JP6518142B2 - Language model generation device and program thereof - Google Patents

Description

  The present invention relates to a language model generation device that mixes a plurality of language models to generate a new language model, and its program.

Conventionally, there is a method of mixing a plurality of language models (statistical language models) generated from an independent learning corpus in order to improve speech recognition accuracy (see, for example, Patent Document 1).
In this method, a linear interpolation coefficient (mixing weight) is determined using the evaluation text similar to the speech recognition target (speech content etc.) so that the generation probability of the evaluation text is maximized, or the linear interpolation coefficient is calculated. It is obtained by the Bayesian learning method, and linear sum interpolation of each language model is performed.

Here, referring to FIG. 6, a conventional general language model mixing method will be described.
As shown in FIG. 6, here, two language models (global language model 20, topic-dependent language model 40) are mixed. The global language model 20 is generated by learning in advance from large-scale learning data (large-scale corpus 200). The topic-dependent language model 40 is generated by learning in advance from small-scale learning data (topic-dependent small corpus 400) depending on the topic (topic) of the speech recognition target.

  For example, according to the conventional method (the first conventional method) described as the background art of Patent Document 1, the generation probability of the evaluation sentence H is maximized using the evaluation sentence H similar to the topic of the speech recognition target The mixed weight λ of the global language model 20 and the topic dependent language model 40 is calculated by maximum likelihood learning. Then, in the first conventional method, a mixed language model is obtained by performing weighted addition (linear sum interpolation) of the global language model 20 and the topic-dependent language model 40 using the mixing weight λ calculated by the linear sum interpolation means M. Generate 80.

Further, according to another method (second conventional method) according to the invention described in Patent Document 1, linear interpolation coefficients are obtained by Bayesian learning using evaluation sentences H smaller than the above-described first conventional method. , Linear sum interpolation to generate a mixed language model 80.
The speech recognition performed by the speech recognition apparatus 100 using the mixed language model 80 generated in this manner can improve the speech recognition accuracy for a specific topic as compared to the case where only the global language model 20 is used.

In the first conventional method, over-learning is suppressed by selecting the evaluation text H as different as possible from the corpus (large-scale corpus 200, topic-dependent small-scale corpus 400). Further, in the second conventional method, over learning is further suppressed by using Bayesian learning to reduce the number of evaluation sentences H more than the first conventional method.
Here, when the text for evaluation exists in the corpus, the overlearning strongly depends on the text (learning data) and the appearance probability of the language model is expected in other texts (unknown data). It means that it will be in the state where the appearance probability can not be obtained.

JP 2005-84179 A

In the first and second conventional methods described above, over-learning can be suppressed by selecting sentences for evaluation as different as possible from the corpus as much as possible, or selecting a small number of sentences for evaluation by using Bayesian learning. I have to.
However, as the corpus is larger, it is practically difficult to select evaluation sentences except for sentences existing in the corpus. That is, in the conventional method, the sentences in the corpus may be used as the sentences for evaluation, and there is a problem that over-learning occurs due to the sentences for evaluation.

  The present invention has been made in view of such a problem, and it is evaluated in advance whether each sentence used for evaluation sentences is appropriate as an evaluation sentence and by using an appropriate evaluation sentence, overlearning It is an object of the present invention to provide a language model generation device capable of mixing language models while suppressing the

  In order to solve the above problem, the language model generation device according to the present invention comprises a topic-dependent language model learned in advance from a learning corpus related to a speech recognition target topic, and a learning corpus having a larger amount of data than the learning corpus. A language model generation device that mixes a learned global language model and generates a mixed language model of a speech recognition target, the evaluation weight generation unit, the first mixture weight generation unit, and the first linear sum interpolation unit. The second mixture weight generation means and the second linear sum interpolation means are provided.

In such a configuration, the language model generation device evaluates the language model by using the evaluation weight generation means, with respect to the global language model, using the whole of the evaluation sentences selected in advance related to the topic of the speech recognition target. An evaluation value of (for example, perplexity) is calculated as an overall evaluation value.
Furthermore, the language model generation apparatus calculates the evaluation value for each divided sentence as an individual evaluation value using the divided sentences obtained by dividing the evaluation sentences according to the predetermined classification with respect to the global language model by the evaluation weight generation means. Do.

Then, the language model generation apparatus generates, as an evaluation weight, the degree of suitability as an evaluation sentence for each divided sentence by the evaluation weight generation means. That is, the evaluation weight generation means increases the evaluation weight of the divided sentence if the evaluation of the global language model is lower when the divided sentence is used than when the entire evaluation sentence is used, and if the evaluation is high, The evaluation weight is generated by reducing the evaluation weight. This means that if the evaluation of the global language model is higher when the divided sentences are used, it is highly likely that the divided sentences are already included in the learning corpus used to learn the global language model, resulting in over-learning. To prevent
As described above, by evaluating the evaluation sentences in division units, the language model generation device can evaluate the degree of overlearning for each classification unit of evaluation sentences.

