JP2006343405A - Speech-understanding device, speech-understanding method, method for preparing word/semantic expression merge database, its program and storage medium - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a speech-understanding device, etc., which achieve speech understanding whose accuracy is higher. <P>SOLUTION: The speech-understanding device 100 is configured, such that a speech understanding result searching part 61 performs speech understanding by using a word/semantic expression merge N-gram model DB 20. Moreover, the speech-understanding device 100 is configured to be equipped with a word/semantic expression merge N-gram model preparing part 30, which prepares the word/semantic expression merge N-gram model DB 20 being the N-gram model of word/semantic expression merging from a language corpus 10 where clear correspondence relation among words and semantic expressions has not been carried out. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a speech understanding device, a speech understanding method, a method for creating a word / semantic expression set database, a program thereof, and a storage medium.

  The problem of speech understanding can be divided into two processes: speech recognition that recognizes speech as a word string, and language understanding that converts a word string into a set of semantic expressions.

  For language understanding, there is a technology that learns rules and probabilistic models for converting word strings to semantic expressions by means of a language corpus that clearly indicates which part of the word string corresponds to which semantic expression. Non-patent documents 1 to 3).

  Recently, language understanding devices have also been developed that perform language understanding from a language corpus that does not have an explicit correspondence between words and semantic expressions. Speech understanding can be performed by connecting this language understanding device and a conventional speech recognition device in series (see Non-Patent Documents 4 and 5).

  Furthermore, it can learn from a language corpus that has no explicit correspondence between words and semantic expressions, and it inputs speech recognition result word strings and the certainty of the words, and realizes speech understanding based on these information. Technology has also been developed (see Non-Patent Documents 6 and 7).

K. Hacioglu and Ward, "A word graph interface for a flexible concept based speech understanding network", in Proc. EUROSPEECH 2001, pp.1775-1778 H.Bonneau-Maynard and F.Lefevre, "Investigating stochastic speech understanding", in Proc.IEEE ASRU, 2001 Y. Esteve et al. "Conceputual decording for spoken dialogue systems", in Proc. EUROSPEECH 2003, pp.617-620 M. Epstein et al. "Statistical Natural Language Understanding using Hidden Clmpings", in Proc.ICASSP, vol.1, pp.176-179, 1996 K. Macherey et al., "Naural Langage Understanding using Statistical Machine Translation", in Proc. EUROSPEECH 2001, pp. 2205-2208 G. Tur et al. "Improving Spoken Language Understanding using Word Confusion Networks" in Proc.ICSLP, pp.1137-1140,2002 G. Tur et al. "Extending Boostig for Call classification using world Confusion Networks", in Proc.ICASSP, vol1, pp.437-440,2004

  In the techniques described in Non-Patent Documents 1 to 3, in order to create a language corpus that clearly shows the correspondence between word strings and semantic expressions, language-related expertise is required, and it takes time to create the language corpus. Therefore, there is a problem that the human cost becomes high.

  Further, when the technologies of Non-Patent Documents 4 and 5 are used in a spoken dialogue system, a language understanding device that is not related to speech is not designed in consideration of the ambiguity of an input word string. For this reason, the language understanding device uses the maximum likelihood speech recognition result (word string) as an object of understanding as it is, or rejects a word with low certainty by using information on the degree of certainty of recognition, Will be either. However, the former has a risk that a speech recognition error remains in the understanding result, and the latter has a problem that a necessary word may be rejected.

  Furthermore, since the technologies described in Non-Patent Documents 6 and 7 are based on the premise that each semantic expression occurs independently, the utterance including a certain semantic expression A is likely to include the semantic expression B. I can't figure out the relationship. That is, if the input voice data is ambiguous or the amount of information is small, the accuracy of voice understanding may be reduced.

  An object of the present invention is to provide a speech understanding device that solves the above-described problems and realizes speech understanding with higher accuracy.

  In order to solve the above-described problem, the speech understanding apparatus of the present invention creates a word / semantic expression database that creates a word / semantic expression set database indicating the probability of occurrence of a combination for each N-grams of combinations of words and semantic expressions. It was set as the structure provided with a set N-gram model preparation part. Further, by using this word / semantic expression set N-gram model, the word string and the semantic expression string that are input are recognized, and the certainty factor of the recognized word string and the semantic expression string is calculated. The voice understanding processing unit is provided. Other configurations will be described in the embodiments described later.

According to the present invention,
(1) Since it is not necessary to explicitly give the correspondence between words and semantic expressions to the language corpus to be used, the cost of creating a speech understanding device can be reduced.
(2) By tightly integrating the processing of speech recognition and language understanding, speech understanding that takes into account various hypotheses such as combinations of word strings and semantic expressions, arrangements of these combinations, and co-occurrence relationships can do. In addition, by calculating the certainty factor in consideration of various hypotheses as described above, it is possible to obtain a more accurate certainty factor for speech understanding.
(3) By using the simultaneous probability of the arrangement of words and semantic expressions as a statistical model, it is possible to express the co-occurrence relationship of semantic expressions. In other words, by using this semantic expression co-occurrence relationship, highly accurate speech understanding can be performed even when the input speech data is ambiguous or the amount of information is small.

