JP2016024325A - Language model generation device, program for the same, and voice recognition device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To generate a language model adapted to a topic, showing wide variations in expression, and highly accurate.SOLUTION: A language model generation device includes: a substitution table storage part which stores a relation between a word and a synonym substituted by the word; a synonym substitution part which reads from the substitution table storage part the synonym of the word included in an n word chain (n is an integer) extracted from text data and generates a new n word chain by substituting the word in the n word chain with the synonym; and a frequency estimation part which estimates appearance frequency of the n word chain after substitution on the basis of appearance frequency of the n word chain before substitution. The substitution language model generation part generates a substitution language model by using the appearance frequency estimated by the frequency estimation part.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a language model generation device, a program thereof, and a speech recognition device.

  In the production of a program such as a television broadcast, a technique for automatically creating a caption text of a program using a voice recognition process is used. This technology recognizes audio produced by program performers, announcers, narrators, etc., outputs text, and uses it for subtitle production.

  Speech recognition processing generally includes an acoustic model that represents acoustic features of language elements such as phonemes for input speech data, and a language model that represents characteristics of appearance frequency of a string of language elements such as characters and words. Is used to obtain the maximum likelihood text by statistical calculation.

  Non-Patent Document 1 describes the following matters. That is, “1. Introduction” describes that a method for constructing a highly versatile language model and adapting a language model using a small amount of task matching data has been proposed. In addition, it is described that there is a technique for automatically generating data with a high degree of fitness as a method for constructing a language model with a high degree of fitness. Further, “2. Language model learning by automatically generating learning data” describes that language data is automatically generated by replacing words in learning data with similar words to increase learning data. Also, in “3. Word substitution taking account of differences in usage”, in order to generate an appropriate sentence even when a word is replaced, the case frame is used as knowledge, and the dependency between nouns and verbs is used. It describes the use of relationships as context. The technique described in Non-Patent Document 1 uses this context to select a replacement word.

Nobuya Shimizu, Daisuke Saito, Masayuki Suzuki, Nobuaki Hamamatsu, Keikichi Hirose, Automatic sentence generation for training language models based on word replacement considering the changes of word usage), National Conference of the Japanese Society for Artificial Intelligence (CD-ROM), Vol. 24, 2G1-OS3-4, 2010

  Some broadcast programs deal with a wide range of genres. In general, information programs have a large topic variation, so it is necessary to adapt the vocabulary and language model in order to produce subtitles by voice recognition processing. For this purpose, it is conceivable to use a script or progress table of a broadcast program. Since the program script and the progress table briefly describe the performers, place names, and information shown in the program, it is effective to use them as information for registering unknown words. However, since the notation in the script does not necessarily match the utterance of the program performer, the adaptation effect may not be sufficiently obtained even if the language model is adapted using the script or the progress table. For example, even if the full name (last name + first name) of the talent who is a program performer is listed in the script or progress sheet, the performer is called by a nickname in the program, or only the first name + first name Sometimes called with "chan". Although many scripts and progress tables are expressed in written language, program performers may speak in a different language from those written words or in a conversational manner.

  In order to reduce the influence of the mismatch between the script and progress table notation and the actual utterance, the words in the script and progress table can be replaced with synonyms to increase the paraphrase variation. Conceivable.

  However, simply generating a sentence in which words are replaced using synonyms results in variations in the number of sentences generated depending on whether or not replacement is performed. That is, the appearance frequency of word chains is biased depending on whether or not the word can be replaced. In addition, if the appearance frequency of the word chain including the portion between the pre-substitution word and the post-substitution word is apportioned, the accuracy of the language model may be lowered.

  For example, when adapting a language model by replacing words with the method described in Non-Patent Document 1, the method replaces words in units of sentences, and the number of sentences generated by the replacement increases. Therefore, while the word chain hit rate for the evaluation data increases, the prediction ability of the language model decreases (perplexity increases). One reason for this is that when there are a plurality of replacement candidates in a sentence, the number of generated sentences increases, which also affects the frequency of word chain including unreplaced words. Another reason is that, in the part where the words in the word chain are replaced, the frequency of the replacement source word is divided by the words before and after the replacement, and the chain frequency of the replacement source word is reduced. . On the other hand, in the method of Non-Patent Document 1, there is a restriction on the number of sentence generations.

  The present invention has been made on the basis of the above problem recognition. For example, a language resource suitable for a topic such as a script or a progress table can be used to cope with variations in expression, and a topic adaptation language with high accuracy. The present invention provides a language model generation device for generating a model, a program thereof, and a speech recognition device to which the language model generation device is applied.