Then, the language model generation device generates, by the first mixture weight generation means, the plurality of individual language models previously learned from the plurality of individual learning corpuses constituting the learning corpus used for learning the global language model, To generate a mixture weight that maximizes the log likelihood when performing linear sum interpolation at the rate of evaluation weight.
As described above, the first mixed weight generation unit suppresses the occurrence of overlearning by adding the evaluation weight as a ratio of performing linear sum interpolation on the language model, and increases the generation probability of the divided sentences in the evaluation sentence. Working mixing weights can be generated.

  Then, the language model generation apparatus generates a mixed global language model by performing linear sum interpolation on a plurality of individual language models at a ratio of the mixed weight generated by the first mixed weight generation unit by the first linear sum interpolation unit. . That is, this mixed global language model is a language model in which the connection probability of the word to the speech recognition target topic is higher than that of the global language model.

Also, the language model generation device is a mixture that maximizes the log likelihood when performing linear sum interpolation between the mixed global language model and the topic dependent language model at the rate of the evaluation weight for each divided sentence by the second mixture weight generation means. Generate weights.
As described above, the second mixed weight generation unit suppresses the occurrence of overlearning by adding the evaluation weight as a ratio of linear sum interpolation of the language model, and increases the generation probability of the divided sentences in the evaluation sentence. Working mixing weights can be generated.

  Then, the language model generation device performs speech recognition by performing linear sum interpolation on the mixed global language model and the topic dependent language model at the ratio of the mixed weight generated by the second mixed weight generation means by the second linear sum interpolation means. Generate a mixed language model of the object.

  Further, in order to solve the above problem, the language model generation device according to the present invention includes a topic-dependent language model learned in advance from a learning corpus related to a speech recognition target topic, and a learning corpus having a larger amount of data than the learning corpus. A language model generation apparatus for generating a mixed language model of the speech recognition target by mixing the global language model learned in advance with the evaluation model, the evaluation weight generation means, the mixture weight generation means, and the linear sum interpolation means It had composition.

In such a configuration, the language model generation device evaluates the language model by using the evaluation weight generation means, with respect to the global language model, using the whole of the evaluation sentences selected in advance related to the topic of the speech recognition target. An evaluation value of (for example, perplexity) is calculated as an overall evaluation value.
Furthermore, the language model generation apparatus calculates the evaluation value for each divided sentence as an individual evaluation value using the divided sentences obtained by dividing the evaluation sentences according to the predetermined classification with respect to the global language model by the evaluation weight generation means. Do.

Then, the language model generation apparatus generates, as an evaluation weight, the degree of suitability as an evaluation sentence for each divided sentence by the evaluation weight generation means. That is, the evaluation weight generation means increases the evaluation weight of the divided sentence if the evaluation of the global language model is lower when the divided sentence is used than when the entire evaluation sentence is used, and if the evaluation is high, The evaluation weight is generated by reducing the evaluation weight.
As described above, by evaluating the evaluation sentences in division units, the language model generation device can evaluate the degree of overlearning for each classification unit of evaluation sentences.

Then, the language model generation device uses the mixture weight generation means to obtain the mixture weight that maximizes the log likelihood when performing linear sum interpolation between the global language model and the proposition dependent language model at the rate of the evaluation weight for each divided sentence. Generate
As described above, the mixing weight generation means suppresses the occurrence of overlearning and acts in a direction to increase the generation probability of the divided sentences in the evaluation sentence by adding the evaluation weight as a ratio of performing linear sum interpolation on the language model. Mixed weights can be generated.

Then, the language model generation device performs linear sum interpolation on the global language model and the topic dependent language model at a ratio of the mixture weight generated by the mixture weight generation means by the linear sum interpolation means, and generates a mixed language model as a speech recognition target. Generate
Thus, the language model generation device suppresses overlearning and generates a language model suitable for a speech recognition target when mixing an existing global language model and a small-scale topic-dependent language model targeted for speech recognition. be able to.
The language model generation device can be operated by a language model generation program for causing a computer to function as each of the above-described means.

The present invention exhibits the following excellent effects.
According to the present invention, in the language model to be mixed, it is evaluated whether or not overlearning is obtained for each of the sentences classified in advance for the evaluation sentences to generate an evaluation weight, and the language model is generated using the evaluation weight. Calculate mixing weights for mixing. Therefore, the present invention can generate language models while suppressing overlearning. Further, by using the language model in which overlearning is suppressed according to the present invention for speech recognition, speech recognition with higher recognition accuracy than before can be performed.