  The best mode for carrying out the present invention (hereinafter referred to as an embodiment) will be described below.

Here, first, the hardware configuration of the speech understanding apparatus according to the present embodiment will be described.
The speech understanding device includes a language corpus 10 to be described later, an input interface that receives input of speech data 40, and an output interface that outputs the calculation processing result of the speech understanding unit 60 to the outside. The input interface is, for example, a network card, and the output interface is, for example, an output interface to a display device.

  The speech understanding device is realized by a computer including a memory such as a RAM (Random Access Memory), a storage unit such as a ROM (Read Only Memory) and a hard disk device, and an arithmetic processing unit such as a CPU (Central Processing Unit). This storage unit implements a word semantic group N-gram model creation program (word / semantic expression group database creation program) and a speech understanding unit 60 for realizing a word / semantic group N-gram model creation unit 30 described later. A voice understanding program or the like is stored. The CPU implements the functions of the word / semantic pair N-gram model creation unit 30 and the speech understanding unit 60 by developing and executing each program stored in the storage unit on the memory. In addition, the voice understanding device may be connected to an input device that inputs an instruction to the voice understanding device and a display device that outputs and displays a result of arithmetic processing of the voice understanding device. These configurations and devices will not be described in the drawings.

FIG. 1 is a block diagram illustrating an expanded function of the voice understanding device according to the present embodiment.
The function of the speech understanding device 100 will be described with reference to FIG.
The speech understanding device 100 receives an input of the language corpus 10 and the speech data 40, outputs a speech understanding result and a speech recognition result, and a word / semantic expression set N− based on the language corpus 10. A word / semantic pair N-gram model creating unit 30 for creating a gram model DB (database) 20 is provided. Further, the speech understanding device 100 refers to the word / semantic set N-gram model DB 20, the acoustic model DB (database) 50, and the conversion dictionary 80, and outputs the speech understanding result of the input speech data 40. The unit 60 is provided. The word / semantic expression set N-gram model DB 20 corresponds to the word / meaning expression set database in the claims. The acoustic model DB 50 corresponds to the acoustic model in the claims.

  The input / output unit 70 is realized by the input interface and the output interface described above. The language corpus 10, the acoustic model DB 50, the conversion dictionary 80, the speech data 40, and the word / semantic set N-gram model DB 20 will be described as being stored in a storage unit (not shown) of the speech understanding device 100. It may be stored in an external storage device and read via the input / output unit 70.

(Language corpus)
The language corpus 10 includes a transcription sentence (word string) that transcribes a human utterance and a set of semantic expressions corresponding to the contents of the sentence. The semantic expression here is such that the meaning of the utterance can be expressed by a set of semantic expression symbols (relationships and structures between symbols are undefined).
The sentence included in this language corpus 10 is, for example,
The word string “From Tokyo Station to Kyoto Station”,
Set of semantic expressions [from = (station = (Tokyo)), to = (station = (Kyoto))],
It is something like that.

  However, unlike the prior art, this language corpus 10 is not given a correspondence between words in a word string and semantic expressions. In other words, in the above example, “from Tokyo Station” corresponds to from = (station = (Tokyo)), and “to Kyoto Station” corresponds to to = (station = (Kyoto)). Absent.

(Word / semantic group N-gram model creation part)
The word / semantic set N-gram model creation unit 30 creates a word / semantic set N-gram model DB 20 based on the language corpus 10. The function of the word / semantic set N-gram model creating unit 30 is divided into a word / semantic expression relevance calculating unit 31, a word / semantic expression associating unit 33, and an N-gram model creating unit 35.

(Word / semantic expression relevance calculator)
The word / semantic expression relevance calculating unit 31 calculates the relevance between the word and the semantic expression for each word and semantic expression appearing in the language corpus 10 and stores the word / semantic expression relevance DB (database) 32. create. This word / semantic expression relevance DB 32 is a database showing how much each word in the language corpus 10 has a relevance to the semantic expression. The created word / semantic expression association degree DB 32 is once stored in a predetermined area of the storage unit and used for the arithmetic processing of the word / semantic expression association unit 33.

Here, the relevance calculated by the word / meaning expression relevance calculation unit 31 uses, for example, a criterion of φ ² described by the following formula (1). This φ ² is a standard for measuring the co-occurrence frequency of two kinds of symbols (word w and semantic expression c in the present embodiment), and measures the level of relevance between them. In the literature.

  W.A.Gale and K.W.Church, "Identifying word correspondences in parallel texts", in Proc. 4th DARPA Workshop on Speech and Natural Language, 1991

  Freq () in Equation (1) is the number of sentences in which the word w, the semantic expression c, or both appear in the language corpus 10, and N is the total number of utterances in the language corpus 10.