  In order to solve the above problem, a language model generation device according to an aspect of the present invention includes a replacement table storage unit that stores a relationship between a word and a synonym that can be replaced with the word, and an n word extracted from text data. Synonym substitution that reads a synonym of a word included in a chain (n is a positive integer) from the replacement table storage unit and generates a new n-word chain by replacing the word in the n-word chain with the synonym A frequency estimation unit that estimates the appearance frequency of the n-word chain after replacement based on the appearance frequency of the n-word chain before replacement, the appearance frequency of the n-word chain extracted from the text data, and A replacement language model generation unit that generates a replacement language model including the appearance frequency of the n word chain after replacement estimated by the frequency estimation unit.

  According to this, since the synonym replacement unit replaces the word included in the n-word chain, the variation of expression is widened. That is, it is possible to generate a language model that is strong against fluctuation of expression from relatively small language resources (input text data). Moreover, since the frequency estimation unit estimates the appearance frequency of the n word chain after replacement, a language model based on the estimation result can be generated.

Further, according to one aspect of the present invention, in the language model generation device, the frequency estimation unit may include the appearance frequency of the n word chain before replacement and the replaced word (n) of the n word chain before replacement ( The m word chain including m (replacement source word) (m is a positive integer, 1 ≦ m <≦ n) and the synonym corresponding to the m word chain in the new n word chain after replacement The appearance frequency of the n word chain after the replacement is estimated based on the ratio of the appearance probability of the m word chain after the replacement including the word.
As an example, the appearance frequency of the n word chain after replacement is estimated by multiplying the appearance frequency of the n word chain before replacement by the above ratio.

Further, according to one aspect of the present invention, in the language model generation device, the frequency estimation unit estimates the appearance frequency of the n word chain before replacement as the appearance frequency of the n word chain after replacement. It is characterized by.
With this configuration, the process for estimating the appearance frequency of the n-word chain after replacement is simplified.

  Further, according to one aspect of the present invention, in the language model generation device described above, the replacement language model generated by the replacement language model generation unit and a small scale based on the appearance frequency of n word chains extracted from the text data A model adaptation unit that generates an adaptive language model by interpolating a topic language model and a large-scale topic language model based on the appearance frequency of n word chains in a large-scale language resource of a general topic It is characterized by.

  According to another aspect of the present invention, there is provided the language model generation device, an acoustic model storage unit that stores acoustic feature data of a language element as an acoustic model, and the adaptive language model generated by the model adaptation unit. The speech recognition apparatus includes a recognition processing unit that performs input speech recognition processing by using the acoustic model read from the acoustic model storage unit while being used as a language model.

  Further, according to one aspect of the present invention, a replacement table storage unit that stores a relationship between a word and a synonym that can be replaced with the word, a word included in an n-word chain (n is a positive integer) extracted from text data A synonym replacement unit that reads a synonym from the replacement table storage unit and generates a new n word chain by replacing the word in the n word chain with the synonym, and an appearance frequency of the n word chain before replacement The frequency estimation means for estimating the appearance frequency of the n word chain after replacement, the appearance frequency of the n word chain extracted from the text data, and the n word after replacement estimated by the frequency estimation means This is a program for causing a computer to function as replacement language model generation means for generating a replacement language model including the occurrence frequency of a chain.

According to the present invention, a language model suitable for a specific topic can be created based on text data. In addition, since the synonym replacement unit replaces words included in the n-word chain, variations in expression are widened. That is, it is possible to generate a language model that is strong against fluctuation of expression from relatively small language resources (input text data). Moreover, since the frequency estimation unit estimates the appearance frequency of the n word chain after replacement, a language model based on the estimation result can be generated.
As a result, a language model adapted to a specific topic or utterance style can be created.

It is a functional block diagram which shows schematic structure of the replacement language model production | generation apparatus by embodiment of this invention. It is a functional block diagram which shows schematic structure of the speech recognition apparatus using the substitution language model production | generation apparatus by the embodiment. It is the schematic which shows the structure and example of a replacement table which the replacement table memory | storage part by the same embodiment memorize | stores. It is a flowchart which shows the procedure of the process in which the replacement language model production | generation apparatus by the embodiment produces | generates a language model. It is a flowchart which shows the more detailed process sequence in the replacement language model production | generation apparatus by the embodiment.

Hereinafter, embodiments of the present invention will be described with reference to the drawings. In this embodiment, a three-word chain is used as a specific example of the n-word chain.
[First Embodiment]
FIG. 1 is a functional block diagram showing a schematic configuration of a replacement language model generation device according to the present embodiment. As shown in the figure, the replacement language model generation device 1 includes a text data acquisition unit 11, a triple generation / frequency calculation unit 12, a synonym replacement unit 13, a frequency estimation unit 14, a replacement language model generation unit 15, It includes a replacement language model storage unit 16, a model adaptation unit 17, a replacement table generation unit 21, a replacement table storage unit 22, a small topic language model storage unit 31, and a large topic language model storage unit 32. Composed.