BRIEF DESCRIPTION OF THE DRAWINGS It is a block block diagram which shows the structure of the language model production | generation apparatus which concerns on 1st Embodiment of this invention. It is explanatory drawing for demonstrating the outline | summary of the language model used with the language model production | generation apparatus which concerns on 1st Embodiment of this invention, Comprising: (a) is a language model learned from a large scale corpus, (b) is a topic We show a language model learned from a dependent small corpus. It is a block diagram which shows the structure of the speech recognition system which performs speech recognition using the language model produced | generated by the language model production | generation apparatus which concerns on 1st Embodiment of this invention. It is a flowchart which shows operation | movement of the language model production | generation apparatus which concerns on 1st Embodiment of this invention. It is a block block diagram which shows the structure of the language model production | generation apparatus which concerns on 2nd Embodiment of this invention. It is explanatory drawing for demonstrating the mixing method of the conventional language model.

Hereinafter, embodiments (first and second embodiments) of the present invention will be described with reference to the drawings.
In the first embodiment, a plurality of language models learned for each of the original materials constituting the large-scale corpus are mixed to be suitable for the language of the speech recognition target, and a topic of the speech recognition target learned from the small-scale corpus ( It is a form that mixes language models depending on the topic).
In the second embodiment, the first embodiment is simplified, and one language model learned from a large corpus is mixed with a language model dependent on a topic for speech recognition to be a small corpus. is there.

Here, the language model is a probability model (statistical language model) which gives the probability that it is a sentence in an arbitrary character string. The language model, for example, N grams a language model, as shown in the following equation (1), a word string w _{1 w} 2 _... w conditional probability of a word w _i after the _i-1 appears (N It is a model given by gram probability).

  In addition, in FIG. 1 etc., Formula (1) is simplified and it describes with P (w | h). That is, h is a word string appearing immediately before the word w.

First Embodiment
[Configuration of Language Model Generation Device]
First, the configuration of the language model generation device 1 according to the first embodiment of the present invention will be described with reference to FIG.

  The language model generation device 1 includes a plurality of topic dependent language models 40 learned in advance from a learning corpus related to a topic of speech recognition target, and a plurality of independent learning corpuses (large scale corpus) having a larger data volume than the learning corpus. Mixed with a plurality of individual language models (here, as an example, a manuscript language model 30, a subtitle language model 31, and a transcription language model 32) previously learned from each learning corpus, and a language model (mixed language) for speech recognition Model 50) is generated.

In FIG. 1, the global language model 20 is a language model in which each learning corpus generated as a manuscript language model 30, a subtitle language model 31, and a transcription language model 32 is learned as one learning corpus (large scale corpus). is there.
Further, in FIG. 1, the mixed global language model 21 is a language model in the middle stage in which the language model generation device 1 generates the original language model 30, the subtitle language model 31, and the transcription language model 32 by mixing. is there.
The learning of the language model is to obtain the probability of the equation (1) from the learning corpus by a general method such as the maximum likelihood estimation method, and the detailed description is omitted here.

Here, with reference to FIG. 2, the relationship of the language model which the language model production | generation apparatus 1 mixes is demonstrated.
As shown in FIG. 2A, the global language model 20 has been learned in advance from learning data (individual learning corpus) for each of the "document", "caption" and "transcription" included in the large-scale corpus 200. It is a language model. The "original" is, for example, original data of a broadcast program such as news. Also, "subtitle" is subtitle data attached to a broadcast program. Also, "transcription" is transcription data in which the sound of a broadcast program is actually transcribed by hand. This large-scale corpus 200 is data obtained by accumulating these data (learning data) for several years, for example.
Further, as shown in FIG. 2A, the manuscript language model 30 is a language model learned in advance from the “manuscript” included in the large-scale corpus 200. The subtitle language model 31 is a language model learned in advance from “subtitles” included in the large-scale corpus 200. The transcription language model 32 is a language model learned in advance from “transcription” included in the large-scale corpus 200.

Also, as shown in FIG. 2 (b), the topic dependent language model 40 is a language model learned in advance from the topic dependent small corpus 400. This topic-dependent small-scale corpus 400 is learning data similar to a topic (topic) to be subjected to speech recognition. For example, when the target of speech recognition is speech of a sports program, the topic-dependent small corpus 400 is learning data or the like transcribed from a sports program broadcasted in the past.
Referring back to FIG. 1, the configuration of the language model generation device 1 will be described in detail.