  By calculating the relevance using this formula (1), a word / semantic expression relevance DB 32 indicating how much each semantic expression has for a certain word is created.

(Word / semantic expression association unit)
The word / semantic expression associating unit 33 associates a word string and a semantic expression for each sentence in the language corpus 10 based on the word / meaning expression relevance DB 32.

Specific method, the word sequence _{(w 1, w 2, ...} , w l) and the columns of the semantic representations _{(c 1, c 2, ...} , c m) to, correspondence with a _i word w _i When the semantic expression to be given is c (a _i ), the optimal correspondence (a ₁ , a ₂ ,..., A _l ) is the relevance of each w _i , c (a _i ) pair in the sentence. It is defined that the product is the maximum (see the following formula (2)).

That is, in order to search for the optimal a ^ (a ₁ , a ₂ ,..., A _l ), the product of relevance is calculated for all combinations of w _i and a _i, and the value is maximized. Search for things and associate them. Then, the associated word w and semantic expression c are rewritten in the form of a symbol (t ₁ , t ₂ ,..., T _l ) combined in the form of “semantic expression: word”, and the correspondence between the word and the semantic expression The attached corpus 34 is created. Here, an empty semantic expression <eps> is associated with a word to which no semantic expression is associated. In addition, it is assumed that the degree of association between <eps> and each word is always 1.

  For example, the word / semantic expression associating unit 33 includes the language corpus 10 described above.

Word string "From Tokyo Station to Kyoto Station"
Set of semantic expressions [from = (station = (Tokyo)), to = (station = (Kyoto))]

As for the combination that calculates the degree of association for each

  from = (station = (Tokyo)): Tokyo <eps>: to to = (station = (Kyoto): Kyoto <eps>:

A corpus 34 with correspondence of the word and meaning expression is created. The corpus 34 with the word / semantic expression correspondence is stored in a predetermined area of the storage unit. This corpus 34 with correspondence between words and meaning expressions is used when the N-gram model creation unit 35 creates the word-meaning expression set N-gram model DB 20.

Here, the optimum association a ^ (a ₁ , a ₂ ,..., A _l ) is assumed to have the maximum product of the relevance of each pair of w _i and c (a _i ). However, the sum of the relevances may be maximized as in the following formula (3).

(N-gram model creation part)
This N-gram model creation unit 35 creates a word / semantic expression set N-gram model DB 20 based on each (t ₁ , t ₂ ,..., T _l ) of the corpus 34 with correspondence between words and semantic expressions.
This word / semantic expression set N-gram model DB 20 is obtained by modeling a combination of a word / semantic expression with an N-gram in the same manner as a word N-gram model generally used in speech recognition technology. That is, for each N-gram combination of words and semantic expressions, the probability of this combination occurring is shown.

Note that the word N-gram model is a model that approximates the conditional probability of the arrangement of (N-1) words before the word when determining the probability that a certain word will occur. By using this word N-gram model, the occurrence probability P (W) of the word string W = (w ₁ , w ₂ ,..., W _l ) is approximated as in the second row of the following formula (4). be able to.

In the prior art, the symbol in the word N-gram model is a word, whereas the word / semantic expression group N-gram model (word / semantic expression group N-gram model DB 20) of the present embodiment The difference is that the symbol is combined with a semantic expression corresponding to the word. That is, the word / semantic expression group N-gram model 20 determines the probability that a pair of a certain word and a semantic expression will occur. This model approximates a sequence with a conditional probability. Here, the joint probability P (W, C) between the word string W and the semantic expression string C, which are associated in the form of (t ₁ , t ₂ ,..., T _l ), is expressed by the following formula (5 ) In the second line.

  For example, first, the word / semantic set N-gram model creation unit 30 is a corpus 34 with correspondence between words / semantic expressions.

  from = (station = (Tokyo)): Tokyo <eps>: to to = (station = (Kyoto)): Kyoto <eps>:

  Is divided by N = 3 (3 grams) as follows.

<Beginning> from = (station = (Tokyo)): Tokyo <eps>: From
from = (station = (Tokyo)): Tokyo <eps>: from to = (station = (Kyoto)): Kyoto
<eps>: to to = (station = (Kyoto): Kyoto <eps>:
to = (station = (Kyoto): Kyoto <eps>: Until <end of sentence>

  Then, the occurrence probability of each combination in the entire language corpus 10 is calculated by the above-described mathematical formula (5), and the word / semantic expression set N-gram model DB 20 is created. Thereafter, the word / semantic set N-gram model creation unit 30 stores the created word / semantic expression set N-gram model DB 20 in a predetermined area of the storage unit.