The text data acquisition unit 11 acquires the text data 91 from the outside. The text data 91 is, for example, a script or progress table data of a broadcast program, and the contents of the text belong to a specific topic.
The triplet generation / frequency calculation unit 12 extracts a three-word chain from the text data acquired by the text data acquisition unit 11 and calculates an appearance frequency for each type of the three-word chain.

  The synonym replacement unit 13 reads a synonym of a word included in the three-word chain extracted from the text data from the replacement table storage unit 22, and replaces the word in the three-word chain with the synonym. Generate a three word chain.

  The frequency estimation unit 14 estimates the appearance frequency of the three-word chain after replacement based on the appearance frequency of the three-word chain before replacement. More specifically, the frequency estimation unit 14 estimates the appearance frequency of the three-word chain after replacement based on the appearance frequency of the three-word chain before replacement and a predetermined ratio described below. The ratio refers to the replaced word in the three word chain with respect to the appearance probability of the m word chain (m is a positive integer, 1 ≦ m ≦ 3) including the replaced word in the three word chain. It is the ratio of the appearance probability of m word chain including. In other words, the ratio is a value with the appearance probability of the m word chain including the replacement source word as the denominator and the appearance probability of the m word chain after replacement corresponding to the m word chain as the numerator. For example, the frequency estimation unit 14 estimates the appearance frequency of the three-word chain after replacement by multiplying the appearance frequency of the three-word chain before replacement by the value of this ratio. In the present embodiment, m = 1. Using a three-word chain is sufficient as the number of chains and is appropriate. Moreover, the scale of the language model data does not become too large. Further, it is difficult for a statistical error to appear in the appearance frequency. In order to estimate the appearance frequency of the three-word chain after replacement, it is appropriate as the scale of calculation to use the ratio of the appearance probability (ratio before and after replacement) of the word unigram (m = 1). Note that a general case of n-word chain that is not three-word chain will be described later (Modification 2).

The replacement language model generation unit 15 generates a replacement language model including the appearance frequency of the n word chain extracted from the text data 91 and the appearance frequency of the n word chain after replacement estimated by the frequency estimation unit 14. .
The replacement language model storage unit 16 stores the replacement language model generated by the replacement language model generation unit 15.

The model adaptation unit 17 includes a replacement language model generated by the replacement language model generation unit 15, a small topic language model based on the appearance frequency of the three-word chain extracted from the text data 91, and a large scale based on a general topic. An adaptive language model is generated by interpolating a large-scale topic language model based on the appearance frequency of a three-word chain in a simple language resource. The model adaptation unit 17 reads the small topic language model from the small topic language model storage unit 31. The model adaptation unit 17 reads the large-scale topic language model from the large-scale topic language model storage unit 32. The model adaptation unit 17 writes the generated adaptive language model in the adaptive language model storage unit 51.
In other words, the model adaptation unit 17 linearly interpolates a plurality of language models (including a language model based on a specific topic and having an expression variation using substitution with synonyms), thereby Adapt.

The replacement table generation unit 21 generates a replacement table based on the data read from the external synonym pair database 92 and writes the replacement table in the replacement table storage unit 22.
The replacement table storage unit 22 stores a correspondence relationship between a word (replacement source word) and a synonym (replacement word) that can be replaced with the word (replacement source word) as a table.

  The small topic language model storage unit 31 stores a language model (small topic language model) based on the appearance frequency calculated by the triplet generation / frequency calculation unit 12. That is, the small-scale topic language model storage unit 31 stores a language model based on the text data 91. That is, this small topic language model represents the feature of the appearance of a three-word chain in a specific topic.

  The large-scale topic language model storage unit 32 stores a general large-scale language model without topic bias. In the present embodiment, the large-scale topic language model storage unit 32 stores the appearance frequency of the three-word chain and also stores the appearance probability of the word unigram.

  FIG. 2 is a functional block diagram showing a schematic configuration of a speech recognition device using the replacement language model generation device 1. As illustrated, the speech recognition device 2 includes a replacement language model generation device 1, an adaptive language model storage unit 51, a pronunciation dictionary storage unit 52, an acoustic model storage unit 53, an input speech acquisition unit 61, and a recognition process. A unit 62 and a recognition result output unit 63 are configured.

The adaptive language model storage unit 51 stores the adaptive language model generated by the model adaptation unit 17 in the replacement language model generation device 1.
The pronunciation dictionary storage unit 52 stores pronunciation dictionary data for each word. The pronunciation dictionary data is obtained in advance from an external database.
The acoustic model storage unit 53 stores acoustic feature data for each language element (phonemes, words, etc.) as an acoustic model.