  As shown in FIG. 1, the language model generation device 1 includes an evaluation weight generation unit 10, a mixture weight generation unit (first mixture weight generation unit 11A, a second mixture weight generation unit 11B), and a linear sum interpolation unit (first 1) linear sum interpolation means 12A, second linear sum interpolation means 12B).

The evaluation weight generation means 10 evaluates the global language model 20 using each sentence (word string) constituting the evaluation sentence H, and evaluates whether each sentence of the evaluation sentence H is appropriate as an evaluation sentence or not. It is generated as a weight.
Here, the evaluation sentence H is a sentence selected as a sentence of content related (similar) to a topic (topic) to be subjected to speech recognition. For example, when the speech recognition target is a broadcast program (information program) providing certain information, a transcript of the same information program in the past may be used as the evaluation sentence H. Here, it is assumed that the evaluation sentences input to the evaluation weight generation means 10 are H = {h ₁ ,..., H _c ,...}, And each h _c is composed of one or more sentences. For example, each h _c may be a transcription of each corner in the program of the information program. That is, it is assumed that the evaluation sentences H are classified in predetermined units, for example, one sentence or one sentence or more (sorted sentences) classified in a predetermined classification.

  The evaluation weight generation unit 10 evaluates the global language model 20 by perplexity (average number of branches) using the evaluation sentence H. The perplexity is an evaluation value that can indicate an average number of words connected next to a certain word, and the smaller the value is, the higher the accuracy of the language model is.

Specifically, the evaluation weight generation means 10 uses perplexity PP (whole evaluation value, using the whole sentence {h ₁ ,..., H _c ,. Overall perplexity PP _all ) and perplexity PP (individual evaluation value, individual perplexity PP _c ) are calculated for each individual sentence (segmented sentence) {h _c } of evaluation text H.

In this equation (2), P _global (w _i | w _{i −N + 1} ... W _i−1 ) represents the conditional probability (N-gram probability) of the global language model 20. Further, n is the number of words of the whole sentence of the evaluation sentence H when calculating the whole perplexity PP _all, and when calculating the individual perplexity PP _c , the individual divided sentences of the evaluation sentence H { h _c } is the number of words.

Then, the evaluation weight generation means 10 compares the overall perplexity PP _all with the individual perplexity PP _c . Here, if the individual perplexity PP _c is larger, that is, if the evaluation of the global language model 20 is lower when the divided sentences are used than when the entire evaluation sentence H is used, the evaluation weight generation unit 10 Increases the weight as the evaluation text of the corresponding individual divided text {h _c }. On the other hand, if the evaluation of the global language model 20 is higher when the divided sentences are used than when the entire evaluation sentence H is used, the evaluation weight generation unit 10 uses the divided sentences {h _c } as evaluation sentences. Reduce the weight of
For example, evaluation weight generating unit 10, as shown in the following equation _(3), if the PP _{C> PP all,} the evaluation weight alpha _c for sentences _{{h c}} "1", with _{PP C} ≦ PP _all If there is, the evaluation weight α _c for the divided sentence {h _c } is set to “0”.

  The evaluation weight generation means 10 associates the evaluation weight with each of the divided sentences of the evaluation sentence H, and outputs it as a weighted evaluation sentence to the first mixed weight generation means 11A and the second mixed weight generation means 11B.

  The first mixed weight generation unit 11A uses the weighted evaluation sentences (evaluation sentences, evaluation weights) generated by the evaluation weight generation unit 10 to generate a plurality of language models (the manuscript language model 30, the subtitle language model 31 and the like). A weighting factor (mixing weight) is generated to mix the transcription language model 32). The first mixed weight generation unit 11A calculates the mixed weight of each language model so that the log likelihood of the weighted evaluation text is maximized.

Specifically, first mixing weight generating unit 11A has the following formula mixture weight lambda _Genko document language model 30 to maximize the log likelihood L (4), the mixture weight lambda _Jimaku subtitle language model 31, write The mixture weight λ _kakiokoshi of the _translation language model 32 is calculated.

In the formula (4), c is the evaluation sentence _{H = {h 1, ...,} h c, ...} is an index pointing to segment text _{h c} of, C is indicative of the sentence total. In ^{_{addition, P genko (w i c |}} w i-N + 1 c ... w i-1 c) shows the conditional probability of the original language model 30 with respect to the sentence _{h c} (N-gram probability). In ^{_{addition, P jimaku (w i c |}} w i-N + 1 c ... w i-1 c) shows the conditional probability of subtitle language model 31 with respect to the division sentence _{h c} a (N-gram probability). In ^{_{addition, P kakiokoshi (w i c |}} w i-N + 1 c ... w i-1 c) shows the conditional probability of the language model 32 transcriptions for the division sentence _{h c} (N-gram probability). Also, n ^c indicates the number of words of the divided sentence h _c , and w ^c indicates the word of the divided sentence h _c .
The first mixing weight generation means 11A can generate (calculate) the mixing weights λ (λ _genko , λ _jimaku , λ _kakiokoshi ) maximizing the equation (4) by the EM algorithm or the like.
The first mixing weight generation means 11A outputs the generated mixing weights λ (λ _genko , λ _jimaku , λ _kakiokoshi ) to the first linear sum interpolation means 12A.

