JP3001334B2 - Language processor for recognition - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a recognition language processing apparatus used for voice recognition, character recognition, and the like.

[0002]

2. Description of the Related Art Conventionally, when recognizing a sentence or a word by an input voice or character string by a recognition device, for example, the following is performed. Standard patterns such as syllables by voice signals and image signals generated in advance and stored in memory, and characteristic patterns of syllables and characters extracted from voice and image signals input from microphones and image reading devices. To obtain a plurality of recognition candidates. Then, the obtained plurality of recognition candidates are developed in the order of similarity to create a phrase candidate sequence or a word candidate sequence, and language processing is performed on all the created phrase candidates or word candidates.

[0003]

However, the above-described conventional recognition method has no problem when the number of recognition candidates is small. However, when the number of recognition candidates increases, the number of phrase candidates and word candidates obtained by expansion increases explosively.

For example, in the case of a speech recognition device that recognizes each syllable, if six syllables are recognized and eight recognition candidates are obtained for each syllable, a word / syllable candidate obtained by a combination thereof is obtained. The number will be 86 ≒ 260,000. When linguistic processing is performed on each of the word / phrase candidates, it takes 260 seconds even if the linguistic processing for one word / phrase candidate is completed in 1 ms, for example. Can not.

An object of the present invention is to provide a recognition language processing method capable of performing high-speed language processing on recognition candidates.

[0006]

According to a first aspect of the present invention, there is provided a recognition language processing apparatus for recognizing speech or characters based on an input speech signal or image signal. In a recognition language processing apparatus that obtains a recognition candidate by performing language processing on a result, a notation of a recognition result from the recognition apparatus is converted into a notation of a smaller number of categories than the number of categories to which the notation of the recognition result belongs according to a predetermined rule. A notation conversion unit, a compression dictionary in which the notation of the headword of the language processing dictionary used when performing the above-described language processing is converted according to the above rules, and the number of categories to which the notation of the headword belongs is compressed; A recognition candidate for performing a preliminary selection of the recognition candidate by performing language processing on the recognition result whose notation is converted by the conversion unit using the compression dictionary A preparatory selection unit, a recognition candidate generating unit that restores the preselection recognition candidates obtained by the recognition candidate preselection unit into a chain of headwords of the language processing dictionary to obtain the recognition candidates, and a recognition from the recognition device. A top-down processing unit that calculates the weight of the recognition candidate obtained by the recognition candidate generation unit in a top-down manner with reference to the result and obtains the final recognition candidate is provided.

A recognition language processing apparatus according to a second aspect of the present invention is the recognition language processing apparatus according to the first aspect, wherein the recognition result from the recognition apparatus includes a weighted recognition unit group, and When calculating the weight of the recognition candidate by the top-down processing unit, if there is a recognition unit not included in the recognition result among a plurality of recognition units constituting the recognition candidate, a predetermined weight is added to the recognition unit. A recognition result adding unit that automatically adds the recognition result to the recognition result is provided.

The recognition language processing apparatus according to a third aspect of the present invention is the recognition language processing apparatus according to the first or second aspect, wherein the notation conversion unit converts the notation of the recognition result from the recognition apparatus. The above-mentioned predetermined rule is converted to a notation that belongs to one category when a notation of a recognition result applicable to the pattern is converted when an erroneous recognition pattern exists in the recognition by the recognition device. The rule is that it is a rule.

[0009]

According to the first aspect of the invention, when a voice signal or an image signal is input to the recognition device, a voice or a character is recognized based on the voice signal or the image signal, and a recognition result is output. Then, the notation of the recognition result output from the recognition device is converted by the notation conversion unit into a notation of a smaller number of categories than the number of categories to which the notation of the recognition result belongs according to a predetermined rule.

Then, with respect to the recognition result after the notation conversion by the notation conversion unit, the recognition candidate preliminary selection unit sets the headword of the language processing dictionary used when performing the language processing to the predetermined word. Language processing is performed using a compression dictionary converted according to the rules. Thus, the preliminary selection of the recognition candidate is performed, and the preliminary selection recognition candidate is obtained.