The input voice acquisition unit 61 acquires input voice data to be subjected to recognition processing from the outside, and supplies it to the recognition processing unit 62.
The recognition processing unit 62 uses the adaptive language model generated by the model adaptation unit 17 as a language model, and uses the acoustic model read from the acoustic model storage unit 53 by using the pronunciation dictionary storage unit 52, thereby Perform recognition processing. The recognition processing unit 62 receives input voice data from the input voice acquisition unit 61. In addition, the recognition processing unit 62 passes the recognition result text data to the recognition result output unit 63.
The recognition result output unit 63 outputs the text data of the recognition result passed from the recognition processing unit 62.

  FIG. 3 is a schematic diagram showing a configuration of the replacement table stored in the replacement table storage unit 22 and data examples. As illustrated, the replacement table stores a replacement source word and a replacement word in association with each other. The replacement source word is a word to be replaced, in other words, a word to be replaced. The replacement word is a new word used for replacement, in other words, a word that replaces the replacement source word. As an example, the replacement table has a two-dimensional tabular data structure. There can be one or more replacement words for one replacement source word. In the figure, the replacement word column stores a plurality of words, which are separated by a slash “/”. For example, the words “television broadcasting”, “television”, “television”, “television set”, “television receiver”, and “receiver” are registered in this substitution table as substitution words for the substitution source word “television”. Has been. This indicates that the replacement source word “TV” can be replaced with any of these replacement words. The same applies to other replacement source words.

Next, a procedure for processing for generating a replacement language model and further generating an adaptive language model will be described.
FIG. 4 is a flowchart illustrating a processing procedure performed by the replacement language model generation device 1. Hereinafter, the operation of the apparatus will be described with reference to this flowchart.

  First, in step S1, the replacement table generation unit 21 reads data from an external synonym pair database 92, and generates a replacement table based on the data. The replacement table generation unit 21 generates a replacement table so that synonyms have a relationship between a replacement source word and a replacement word. As an example of the synonym pair database 92, “Japanese WordNet synonym pair database” published by the National Institute of Information and Communications Technology (NICT) can be used.

  Next, in step S <b> 2, the replacement table generation unit 21 additionally registers the pronunciation data of the words read from the external synonym pair database 92 in the pronunciation dictionary storage unit 52.

  Next, in step S3, the large-scale topic language model is written in the large-scale topic language model storage unit 32 to be usable. The large-scale topic language model is a language model constructed from language resources (newspaper articles, magazine articles, web documents, etc.) related to general-purpose topics, without being biased toward specific topics. The large-scale topic language model has a value of (appearance probability) for each N-gram of a word. In the present embodiment, a chain of three words is used, and hereinafter referred to as “three-word chain”.

  Next, in step S4, the text data acquisition unit 11 reads the text data 91. The text data 91 is obtained from a program script or a progress table.

  Next, in step S5, the triple generation / frequency calculation unit 12 generates a triple (three word chain) from the text data read by the text data acquisition unit 11, and calculates the appearance frequency for each type of three word chain. To do. The triplet generation / frequency calculation unit 12 constructs a small topic language model based on the calculated appearance frequency. The small-scale topic language model has a value of (appearance probability) for each N-gram of a word. The triplet generation / frequency calculation unit 12 writes the constructed small topic language model in the small topic language model storage unit 31.

  Next, in step S <b> 6, the synonym replacement unit 13 replaces words included in the three-word chain obtained by the triplet generation / frequency calculation unit 12 using the replacement table read from the replacement table storage unit 22. , Generate a three-word chain after replacement.

Next, in step S <b> 7, the frequency estimation unit 14 estimates the appearance frequency for the three-word chain expanded by the synonym replacement unit 13. At this time, the frequency estimation unit 14 uses the appearance probability read from the large-scale topic language model storage unit 32. A specific method of frequency estimation will be described later. Here, the “extended three-word chain” is a three-word chain obtained by the replacement in step S6. That is, when at least one of the words included in the three-word chain is replaced with a synonym, it means that variations in the notation on the face are increased.

  Next, in step S <b> 8, the replacement language model generation unit 15 generates a replacement language model based on the frequency estimated by the frequency estimation unit 14 and writes the replacement language model in the replacement language model storage unit 16. A specific method for estimating the replacement language model will be described later.