  The first linear sum interpolation unit 12A mixes a plurality of language models (a manuscript language model 30, a subtitle language model 31, and a transcription language model 32) using the mixing weights λ generated by the first mixing weight generation unit 11A. It is

Specifically, the first linear sum interpolation means 12A, as shown in the following equation (5), between the language model be mixed, each same word _{w i,} mixture weight _{λ (λ genko, λ jimaku,} λ _A mixed global language model 21 is generated by _performing weighted addition (linear sum interpolation) of N-gram probabilities using _kakiokoshi ). Here, P _mix (w _i | w _{i−N + 1} ... W _i−1 ) indicates the N-gram probability of the mixed global language model 21 to be generated.

The mixed global language model 21 generated in this manner is a language model in which the conditional probability for the expression of the speech recognition target is higher than that of the global language model 20.
The first linear sum interpolation means 12A writes and stores the generated mixed global language model 21 in storage means (not shown). Further, this mixed global language model 21 is referred to by the second mixed weight generation means 11B and the second linear sum interpolation means 12B described later.

The second mixed weight generation unit 11B uses the weighted evaluation sentences (evaluation sentences, evaluation weights) generated by the evaluation weight generation unit 10 to generate a plurality of language models (mixed global language model 21 and topic dependent language model). 40) to generate a weighting factor (mixing weight). The second mixed weight generation unit 11B calculates the mixed weight of each language model so that the log likelihood of the weighted evaluation text is maximized.
The method of generating mixture weights in the second mixture weight generation means 11B is the same as the first mixture weight generation means 11A except that the language model to be mixed is different.

Specifically, the second mixture weight generation means 11 B mixes the mixture weights λ _mix of the mixed global language model 21 which maximizes the log likelihood L of the following equation (6), and mixes the mixture weights λ _wadai of the topic dependent language model 40. Calculate

In the formula ^{_{(6), P mix (w}} i c | w i-N + 1 c ... w i-1 c) illustrates a conditional probability of mixing for sentence _{h c} global language model 21 (N-gram probability). In _{^{addition, P wadai (w i c |}} w i-N + 1 c ... w i-1 c) shows the conditional probability of topic dependent language model 40 with respect to the sentence _{h c} (N-gram probability). Other variables are the same as in the equation (4).
The second mixing weight generation means 11B outputs the generated mixing weights λ (λ _mix , λ _wadai ) to the second linear sum interpolation means 12B.

The second linear sum interpolation unit 12B mixes a plurality of language models (a mixed global language model 21 and a topic dependent language model 40) using the mixture weights λ generated by the second mixture weight generation unit 11B. .
The mixing method in the second linear sum interpolation means 12B is the same as the first linear sum interpolation means 12A except for the language model to be mixed.

Specifically, the second linear sum interpolation unit 12B mixes the mixing weights λ (λ _mix , λ _wadai ) for each word w _i among the language models to be mixed, as shown in the following equation (7). A mixed language model 50 is generated by using weighted addition (linear sum interpolation) of N-gram probabilities. Here, P _mix2 (w _i | w _{i−N + 1} ... W _i−1 ) indicates the N-gram probability of the mixed language model 50 to be generated.

As a result, the mixed language model 50 is a language model in which the conditional probability of the speech recognition target topic is raised with respect to the mixed global language model 21 in which the conditional probability for the expression of the speech recognition target is increased. It becomes.
The mixed language model 50 generated by the language model generation device 1 can be used in a general speech recognition device. In that case, for example, as shown in FIG. 3, the speech recognition apparatus 100 recognizes and recognizes speech using the mixed language model 50 generated by the language model generation apparatus 1 and the existing pronunciation dictionary 60 and the acoustic model 70. Output the result.

By configuring the language model generation device 1 as described above, the language model generation device 1 provides an evaluation weight to the evaluation sentence H to be used when mixing a plurality of language models, and mixes the mixed weights. Since it calculates and mixes, over-learning can be suppressed.
Further, in the language model generation device 1, since the conditional probability with respect to the expression of the speech recognition target is enhanced in the mixed global language model 21, as in the second embodiment (FIG. 5) described later, the global language model 20 As compared with the case where the topic dependent language model 40 is directly mixed, the conditional probability for the expression of the speech recognition target can be further enhanced.