[0011] Thereafter, the recognition candidate generating section restores the preselected recognition candidates to a chain of headwords in the language processing dictionary to generate recognition candidates. Then, the top-down processing unit refers to the recognition result, calculates the weight of the generated recognition candidate in a top-down manner, and obtains a final recognition candidate.

As described above, the language processing dictionary and the recognition result are converted in advance into a notation of a smaller number of categories, so that the language processing using the compression dictionary for the recognition result after the notation conversion can be performed in a very short time. Will be implemented. Thereafter, since the weights of the recognition candidates generated based on the thus-obtained preliminary selection recognition candidates are simply calculated from the top down, the bottom-up language processing is not required. The final recognition candidate is obtained in a very short time.

Further, in the second invention, a recognition result composed of a weighted recognition unit group is output from the recognition device. When the weight of the recognition candidate is calculated by the top-down processing unit, if there is a recognition unit not included in the recognition result among the plurality of recognition units constituting the recognition candidate, the recognition result adding unit performs the recognition. A predetermined weight is added to the unit and automatically added to the recognition result. By doing so, there is no case where the recognition candidate is rejected, and the weights of all the recognition candidates generated by the recognition candidate generator are calculated from the top down.

[0014] In the third invention, the notation based on the recognition result applicable to the erroneous recognition pattern existing in the recognition by the recognition device is converted into a notation that belongs to one category. The notation of the recognition result from the recognition device is converted by the conversion unit. Thus,
The above-mentioned recognition result is efficiently transcribed according to the tendency of recognition errors by the recognition device.

[0015]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, the present invention will be described in detail with reference to the illustrated embodiments. In this embodiment, syllables are used as recognition units,
A description will be given based on a case where language processing is performed in units of phrases.

<First Embodiment> FIG. 1 is a block diagram of a recognition language processing apparatus according to this embodiment. The input / output unit 2 sends the input speech signal to the speech recognition device 10, receives a syllable candidate that is a recognition result in syllable units from the speech recognition device 10, and outputs the syllable candidate to the lattice creation unit 3. The lattice creation unit 3 creates a syllable lattice from the input syllable candidates.

The reduced notation conversion unit 4 receives the syllable lattice from the lattice creation unit 3 and converts the notation of each syllable candidate constituting the syllable lattice into reduced notation according to a predetermined rule to obtain a reduced notation syllable lattice. The above rule is a notation representing a vowel as described in detail later on the notation of syllable candidates constituting the syllable lattice (hereinafter this notation is called the original notation).
(Hereinafter, this notation is referred to as reduced notation.) Note that the rules in this case may be tabulated in a table representing the relationship between the original notation and the reduced notation and stored in a memory.

The reduced notation language processing unit 5 expands the reduced notation syllable lattice obtained by the reduced notation conversion unit 4 and performs reduced notation language processing using the reduced notation dictionary 6. Then, the candidate deleting unit 7 refers to the result of the reduced notation language processing from the reduced notation language processing unit 5, and selects an impossible syllable candidate from the syllable candidates constituting the syllable lattice created by the lattice creating unit 3. Remove. Thereafter, the original notation language processing unit 8 develops the syllable lattice in which the impossible syllable candidates are deleted and the number of constituent elements is reduced to create a phrase candidate sequence, and uses the original notation dictionary 9 to perform the original notation language processing. Is implemented.

The phrase candidate sequence successfully processed in the original notation language is output to an external device (a device for obtaining a recognition result, not shown) via the input / output unit 2.

Next, the contents of the original notation dictionary 9 and the reduced notation dictionary 6, the contents of the syllable lattice created by the lattice creation unit 3, the reduced notation conversion unit 4, the reduced notation language processing unit 5, and the candidate deletion unit 7 and each processing performed by the original notation language processing unit 8 will be described in detail.

