than the upper Symbol threshold
The language model creation device according to claim 1 . Data classification means for dividing the input text data into a plurality of text data clusters according to the classification criteria and outputting them,
Each text data cluster stored in the storage means is output by the data classification means.
The language model creation device according to claim 1 or claim 2 , wherein The storage means of the language model creation device stores a plurality of text data clusters, evaluation data that is data used for evaluation of the language model, and an evaluation acoustic model that is an acoustic model used for evaluation of the language model. And
A cluster language model creating means for creating a language model corresponding to each text data cluster (hereinafter referred to as a “cluster language model”) from each text data cluster;
The model synthesizing means of the language model creation device is referred to as a synthesized cluster language model (hereinafter referred to as “partial selection synthesized cluster language model”) from combinations of the remaining cluster language models excluding some cluster language models from all cluster language models. ) To synthesize each,
The model evaluation unit of the language model creation device evaluates each partial selection synthesis cluster language model using the evaluation data and the evaluation acoustic model, and outputs an evaluation result of each partial selection synthesis cluster language model. A model evaluation step;
When the selection means of the language model creation device selects a cluster language model excluded in the synthesis of the partial selection synthesis cluster language model that gave a low evaluation result among the above evaluation results, and the selected cluster language model is one A step of outputting the cluster language model as a created language model to
A language model creation method, comprising: a model integration step in which the model integration unit of the language model creation apparatus synthesizes and outputs one language model from a plurality of the selected cluster language models. Language modeling program for causing a computer to function claims 1 as the language model creating apparatus according to claim 3. A computer-readable recording medium on which the language model creation program according to claim 5 is recorded.

Images