  The language model generation device 1 is a computer whose illustration is omitted, evaluation weight generation means 10, first mixed weight generation means 11A, first linear sum interpolation means 12A, second mixed weight generation means 11B, second linear sum It can be operated by a program (language model generation program) to function as the interpolation means 12B.

[Operation of language model generation device]
Next, the operation of the language model generation device 1 according to the first embodiment of the present invention will be described with reference to FIG. 4 (refer to FIG. 1 for the configuration as appropriate).
First, the language model generation device 1 generates the evaluation weight of the evaluation sentence H from the global language model 20 learned using the large scale corpus by the evaluation weight generation means 10 (step S1).
Specifically, the evaluation weight generation means 10 calculates perplexity PP (overall perplexity PP _all ) using the whole sentence {h ₁ ,..., H _c ,. Perplexity PP (individual perplexity PP _c ) is calculated for each individual sentence {h _c } of sentence H (see equation (2) above). Then, if the individual perplexity PP _c is larger than the whole perplexity PP _all , the evaluation weight generation means 10 increases the weight of the corresponding individual sentence {h _c } as an evaluation sentence, Otherwise, an evaluation weight is generated so as to reduce the weight of the individual sentence {h _c } as an evaluation sentence (see the equation (3)).

Then, the language model generation device 1 uses the evaluation weight generated in step S1 by the first mixed weight generation unit 11A to read the original, the subtitle, and the original learned from the respective learning data of the transcription and transcription. A mixed weight λ (λ _genko , λ _jimaku , λ _kakiokoshi ) for mixing the language model 30, the subtitle language model 31, and the transcribed language model 32 is generated (step S2).
Specifically, the first mixing weight generation unit 11A calculates the mixing weight of the document language model 30, the subtitle language model 31, and the transcription language model 32 such that the log likelihood of the evaluation text is maximized (described above. Formula (4) reference).

Then, the language model generation device 1 mixes the original language model 30, the subtitle language model 31, and the transcription language model 32 by the first linear sum interpolation means 12A using the mixing weights calculated in step S2, and mixes them. A global language model 21 is generated (step S3).
Specifically, the first linear sum interpolation unit 12A calculates the mixture weights λ (λ _genko , λ _jimaku , λ _kakiokoshi ) of the original language model 30, the subtitle language model 31, and the transcription language model 32 calculated in step S2. The mixed global language model 21 is generated by performing weighted addition (linear sum interpolation) of N-gram probabilities using the above (see the above-mentioned equation (5)).

Then, the language model generation device 1 learns from the mixed global language model 21 generated in step S3 and the topic-dependent small-scale corpus, using the evaluation weight generated in step S1 by the second mixture weight generation unit 11B. A mixed weight λ (λ _mix , λ _wadai ) is generated to be mixed with the topic dependent language model 40 (step S4).
Specifically, the second mixed weight generation unit 11B calculates the mixed weight of the mixed global language model 21 and the topic-dependent language model 40 such that the log likelihood of the evaluation text is maximized (the above-mentioned equation (6)). reference).

Then, the language model generation device 1 mixes the mixed global language model 21 and the topic dependent language model 40 by the second linear sum interpolation means 12 B using the mixing weights calculated in step S 4, and sets the mixed language model 50. Generate (step S5).
Specifically, the second linear sum interpolation unit 12B uses the mixed weights λ (λ _mix , λ _wadai ) of the mixed global language model 21 and the topic dependent language model 40 calculated in step S4 to calculate the N-gram probability. The mixed language model 50 is generated by performing weighted addition (linear sum interpolation) (see the above equation (7)).
By the above operation, the language model generation device 1 can generate a language model in which overlearning is suppressed and the recognition accuracy of the speech recognition target is enhanced.

[Performance evaluation]
Next, an evaluation result of evaluating the language model generation device 1 will be described.
Corpus (script, subtitles, transcription) constituting the large-scale corpus 200 (see FIG. 2) used for this evaluation, and the topic-dependent small-scale corpus 400 are data used in a past broadcast program, The corpus size is shown in [Table 1] below.

  In the language model generation device 1, the manuscript language model 30, the subtitle language model 31, the transcription language model 32, and the topic dependent language model 40 which are learned using the corpus shown in [Table 1] are mixed.