  The acoustic model DB 50 shows correspondence between voice feature values and phonemes. Moreover, the conversion dictionary 80 shows the word by the combination of phoneme for every word. The acoustic model DB 50 and the conversion dictionary 80 are referred to when a voice understanding result search unit 61 described later performs voice recognition or semantic expression recognition of the voice data 40. The acoustic model DB 50 and the conversion dictionary 80 are the same as those used in the conventional speech recognition device.

(Speech Understanding Department)
Next, the voice understanding unit 60 will be described. When the speech understanding unit 60 receives the input of the speech data 40 via the input / output unit 70, the speech understanding unit 60 refers to the acoustic model DB 50, the conversion dictionary 80, and the word / semantic expression set N-gram model DB 20 of the storage unit, and obtains the speech understanding result. (Semantic expression recognition result and speech recognition result) are output. The voice understanding unit 60 is divided into a voice understanding result search unit 61 and an output shaping unit 63.

(Speech understanding result search part)
When the speech understanding result searching unit 61 receives the input of the speech data 40, the speech understanding result searching unit 61 refers to the acoustic model DB 50 and the word / semantic expression set N-gram model DB 20 to obtain a word and a semantic expression corresponding to the input speech data. Output a sequence of symbols in pairs. The speech understanding result search unit 61 also outputs a certainty factor in recognition of each symbol. The speech understanding result search unit 61 corresponds to the speech understanding processing unit in the claims.

  The certainty factor is a value indicating the degree to which no other candidate that could compete with the word (semantic expression) was found. That is, as the certainty factor approaches 1, it indicates that there is no other candidate that competes with the word (semantic expression), and as the certainty factor approaches zero, the word (semantic expression) in the speech understanding process. This indicates that many other words (semantic expressions) candidates having similar scores were competing. The speech understanding result search unit 61 can be realized by applying speech recognition processing in the conventional speech recognition technology.

The mathematical problem setting in the speech understanding result search unit 61 is as follows.
X represents a time-series feature amount obtained from speech data, W represents a word string, and C represents a semantic expression string. P (X | W) is called an acoustic model, and P (W) is called a language model. In the speech recognition apparatus as the prior art, a word string W ^ that maximizes the conditional probability P (W | X) of the following formula (6) (or maximizes the likelihood of the feature amount) is searched. And output as a speech recognition result.

  Here, the calculation formula that the speech understanding result search unit 61 of the present embodiment outputs a sequence of semantic expressions directly from the voice data 40 is expressed by the following formula (7) by applying the formula (6). .

In addition, the approximation of the 3rd line of Formula (7) is
(1) From the assumption that X is independent of C, P (X | W, C) is replaced with P (X | W);
(2) Probability P (X | W) P (W for the highest likelihood X, W of the summation type of Σ _W P (X | W) P (W | C) P (C) by Viterbi approximation | C) Replacement with P (C),
It is carried out.

  That is, in the present embodiment, Formula (4) and Formula (6) and Formula (5) and Formula (7) are used in the same form by using the same form. The word N-gram model is replaced with a word / semantic expression group N-gram model (word / semantic expression group N-gram model DB 20). Thereby, from the speech recognition apparatus in which the model is replaced, a symbol in which a word and a semantic expression are paired with respect to the input of the speech data 40 is obtained as a recognition result.

  The certainty factor of the recognized symbol is expressed as a posterior probability that the symbol appears in the recognition result. Note that a method for calculating the posterior probability in speech recognition is described in the following document.

  Frank Wessel et al., "Confidence measures for large vocabulary continuous speech recognition", IEEE Tansactionsons Speech and Audio Processing, Vol. 9, no. 3, pp. 288-298

  Moreover, the technique which incorporated calculation of posterior probability in the speech recognition apparatus is described in the following documents.

Lee Yong-sin, et al., “High-speed reliability calculation method based on word posterior probabilities in the 2-pass deep search algorithm”, Information Processing Society of Japan vol.2003, No. 124, 2003-SLP-9, pp.281-286

  When the speech understanding result searching unit 61 according to the present embodiment calculates the certainty factor using these techniques, the certainty factor is added to the symbol string of the word / semantic expression group and the word / semantic expression pair symbol with the certainty degree is attached. Column 62 is created. Then, this certain word / meaning expression combination symbol string 62 with certainty is stored in a predetermined area of the storage unit. Thereafter, the word / semantic expression combination symbol string 62 with certainty is shaped into a predetermined format by the output shaping unit 63.

(Output shaping section)
The output shaping unit 63 selects a symbol having a certainty level higher than a certain level (threshold) from the word / semantic expression pair symbol string 62 with certainty level. Then, a word and a semantic expression are extracted from the symbol, and are formatted and output as a set of a word string and a semantic expression corresponding to the input voice data 40.

  The symbol string of the output at this time uses a symbol in which each symbol is in the form of “semantic expression: word” and a pair of a word and a semantic expression. The output shaping unit 63 can obtain a semantic expression sequence that is a result of the semantic expression recognition by extracting the semantic expression symbol from the “semantic expression: word”, and can extract the symbol of the word from the “semantic expression: word”. Thus, a word string that is a voice recognition result can be obtained.