  Next, in step S9, the model adaptation unit 17 replaces the large-scale topic language model read from the large-scale topic language model storage unit 32 with the small-scale topic language model read from the small-scale topic language model storage unit 31. Model adaptation processing is performed based on the replacement language model read from the language model storage unit 16. In the model adaptation process, linear interpolation based on the appearance probability value read from each language model is performed for the appearance probability of a certain three-word chain. The language model generated by the model adaptation unit 17 is an adaptive language model adapted to a specific topic corresponding to the input text data 91. The model adaptation unit 17 writes the generated adaptive language model in the adaptive language model storage unit 51.

  As described with reference to FIG. 2, the adaptive language model generated by the processing procedure as described above and written in the adaptive language model storage unit 51 is read out by the recognition processing unit 62 and is used for recognition processing. Used for That is, the recognition processing unit 62 uses the adaptive language model read from the adaptive language model storage unit 51 and the acoustic model read from the acoustic model storage unit 53 while referring to the pronunciation dictionary read from the pronunciation dictionary storage unit 52. The input voice is recognized. The recognition result output unit 63 outputs the recognition result obtained by the recognition processing unit 62. Specifically, as described above, the recognition processing unit 62 applies the language model and the acoustic model, and performs statistical processing, thereby outputting the maximum likelihood text for the input speech acquired by the input speech acquisition unit 61. To do. Note that the recognition processing itself by the recognition processing unit 62 can be performed using existing technology.

Next, a more detailed processing procedure for generating a replacement language model will be described.
FIG. 5 is a flowchart showing a more detailed procedure in the replacement language model generation device 1. The procedure shown in FIG. 5 explains the processing from step S5 to step S9 in FIG. 4 in more detail.

First, in step S51, the triplet generation / frequency calculation unit 12 extracts a three-word chain from the topic-adapted text data acquired by the text data acquisition unit 11, and obtains the frequency of the three-word chain. For example, when a three-word chain “reinforce costume” appears twice in the text data, the three-word chain is extracted and the frequency of appearance of the three-word chain is counted,
“Costumes-Reinforcement: Frequency 2”
Is temporarily generated. Similarly, the frequency is obtained for all three word chains in the text data.

  In step S52, the synonym replacement unit 13 replaces words included in the three-word chain obtained in step S51 above. Specifically, the synonym replacement unit 13 acquires a set of three word chains (with frequency information) generated by the triplet generation / frequency calculation unit 12, and uses each of the words included in each three word chain as a key. The replacement table stored in the replacement table storage unit 22 is searched. If the key word corresponds to a replacement source word in the replacement table, one or more replacement words corresponding to the replacement word are read from the replacement table, and these three replacement words are used in the three-word chain. Replace contained words. As a result, a new three-word chain is generated. When there are a plurality of replacement words corresponding to a certain replacement source word, the synonym replacement unit 13 performs replacement using each replacement word and generates a new three-word chain. Moreover, the synonym replacement part 13 produces | generates a new 3 word chain | strand by the combination of all substitution, when the some of the words contained in the original 3 word chain | part can be replaced. For example, the original three-word chain is “A-B-C” (A, B, and C are words respectively), the word A can be replaced in five ways, and the word B can be replaced in three ways. In the case where the word C can be replaced in four ways, normally, 119 ways of substitution are performed with (5 + 1) × (3 + 1) × (4 + 1) −1. In other words, in this case, the original one three-word chain and the newly generated 119 three-word chains are combined and expanded to 120 three-word chains.

As a specific example, the replacement table illustrated in FIG. 3 is applied to a three-word chain “reinforcing costume”. The word “costume” corresponds to the replacement source word, and there are “sales”, “kimono”, “wear”, “clothing”, “clothes” (following) as replacement words corresponding to the replacement source word. To do. Then, the synonym replacement unit 13 generates a new three-word chain using all replacement words. The original 3-word chain is
“Costumes-Reinforcement: Frequency 2”
Therefore, the newly generated 3-word chain is
"Observations-Reinforcement: Frequency 2"
"Kimono-Reinforcement: Frequency 2"
“Wear-reinforcement: Frequency 2”
“Clothing-reinforcement: Frequency 2”
"Clothing-reinforced-frequency 2" (followed)
It is. Here, the synonym replacement unit 13 takes over the appearance frequency of the original three-word chain and applies it to the newly generated three-word chain. In the above example, the frequency is 2.

Next, in step S53, the frequency estimation unit 14 estimates the frequency of the three-word chain newly generated by the synonym replacement unit 13.
In the case where there is only one replacement target word in the word chain, specifically, the frequency estimation unit 14 calculates the frequency C _syn of the three-word chain newly generated by replacement by the following formula (1 ). In other words, the frequency C _syn is a frequency obtained by estimation.