First, about the effect that the language model generation device 1 mixes language models (here, the original language model 30, the subtitle language model 31, and the transcription language model 32) using the evaluation weights generated by the evaluation weight generation unit 10. explain.
In the following [Table 2], the global language model 20 (P _global (w | h)) generated by simply learning a large scale corpus and the evaluation weight are not used, and the evaluation weight α in the above equation (4) A language model (P _mixtest (w | h)) generated with _c being always “1” and a mixed global language model 21 (P _mix (w | h)) generated using the evaluation weights of the present invention Shows the perplexity value of.

As shown in [Table 2], in the language model generation device 1 according to the present invention, the mixed global language model 21 (P _mix (w | h) generated by mixing the evaluation weights of the evaluation sentences H is Compared to the language models (P _global (w | h) and P _mixtest (w | h)), the perplexity value is smaller, which indicates that a language model with high accuracy is generated.

Next, the accuracy of speech recognition using the language model (mixed language model 50) generated by the language model generation device 1 will be described.
In the following [Table 3], the word error rate when speech recognition is performed using the mixed language model 50 (P _mix2 (w | h)) generated using the evaluation weight of the present invention, and without using the evaluation weight (4) shows the word error rate when speech recognition is performed using the language model (P _mix2test (w | h)) generated with the evaluation weight α _c always set to “1” in the equation (4).

As shown in [Table 3], in the language model generation device 1 according to the present invention, the mixed language model 50 (P _mix2 (w | h) generated and mixed with the evaluation weight of the text for evaluation H is an evaluation weight. As compared with the language model (P _mix2test (w | h)) generated without using, the word error rate is smaller, and the accuracy of speech recognition can be improved.

Second Embodiment
Next, the configuration of the language model generation device 1B according to the second embodiment of the present invention will be described with reference to FIG.

  Similar to the language model generation device 1 (see FIG. 1), the language model generation device 1B adds a small scale language model (topic dependent language model 40) for speech recognition to a large scale language model (global language model 20). Are weighted and added. The language model generation device 1B is different from the language model generation device 1 (see FIG. 1) in that a plurality of language models (a manuscript language model 30, a subtitle language model 31, and a transcription language model) are learned in advance independently in a large scale corpus. 32 [see FIG. 1]) is not mixed.

  As shown in FIG. 5, the language model generation device 1B includes an evaluation weight generation unit 10, a mixture weight generation unit 11, and a linear sum interpolation unit 12. The evaluation weight generation unit 10 is the same as the configuration of the language model generation apparatus 1 described with reference to FIG.

The mixed weight generation unit 11 uses the weighted evaluation sentences (evaluation sentences, evaluation weights) generated by the evaluation weight generation unit 10 to generate a plurality of language models (global language model 20 and topic dependent language model 40). It generates weighting factors (mixing weights) to be mixed. The mixing weight generation unit 11 calculates the mixing weight of each language model so as to maximize the log likelihood of the weighted evaluation sentence.
The method of calculating the mixture weight based on the log likelihood is the same as the method of the first mixture weight generation means 11A and the second mixture weight generation means 11B described with reference to FIG.
The mixing weight generation means 11 outputs the generated mixing weights λ (λ _global , λ _wadai ) to the linear sum interpolation means 12.

The linear sum interpolation means 12 mixes a plurality of language models (the global language model 20 and the topic dependent language model 40) using the mixing weight λ generated by the mixing weight generation means 11. The linear sum interpolation means 12 outputs the generated mixed language model 50B to the outside.
The method of mixing language models using the mixing weights is the same as the method of the first linear sum interpolation means 12A and the second linear sum interpolation means 12B described in FIG. .

  As described above, the language model generation device 1 B mixes the topic dependent language model 40 generated as a corpus of the speech recognition target topic with the global language model 20 generated by the existing large-scale corpus, The recognition accuracy of the target speech can be improved. In addition, at this time, the language model generation device 1B is already included in the corpus and learned by increasing the weight of the sentence suitable for evaluation when calculating the mixture weight in each sentence of the evaluation sentence H. Over-learning can be suppressed.

  The language model generation device 1B can operate a computer (not shown) with a program (language model generation program) that functions as the evaluation weight generation means 10, the mixing weight generation means 11, and the linear sum interpolation means 12.

As mentioned above, although embodiment (1st, 2nd embodiment) of this invention was described, this invention is not limited to these embodiment, It can deform | transform variously as follows.
«Other modifications»
Here, the evaluation weight α _c generated by the evaluation weight generation means 10 is binary (“0”, “1”) as shown in the equation (3).
However, the evaluation weight generation means 10 sets the evaluation weight α _c on the basis of the difference between the entire perplexity PP _all of the whole sentence of the evaluation sentence H and the individual perplexity PP _{c of} each sentence, etc. The value may be in the range of “more than“ 1 ”or less. For example, the entire whole sentences perplexity PP _all, the evaluation weights for sentences difference obtained by subtracting the individual perplexity PP _c of each sentence is maximum "1", the evaluation weights for sentences having the smallest "0" For the evaluation weights for other sentences, values may be assigned according to the ratio of the magnitude of the difference.