  Here, operation | movement of each component of the speech understanding apparatus 100 is demonstrated easily using FIG.

First, the word / semantic expression relevance calculating unit 31 calculates the relevance between a word and a semantic expression for the language corpus 10 input via the input / output unit 70 and creates a word / semantic expression relevance DB 32 ( S1). Next, the word / semantic expression associating unit 33 refers to the word / semantic expression association degree DB 32 and calculates a combination that maximizes the degree of association between the word of the language corpus 10 and the semantic expression. Then, the word / semantic expression association unit 33 creates a corpus 34 with word / semantic expression correspondence by the combination (S2). Next, the N-gram model creation unit 35 calculates the occurrence probability of the word / semantic expression group N-gram based on the corpus 34 with correspondence between the words / semantic expressions, and the word / semantic expression group N-gram model DB 20. Is created (S3). When the speech understanding result search unit 61 receives input of the speech data 40 via the input / output unit 70, the speech understanding result search unit 61 refers to the acoustic model DB 50 of the storage unit, the conversion dictionary 80, and the word / semantic expression set N-gram model DB 20 created in S3. Then, a word / semantic expression pair symbol string 62 with certainty corresponding to the input voice data 40 is created (S4). Then, the output shaping unit 63 selects a symbol having a certainty level higher than a certain level (threshold) from the word / semantic expression set symbol string 62 with certainty level. Next, the output shaping unit 63 extracts words and semantic expressions from the selected symbols, shapes them as a set of word strings and semantic expressions corresponding to the input voice data 40, and passes through the input / output unit 70. Output (S5).
In this way, the voice understanding device 100 calculates and outputs the voice understanding results (semantic expression recognition results and voice recognition results).

  The voice understanding device 100 according to the present embodiment can be realized by a voice understanding program and a word / semantic expression set N-gram model creation program for executing the processing as described above, and these programs can be read by a computer. It can be stored in a simple storage medium (CD-OM or the like) and provided. It is also possible to provide the program through a network.

  Next, examples of the present invention will be described. In the present embodiment, utterances collected by a spoken dialogue system for a railway route guidance domain (field) are used as the language corpus 10 input to the speech understanding device 100.

  In the speech dialogue system described above, yes (affirmation), no (denial), backchannel (departure), departtime (departure time: upper class of time), arrivaltime (arrival time: upper class of time), time (time) (No departure / arrival designation)), from (departure station: upper class of station), to (arrival station: upper class of station), station (station name (no departure / arrival designation)), no = (X) Semantic expressions such as (Negation of X) are defined.

  Here, the semantic expression of the upper class includes the lower semantic expression. For example, when simply referring to “Tokyo Station”, it is represented as station = (Tokyo), and when referring to “Tokyo Station as a departure station”, it is represented from from = (station = (Tokyo)).

  The language corpus 10 used in the present embodiment is composed of a voice file name: a written sentence (word string): a corresponding semantic expression string (partially extracted) as follows.

(Language corpus)
/norikae/20030807/000/20030807-000-000.wav:From Takebashi to Takasaki ： from
= (station = (Takehashi)) to = (station = (Takasaki))
/norikae/20030807/000/20030807-000-001.wav:18:20 Arrival:
arrivetime (hour = (18), minute = (20))
/norikae/20030807/000/20030807-000-002.wav: Please, yes
/norikae/20030807/000/20030807-000-003.wav:From Takebashi to Takasaki ： from = (station = (Takebashi)) to = (station = (Takasaki))
/norikae/20030807/000/20030807-000-006.wav：18:20, arrive at Takasaki ： to = (station = (Takasaki)) arrivetime (hour = (18), minute = (20))
/norikae/20030807/000/20030807-000-007.wav: 18:20 Arrival at Takasaki: to = (station = (Takasaki)) arrivetime (hour = (18), minute = (20))

The word / semantic group N-gram model creation unit 30 calculates the relevance (φ ² ) of the above-described mathematical formula (1) for each pair of the word and the semantic expression in the language corpus 10, The following file to be the semantic expression set N-gram model DB 20 is created (partially extracted). Here, each line of the file indicates a word in the language corpus 10 and the degree of association between each semantic expression for the word. The format here is “semantic expression: relevance (φ ² )”.

(File)
Daiba
to = (station = (daiba)): (0.111038)
station = (Daiba): (0.110904)
arrivetime (hour = (9), minute = (40)): (0.002822)
from = (station = (Akabane)): (0.002662)
from = (station = (prince)): (0.000966)
backchannel: (0.000039)
Until
to = (station = (Yokohama)) :( 0.001614)
no: (0.01138)
to = (station = (Soga)): (0.00001059)
to = (station = (daiba)): (0.0002)

The second to seventh lines of this file indicate the degree of association (φ ² ) of each semantic expression with respect to the word “Daiba”. For example, the word “daiba” and the semantic expression “to = (station = (daiba))” (“daiba” as the destination) indicate that they have a relevance of “0.111038”. (See line 2).