Here, C _orig is the frequency of the original three-word chain before replacement, and is calculated by the triplet generation / frequency calculation unit 12. Further, w _orig and w _syn are a replacement source word and a replacement word corresponding thereto. Further, p (w _orig ) and p (w _syn ) are word unigram probabilities of the replacement source word and the replacement word, respectively. The frequency estimation unit 14 acquires the values of p (w _orig ) and p (w _syn ) from the large-scale topic language model storage unit 32.

Even if the replacement source word w _orig and the replacement word w _syn are registered in association with each other in the replacement table, there is a case where any of these is an unknown word and its appearance probability cannot be obtained from the language model. For example, when the replacement source word w _orig is an unknown word, C _syn = 10 is set regardless of the calculation in the above equation (1). Further, when the replacement word w _syn is an unknown word, this is also set to C _syn = 0.01 regardless of the calculation in the equation (1).

In the above description, the case where only one word in the three-word chain is to be replaced has been described. However, depending on the type of the three-word chain, a plurality of words included in the word may be subject to replacement. At most, three words can be replaced. In this way, when a plurality of words in the three-word chain are replaced, the value of C _syn in the case where each one word is replaced is calculated by the previous equation (1), The minimum value among the values of C _syn of is set as the frequency of the three-word chain after replacement. That is, the frequency estimation unit 14 calculates the frequency C _syn by the following equation (2).

Here, i is an index for a word to be replaced among words included in the three-word chain. P (w _{orig, i} ) is a word unigram probability of the i-th replacement source word in the three-word chain. P (w _{syn, i} ) is the word unigram probability of the i-th replacement word. The unigram probability of each word is read from the large-scale topic language model storage unit 32 and used for calculation.

Here, a calculation example of the frequency of one three-word chain obtained by word replacement will be described. Taking the case where the word “costume” included in the original three-word chain “costume-reinforcement” is replaced with the replacement word “kimono”, for example, the newly generated three-word chain “kimono-” The frequency C _syn of “reinforcement” is calculated as follows. That is, the replacement source word w _orig is “costume”. The replacement word w _syn is “kimono”. The word unigram probabilities of each of these words are stored in the large-scale topic language model storage unit 32 and are as follows.
p (w _orig ) = 2.81 × 10 ⁻⁵
p (w _syn ) = 4.07 × 10 ⁻⁵
That means
p (w _syn ) / p (w _orig ) = 1.45
And also
C _orig = 2
Because
C _syn = 2 × 1.45 = 2.90
It is.

  In order to make it easy to handle as a value, all three-word chain frequencies are uniformly multiplied by 100 to obtain an integer value. In other words, the frequency of the upper three-word chain “kimono-reinforce” is also set to 290 by multiplying 2.90 by 100.

  In step S54, the replacement language model generation unit 15 constructs a replacement language model using the three-word chain frequency calculated as described above, and writes the replacement language model in the replacement language model storage unit 16.

  In step S55, the model adaptation unit 17 then reads the large-scale topic language model read from the large-scale topic language model storage unit 32, the small-scale topic language model read from the small-scale topic language model storage unit 31, and the replacement language. The topic adaptive language model is calculated by linear interpolation using the replacement language model read from the model storage unit 16. Specifically, the model adaptation unit 17 performs calculation according to the following expression (3) to calculate an adaptive language model.

Here, p (w _i ) is a probability value after adaptation for a certain three-word chain, and is obtained by interpolation calculation. In addition, p _L (w _i ) is a probability value in the large-scale topic language model for the three-word chain. Further, p _S (w _i ) is a probability value in the small topic language model for the three-word chain. P _R (w _i ) is a probability value in the replacement language model for the three-word chain. Α, β, and γ are respective weighting factors. These coefficients α, β, and γ are appropriately set in advance. As an example, an optimum value of these coefficients α, β, γ, or a value equivalent to the optimum value can be obtained by machine learning processing based on teacher data.

  In addition, you may make it implement | achieve the function of the replacement language model production | generation apparatus and speech recognition apparatus in embodiment mentioned above with a computer. In that case, the program for realizing these functions may be recorded on a computer-readable recording medium, and the program recorded on the recording medium may be read into a computer system and executed. Here, the “computer system” includes an OS and hardware such as peripheral devices. The “computer-readable recording medium” refers to a storage device such as a flexible disk, a magneto-optical disk, a portable medium such as a ROM and a CD-ROM, and a hard disk incorporated in a computer system. Furthermore, a “computer-readable recording medium” dynamically holds a program for a short time, like a communication line when transmitting a program via a network such as the Internet or a communication line such as a telephone line. In this case, a volatile memory inside a computer system serving as a server or a client in that case may be included, and a program that holds a program for a certain period of time. The program may be a program for realizing a part of the above-described functions, or may be a program that can realize the above-described functions in combination with a program already recorded in a computer system.