Also, here, the evaluation weight generation unit 10 uses perplexity as an index for evaluating a language model.
However, the evaluation weight generation means 10 does not necessarily have to use perplexity as long as it is an index that can evaluate the language model numerically. For example, entropy (E in the above equation (2)) and log likelihood (from Σ in the above equation (2)) may be used.

  In addition, although a plurality of language models are illustrated (for example, the manuscript language model 30, the subtitle language model 31, the transcription language model 32, etc.) here, the language models to be mixed are limited to these. Absent. For example, the manuscript language model 30 may be a newspaper manuscript for several years in addition to a broadcast program manuscript.

1, 1 B language model generation device 10 evaluation weight generation means 11 mixture weight generation means 11 A first mixture weight generation means 11 B second mixture weight generation means 12 linear sum interpolation means 12 A first linear sum interpolation means 12 B second linear sum interpolation means 20 Global Language Model 21 Mixed Global Language Model 30 Manuscript Language Model (Individual Language Model)
31 Subtitle Language Model (Individual Language Model)
32 Transcript language model (individual language model)
40 Topic-Dependent Language Model 50, 50B Mixed Language Model

Claims (5)

A mixed language of speech recognition target by mixing a topic dependent language model learned in advance from a learning corpus related to a speech recognition target topic and a global language model learned in advance in a learning corpus having a larger data volume than the learning corpus A language model generation device for generating a model,
The global language model using the overall evaluation value obtained by evaluating the global language model using the whole of the evaluation sentences associated with the topic in advance and the divided sentences obtained by classifying the evaluation sentences according to predetermined classifications Evaluation weight generation means for calculating an individual evaluation value for each of the divided sentences obtained by evaluating the above, and generating a degree of appropriateness as the evaluation sentence as an evaluation weight for each of the divided sentences;
Logarithm of linear sum interpolation of a plurality of individual language models learned in advance from a plurality of individual learning corpuses constituting a learning corpus used to learn the global language model, for each of the divided sentences at the ratio of the evaluation weight First mixing weight generation means for generating mixing weights with maximum likelihood;
First linear sum interpolation means for generating a mixed global language model by performing linear sum interpolation on the plurality of individual language models at a ratio of the mixture weights generated by the first mixture weight generation means;
A second mixture weight generation unit configured to generate mixture weights that maximize the log likelihood when performing linear sum interpolation on the mixture global language model and the topic dependent language model at the ratio of the evaluation weight for each of the divided sentences;
A second linear sum interpolation that generates a mixed language model of the speech recognition target by performing linear sum interpolation on the mixed global language model and the topic dependent language model at a ratio of the mixed weight generated by the second mixed weight generation unit. Means,
A language model generation apparatus comprising: A mixed language of speech recognition target by mixing a topic dependent language model learned in advance from a learning corpus related to a speech recognition target topic and a global language model learned in advance in a learning corpus having a larger data volume than the learning corpus A language model generation device for generating a model,
The global language model using the overall evaluation value obtained by evaluating the global language model using the whole of the evaluation sentences associated with the topic in advance and the divided sentences obtained by classifying the evaluation sentences according to predetermined classifications Evaluation weight generation means for calculating an individual evaluation value for each of the divided sentences obtained by evaluating the above, and generating a degree of appropriateness as the evaluation sentence as an evaluation weight for each of the divided sentences;
Mixing weight generation means for generating a mixture weight that maximizes the log likelihood when performing linear sum interpolation between the global language model and the topic dependent language model at a rate of the evaluation weight for each of the divided sentences;
Linear sum interpolation means for performing linear sum interpolation on the global language model and the topic dependent language model at a ratio of the mixture weight to generate a mixed language model of the speech recognition target;
A language model generation apparatus comprising:   The evaluation weight generation means calculates a perplexity of the global language model as the overall evaluation value and the individual evaluation value, and for the divided sentence in which the individual evaluation value is larger than the overall evaluation value. The language model generation device according to claim 1, wherein the evaluation weight is set large.   The evaluation weight generation means calculates a perplexity of the global language model as the overall evaluation value and the individual evaluation value, and for the divided sentence in which the individual evaluation value is larger than the overall evaluation value. 3. The language model generation device according to claim 1, wherein the evaluation weight is set to “0” for the divided sentences other than the evaluation weight “1”.   A language model generation program for causing a computer to function as the language model generation device according to any one of claims 1 to 4.

Images