The ninth to twelfth lines show the degree of association (φ ² ) of each semantic expression for the word “to”. For example, the word “to” and the semantic expression “to = (station = (daiba))” indicate that they have a degree of association of “0.000002” (see the 12th line).

  Subsequently, the word / semantic expression association unit 33 searches for an optimum association between the word string and the semantic expression string of the language corpus 10 using the file data described above. The search here is solved as a search problem of association that maximizes a ^ in the above formula (2). That is, the product of the relevance levels for all combinations of associations is obtained in each sentence, and the product having the maximum value is selected as the optimum association.

For example, if the language corpus 10 is
Word string: to Daiba
Semantic expression sequence: to = (station = (daiba)),
The possible correspondences are as follows:
(1) to = (station = (Daiba)): up to Daiba <eps>:
(2) <eps>: Daiba to = (station = (Daiba)): Until
There are two types.
Here, the product of the relevance level of (1) is “0.11038”, and the product of the relevance level of (2) is “0.000002”, so (1) is selected as the optimum association.

As a result of such processing, the corpus 34 with the correspondence between the word and the semantic expression is created by rewriting the language corpus 10 as follows by using a new symbol in which the optimally matched word and the semantic expression are paired. Yes (partial excerpt). Here, a symbol that is a combination of a word and a semantic expression is shown separated by a space.
<S>: <s> is a symbol indicating the start of a sentence, and </ s>: </ s> is a symbol indicating the end of the sentence.

(Corpus with word / semantic expressions)
<s>: <s> from = (station = (Yokohama)): Yokohama <eps>: <eps>: Nine <eps>: Time \\
departtime = (hour = (9), minute = (45)): forty-five <eps>: minute <eps>: to <eps>: departure \\
to = (station = (Soga)): Soga <eps>: Until </ s>: </ s>
<s>: <s> departtime = (hour = (21), minute = (50)): 21 <eps>: hour <eps>: fifty \\
<eps>: minutes <eps>: to <eps>: ride <eps>: then <eps>: mass </ s>: </ s>
<s>: <s> departtime = (hour = (14), minute = (30)): Fourteen <eps>: Time <eps>: Thirty \\
<eps>: Minutes <eps>: from = (station = (Shinjuku)) <eps>: <eps>: Departure <eps>: Shi \\ <eps>: Mas </ s>: </ s>

The word / semantic expression N-gram model creation unit 30 converts the corpus 34 with the word / semantic expression correspondence thus created into a word / semantic expression group N-gram model DB 20.
In this embodiment, the CMU-cambridge SLM Toolkit (P. Clarkson and R. Rosenfeld, “Staistica1 1anguage modeling using the CMU-”, which is a technology released as freeware as the word / semantic pair N-gram model creation unit 30. N-gram language model creation software, “Cambridgetoolkit”, in Proc. EUROSPEECH 1997, pp. 2707-2710) was used.

  When this corpus 34 with word / semantic expression correspondence is processed in the same manner as in the case of a language corpus with only normal word strings, a file in the following format (ARPA format) can be output. (Excerpt). In this example, a trigram model with N = 3 and a bigram model with N = 2 were created. A part of the N = 3 trigram model created in this example is shown below. The value on the left is the likelihood score (logarithmic notation) for the three sets of symbols on the right.

  For example, in the following word / semantic expression group N-gram model DB 20, “to = (station = (Asakusa)): Asakusa <eps>: to from = (station = (Shinjuku)): Shinjuku” for the language corpus 10. The likelihood score of the three sets indicates “−0.5661”.

(Word / semantic expression group N-gram model DB)
-0.5146 to = (station = (Kawagoe)): Kawagoe backchannel: Ya <s>
-1.1167 to = (station = (Asakusa)): Asakusa <eps>: to <s>
-0.5661 to = (station = (Asakusa)): Asakusa <eps>: From = (station = (Shinjuku)): Shinjuku
-1.1167 to = (station = (Asakusa)): Asakusa <eps>: from = (station = (Okubo)): Okubo
-0.2651 to = (station = (Asakusa)): Asakusa <eps>: te <eps>:
-0.5146 to = (station = (Asakusa)): Asakusa <eps>: Departure <eps>: Minutes

  The word / semantic expression set N-gram model DB 20 created in this way is used in the speech recognition apparatus. In this embodiment, as the speech understanding result search unit 61, Julius (A. Lee et al., “Julius-an open source real-time large vocabllary recognition engine”, in Proc. Speech recognition software called EUROROSPEECH 2001, p.1691-1694) is used.