  By using the present embodiment, it is possible to construct a language model corresponding to variations of expressions as well as expressions that appear in input text data.

  That is, programs such as information programs that deal with a wide range of genres have different topics for each broadcast program, so it is necessary to adapt the vocabulary and language model used for speech recognition. As this adaptive language model, a language model in which a general large-scale language model without topical bias and a small-scale language model for each program using a script or a progress table of each broadcast program is useful. In the script / progress table, the performers, place names, and information shown in the program are succinctly written, which is effective for registering unknown words, etc. In some cases, the effect of adaptation is not sufficient. In addition, many scripts and progress tables are expressed in written language, and performers may utter in different phrases and conversations from the scripts and progress tables. In order to reduce the mismatch between the script / progress table and the utterance, more effective adaptation data can be generated by replacing words in the script / progress table with synonyms and increasing paraphrase variations. According to this embodiment, even when a word is replaced with a synonym, the frequency estimation unit can estimate an appropriate frequency based on the appearance frequency of the word chain before replacement. Therefore, the accuracy of the language model can be improved. In this embodiment, based on synonym substitution from text data such as a script and a progress table, the word chain variation is increased only between the preceding and following n sets of words, and prior knowledge (n in a large-scale topic language model, n The frequency of occurrence of the partial chain of the word chain (typically, the appearance probability of the word unigram) is used to estimate the frequency of the word chain including the replacement word. This makes it possible to generate a topic adaptive language model for improving speech recognition performance.

  Although the embodiment has been described above, the present invention can also be implemented in the following modified example. In addition, you may implement combining some of the following modifications.

[Modification 1]
In the above embodiment, a language model based on a three-word chain is used. In this modification, not only three word chains but generally n word chains (n = 1, 2, 3, 4,...) Are used. That is, the language model is data that statistically represents the appearance frequency of n word chains. Also, the triplet generation / frequency calculation unit 12, the synonym replacement unit 13, the frequency estimation unit 14, the replacement language model generation unit 15, the model adaptation unit 17, the recognition processing unit 62, and the like are not limited to the three-word chain, The word chain is processed.

[Modification 2]
In the above embodiment, when estimating the frequency of the three-word chain newly generated by the replacement, the frequency estimation unit 14 uses the equation (1) to calculate the word unigram probability p (w _orig ) of the replacement source word. The word unigram probability p (w _syn ) of the replacement word was used. In this modification, the ratio of the appearance probability (bigram probability) of the replacement two-word chain and the appearance probability of the replacement two-word chain is used instead of the word unigram probability ratio.

When modification 1 and modification 2 are combined, that is, when a three-word chain is generalized to an n-word chain, a language model based on the n-word chain is used, and the appearance of m-word chains before and after synonym substitution Use the probability ratio. Here, 1 ≦ m ≦ n (m and n are integers).
That is, in the present modification, the frequency estimation unit 14 modifies Equation (1) to estimate the appearance probability of the n word chain newly generated by the replacement using the ratio of the appearance probability of the m word chain. To do. Specifically, the frequency estimation unit 14 generates the appearance frequency of the n-word chain before replacement, the appearance probability of the m-word chain including the replacement source word in the n-word chain before replacement, and the new n after replacement. Based on the ratio of the appearance probability of the m word chain after replacement including the synonym corresponding to the m word chain in the word chain, the appearance frequency of the n word chain after replacement is estimated. In other words, the ratio is a value with the appearance probability of the m word chain including the replacement source word as the denominator and the appearance probability of the m word chain after replacement corresponding to the m word chain as the numerator. For example, the frequency estimation unit 14 estimates the appearance frequency of the n word chain after replacement by multiplying the appearance frequency of the n word chain before replacement by the value of this ratio.
The appearance probability of the m word chain is read from the large-scale topic language model storage unit 32.

  In the case of this modification, the following formula (4) is used instead of formula (1).

In equation (4), p (w _orig | w _m−1 , w _m−2 ,...) Is the appearance probability of the m word chain including the replacement source word. Further, p (w _syn | w _m−1 , w _m−2 ,...) Corresponds to this, and is the appearance probability of the m word chain including the replaced word.

[Modification 3]
In this modification, the frequency estimation unit 14 estimates that the appearance frequency of the three-word chain newly generated by the replacement is the same as the appearance frequency of the three-word chain before the replacement.
When combining this modification and Modification 1, the frequency estimation unit 14 estimates that the appearance frequency of the n word chain newly generated by the replacement is the same as the appearance frequency of the n word chain before the replacement. To do. In other words, the frequency estimation unit 14 uses the appearance frequency of the n word chain before replacement as it is, and estimates that it is the appearance frequency of the n word chain after replacement.