  The symbol of the word / semantic expression group is defined as a recognized vocabulary having a word part reading (phoneme string) in the conversion dictionary 80 as a vocabulary file having the following format.

from = (station = (Takebashi)) ： Takebashi from = (station = (Takebashi)) takeba sh i
from = (station = (Takasaki)) ： Takasaki from = (station = (Takasaki)) takasaki

  The first column of each line separated by a space is a symbol of a word / semantic expression pair, the second column is a notation symbol that the speech recognition software outputs as a standard for the symbol, and the third column is a recognition vocabulary. It is a phoneme sequence.

  Here, when a voice file recorded in the WAV format as voice data 40 is input to the voice understanding device 100 having the recognition vocabulary (conversion dictionary 80) (in this embodiment, a voice recognition device which is a conventional technology is used). The speech understanding result search unit 61 outputs the following recognition result with reference to the recognition vocabulary and the word / semantic expression set N-gram model DB 20. This recognition result corresponds to the word / semantic expression pair symbol string 62 with certainty in the present embodiment.

(Confirmed word / semantic expression set symbol string)
sentencel: from = (station = (Takehashi)) to = (station = (Takasaki))
wseq1: <s> from = (station = (Takebashi)): Takebashi <eps>: to to = (station = (Takasaki)): Takasaki <eps>: To </ s>
phseq1: silB ｜ takeba sh i | kara | takasa ki | made | silE
cmscore1: 0.984 0.982 0.955 0.898 0.510 1.000
score1: -5849.820801

  Among these, the first line (sentencel) is the speech understanding result obtained in this embodiment, the second line to the third line (wseq1) are the symbol strings of the recognized word / semantic expression group, and the fourth line (phseq1). Is the phoneme string, the fifth line (cmscore1) is the certainty, and the sixth line (score1) is the recognition score.

Next, the output shaping unit 63 uses the word / semantic expression pair symbol string 34 of the word / semantic expression pair symbol string 34 with the certainty factor and the certainty factor (cmscore1).
(1) Remove the symbols <s> and </ s> representing the beginning and end of the sentence.
(2) Symbols below the certainty level (threshold value) set in advance are not output (in this embodiment, the certainty threshold is set to 0.5).
(3) Each symbol is divided by a symbol “:” that separates the semantic expression and the word, and the symbol string of the semantic expression and the word string are output separately.

This
As a result of semantic expression recognition,

  from = (station = (Takehashi)) (confidence 0.982) to = (station = (Takehashi)) (confidence 0.898)

Is output,
As a result of speech recognition,

  From Takebashi (confidence 0.982) to (confidence 0.955) Takasaki (confidence 0.898) (confidence 0.510)

Is output.

As described above, the voice understanding device 100 can perform voice understanding of the input voice data 40. Here, the certainty threshold set in the output shaping unit 63 is stored in the storage unit and can be changed via the input / output unit 70. For example, if the certainty threshold is set to 0.6 for the speech understanding result, a symbol with confidence 0.6 or less is not output, so the speech recognition result “until (confidence 0.510)” is not output. .
That is, when the user of the voice understanding device 100 wants to obtain a voice understanding result with a higher certainty level, the threshold is set higher, and when the user wants to obtain a voice understanding result including a lower confidence level, the threshold is set lower. do it.

"Experimental result"
As an experiment for showing the effect obtained by the present invention, a word N-gram model (word N-gram model DB) and a word / semantic expression set N- created using the same language corpus (consisting of about 9000 sentences). FIG. 2 shows the result of speech understanding using the following methods for about 3000 sentences of speech data 40 using the gram model DB 20 (see FIG. 2). In this experiment, speech understanding was performed by the following four methods (1) to (4), and the performance of each method was compared. Here, the rejection level of the speech understanding result is changed from 0 (not rejected at all) to 1 (all rejected) according to the certainty level, and the performance is compared.
(1) Speech recognition is performed using a word N-gram, and a semantic expression string having the highest likelihood is searched for and obtained from the obtained word string using the word-semantic expression group N-gram model DB 20 and output. How to do (baseline).
(2) Speech recognition is performed using a word N-gram, and after the word with high certainty is rejected and replaced with a symbol representing an unknown word in the obtained word string, the maximum likelihood for the word string is obtained. A method of searching and outputting a high-level semantic expression sequence using the word / semantic expression group N-gram model DB 20 (a method of ignoring an uncertain word using a certainty factor of a word) (WordReject).
(3) The voice understanding method (Proposed) of this embodiment.
(4) (For comparison) Assuming that all speech recognition is performed correctly, the most likely semantic expression sequence for the correct word sequence is determined using the word / semantic expression set N-gram model DB 20. Search and output method (Transcription).
The horizontal axis (Precision) in FIG. 2 indicates the ratio (accuracy rate) of correct answers among the semantic expressions output as the speech understanding results. The vertical axis (Recall) indicates the ratio (recall rate) of correct answers to be obtained for the speech data 40 that are output as speech understanding results. In both cases, the unit is%.

  As shown in FIG. 2, for example, when the precision ratios of the output results obtained by the methods (1) to (3) are compared around the recall ratio of 80%, all are about 80%. However, when compared around 70% recall, (3) Proposed method (this example) can perform speech understanding with higher precision than (1) baseline and (2) WordReject methods. I understand.

  Further, for example, when the reproduction rate of the output result by the methods (1) to (3) is compared around the matching rate of 86%, the (1) Precision method is about 37%, and (2) the WordReject method is It is about 37% to 70%, and it can be seen that (3) Proposed method (Example) is about 78%. That is, it can be seen that (3) Proposed method (example) can perform speech understanding with a higher recall than (1) baseline and (2) WordReject methods.

  Based on this result, the present invention integrates the processes of speech recognition and language understanding, calculates confidence, and rejects speech understanding results that are below this confidence level. It has been shown that the precision can be increased when the recall is obtained. In addition, it was shown that the reproducibility can be increased when the same precision as other methods can be obtained.

It is the block diagram which expanded and demonstrated the voice | voice understanding apparatus of this Embodiment. It is the graph which showed the performance change when changing the rejection level of a speech comprehension result with certainty.

Explanation of symbols

10 Language Corpus 20 Semantic / Expression Group N-Gram Model DB (Word / Semantic Expression Group Database)
30 Word / Semantic Pair N-Gram Model Creation Unit 31 Word / Meaning Expression Relevance Calculation Unit 32 Word / Meaning Expression Relevance DB
33 Word / Semantic Expression Corresponding Unit 34 Corpus with Word / Semantic Expression Correspondence 35 N-gram Model Creation Unit 40 Audio Data 50 Acoustic Model DB (Acoustic Model)
60 Speech understanding unit 61 Speech understanding result search unit (speech understanding processing unit)
62 Word / semantic expression set symbol string with certainty factor 63 Output shaping unit 70 Input / output unit 80 Conversion dictionary 100 Speech understanding device

Claims (8)

An input / output unit for inputting / outputting various data;
A storage unit for storing a word / semantic expression group database indicating a probability of occurrence of the combination and an acoustic model indicating acoustic features of phonemes constituting the word for each N-grams of combinations of words and semantic expressions;
The speech data input via the input / output unit is recognized with reference to the acoustic model and the word / semantic expression set database to recognize a word string and a semantic expression string expressed by the speech data, and the recognized word A speech understanding processor that calculates the certainty of the sequence and the semantic representation sequence;
A speech understanding device comprising: A speech understanding method using a speech understanding device for recognizing a word string and a semantic expression string that the voice data means based on input voice data,
The voice understanding device
Receiving the voice data input;
Refer to the word / semantic expression group database indicating the probability of occurrence of the combination and the acoustic model indicating the acoustic features of the phonemes constituting the word for each N-gram of the combination of the word and the semantic expression. A word string and a semantic expression string that are data recognition results, and a certainty factor that is a degree that no other candidate that competes with the word string and the semantic expression string is found for each of the word string and the semantic expression string. Speech understanding step to output,
A speech understanding method characterized by executing   In the speech understanding step, the word string and the semantic expression string, which are the recognition results, are obtained from the word string and the semantic expression string that maximize the likelihood of the time-series feature amount of the input speech data, or an approximate expression thereof. The speech understanding method according to claim 2, wherein the speech understanding method is a word string and a semantic expression string. The voice understanding device
The speech understanding method according to claim 2 or 3, wherein a word string and a semantic expression string with the certainty factor exceeding a predetermined threshold value are output. A method for creating a word / semantic expression set database used in the speech understanding method according to claim 2,
The voice understanding device
For each sentence, accepting an input of a language corpus indicating a set of word strings and semantic expressions constituting the sentence;
Calculating the degree of association between each word and the semantic expression based on the co-occurrence frequency of the word and the semantic expression included in the language corpus, and creating a word / semantic expression relation degree DB;
Referring to the created word / semantic expression relevance DB, for each possible combination of words and semantic expressions in each sentence of the language corpus, a combination that maximizes the sum or product of the relevance is calculated. Steps,
Creating a corpus with word / semantic expression correspondence that associates words and semantic expressions of the language corpus with the calculated combinations;
In the corpus with correspondence of the word / semantic expression, calculating a joint probability of the combination of the word and the meaning, and creating a word / semantic expression pair database including the joint probability;
A method of creating a word / semantic expression set database characterized by   The simultaneous probability of the combination of the word and the semantic expression indicates the probability that the combination of the word and the semantic expression occurs, with the condition of the arrangement of the combination of the word and the semantic expression before (N−1) pieces of the combination. 6. The method of creating a word / semantic expression set database according to claim 5, wherein the word / semantic expression set database is approximated by probability.   A program for causing a computer to execute the method according to any one of claims 2 to 6.   A computer-readable storage medium storing the program according to claim 7.

Images