[Modification 4]
In this modification, the replacement language model generation device 1 does not include the model adaptation unit 17. In the replacement language model generation device 1, the replacement language model generation unit 15 generates a replacement language model and writes it in the replacement language model storage unit 16 as described in the above embodiment. This replacement language model is a language model specialized for a specific topic. This replacement language model is based on the appearance frequency of word chains appearing in the original text data 91, and the appearance frequency (estimated appearance frequency) of word chains obtained by replacing words included in those word chains with synonyms. ) Based on the frequency of appearance.

[Modification 5]
In the above embodiment, the script data of the broadcast program and the text of the progress table are used as the text data 91, but the text data to be used is not limited to this. In this modified example, a presentation summary or presentation material that can be obtained in advance at an academic conference or a lecture is used as the text data 91. Then, this modified example can be applied to generate a transcription text by speech recognition of the academic society or lecture. Moreover, the example of the text data 91 to utilize is not limited to these, but is arbitrary.

  As mentioned above, although embodiment of this invention and its modification were explained in full detail with reference to drawings, the concrete composition is not restricted to this embodiment, the design of the range which does not deviate from the gist of this invention, etc. included.

  The present invention can be used for voice recognition processing in general. Also, for example, it can be used for recognizing the sound of content such as TV broadcast programs whose topics vary greatly depending on the program. As an example, it can be used for automatic generation or semi-automatic generation of subtitle text for broadcasting.

1 Replacement language model generator (Language model generator)
2 Speech recognition device 11 Text data acquisition unit 12 Triple generation / frequency calculation unit 13 Synonym replacement unit 14 Frequency estimation unit 15 Replacement language model generation unit 16 Replacement language model storage unit 17 Model adaptation unit 21 Replacement table generation unit 22 Replacement table storage Unit 31 Small topic language model storage unit 32 Large topic language model storage unit 51 Adaptive language model storage unit 52 Pronunciation dictionary storage unit 53 Acoustic model storage unit 61 Input voice acquisition unit 62 Recognition processing unit 63 Recognition result output unit

Claims (6)

A replacement table storage unit that stores a relationship between a word and a synonym that can be replaced with the word;
A synonym of a word included in an n-word chain (n is a positive integer) extracted from text data is read from the replacement table storage unit, and a new n is obtained by replacing the word in the n-word chain with the synonym. A synonym replacement for generating word chains;
A frequency estimation unit that estimates an appearance frequency of the n word chain after the replacement based on an appearance frequency of the n word chain before the replacement;
A replacement language model generation unit that generates a replacement language model including the appearance frequency of the n word chain extracted from the text data and the appearance frequency of the n word chain after replacement estimated by the frequency estimation unit;
A language model generation apparatus comprising: The frequency estimation unit includes an m word chain including an occurrence frequency of the n word chain before replacement and a replacement source word of the n word chains before replacement (m is a positive integer, 1 ≦ m ≦ n) And the ratio of the occurrence probability of the replaced m word chain including the synonym corresponding to the m word chain in the new n word chain after replacement, based on the post-replacement Estimating the frequency of occurrence of n word chains of
The language model generation apparatus according to claim 1. The frequency estimation unit estimates the appearance frequency of the n word chain before replacement as the appearance frequency of the n word chain after replacement.
The language model generation apparatus according to claim 1. The replacement language model generated by the replacement language model generation unit, a small topic language model based on the appearance frequency of n word chains extracted from the text data, and n in a large-scale language resource of a general topic A model adaptation unit that generates an adaptive language model by interpolating a large-scale topic language model based on the appearance frequency of word chains,
The language model generation device according to any one of claims 1 to 3, further comprising: The language model generation device according to claim 4,
An acoustic model storage unit that stores acoustic feature data of language elements as an acoustic model;
Using the adaptive language model generated by the model adaptation unit as a language model, and using the acoustic model read from the acoustic model storage unit, a recognition processing unit for performing recognition processing of input speech;
A speech recognition apparatus comprising: A replacement table storage means for storing a relationship between a word and a synonym replaceable with the word;
A synonym of a word included in an n-word chain (n is a positive integer) extracted from text data is read from the replacement table storage means, and the word in the n-word chain is replaced with the synonym to create a new n Synonym replacement means for generating word chains,
A frequency estimating means for estimating an appearance frequency of the n-word chain after the replacement based on an appearance frequency of the n-word chain before the replacement;
A replacement language model generating means for generating a replacement language model including the appearance frequency of the n word chain extracted from the text data and the appearance frequency of the n word chain after the replacement estimated by the frequency estimation means;
As a program to make the computer function as.

Images