FIG. 2 shows the contents of the original notation dictionary 9 as the language processing dictionary handled in this embodiment. In the original notation dictionary 9, three prefixes, five independent words, and three attached words are registered. Then, it is assumed that all independent words have a possibility of being prefixed with a prefix, while others have a possibility of having an attached word.

FIG. 3 shows a syllable lattice obtained by the lattice creation unit 3 as a result of recognizing the utterance of the five-syllable "national". However, the numbers in parentheses indicate the matching distance of each syllable candidate (hereinafter simply referred to as distance). In this example, there are 3 to 5 syllable candidates for each syllable, and when these syllable candidates are expanded for language processing, a total of 720 syllable candidate sequences are obtained.

FIG. 4 shows a case where the original spelling language processing is performed using the original spelling dictionary 9 shown in FIG. 2 with respect to the above 720 candidate phrase strings obtained by expanding the syllable lattice shown in FIG. FIG. The original notation language processing here refers to general kana-kanji conversion processing, and a recognition result sequence in which kana-kanji conversion has failed is rejected.
In FIG. 4, recognition result strings that have been successfully converted into kana-kanji characters are arranged in ascending order of distance. This distance is the total value of the distances (see FIG. 3) of each syllable candidate constituting the recognition result sequence.

The present invention relates to how to obtain the recognition result sequence shown in FIG. 4 from the syllable lattice (recognition result) shown in FIG. 3 using the original notation dictionary 9 shown in FIG.

FIG. 5 shows the contents of the reduced notation dictionary 6 as the compression dictionary, which is a feature of the present invention. Here, the reduced notation refers to a syllable candidate obtained as a result of recognition converted to a lesser notation. That is, in the present embodiment, a total of 111 syllables “a / ka / sa / ta / na / ...” are grouped for each syllable having the same vowel, and reduced to syllables belonging to the vowel / a / group. Give the notation / A /. Similarly, the vowel / i / group is given a reduced notation / I /
The / u / group is given reduced notation / U /, the vowel / e / group is given reduced notation / E /, and the vowel / o / group is given reduced notation / O /. In addition, sound-repellent / n / reduced notation
Give / X /. Thus, the 111 syllables are / A /, / I
Expressed in six types of reduced notation of /, / U /, / E /, / O / and / X /, the number of categories “111” in the original notation is compressed to “6”.

In the reduced notation dictionary 6, each headword in the original notation dictionary 9 shown in FIG. 2 is represented by a smaller reduced notation (in the case of the present embodiment, the above-described six reduced notations representing vowel strings and pronunciation). It was done. For example, in the independent word dictionary of the original notation dictionary 9 in FIG. 2, five types of headwords “Kokumin”,
"Komin", "Tomin", "Funin" and "Fumin" are registered, but when they are replaced with reduced notation, "Kokumin",
"Komin" and "Tomin" have the same notation "OUIN", and "Fuin" and "Fumin" have the same notation "UX".
As a result, the number of entry words of the independent word dictionary in the reduced notation dictionary 6 is reduced to two types.

Actually, since each item of the original notation dictionary 9 has various flags indicating a connection relation, a utilization type, and the like, it is not possible to integrate headings simply because the reduced notation is the same. Can not. Even in such a case, similar headwords can be integrated by performing operations such as newly setting a new flag or providing a plurality of flag areas for one headword.

FIG. 6 shows a reduced syllable lattice obtained by performing a reduced notation conversion by the reduced notation conversion unit 4 on the syllable lattice shown in FIG. In parentheses,
Describe the minimum distance among the distances of the original syllable in the same reduced notation.

FIG. 7 shows a result obtained by performing reduced notation language processing on the reduced notation syllable lattice shown in FIG. 6 by the reduced notation language processing unit 5 using the reduced notation dictionary 6 shown in FIG. . In FIG. 7, reduced notation phrase candidate strings obtained from all combinations of the reduced notation syllable lattices shown in FIG. 6 are listed, and the reduced notation phrase candidate strings that have been successfully subjected to the reduced notation language processing are marked with a circle. The reduced notation phrase candidate sequence that is rejected due to failure in the notation language processing is marked with x.

Here, in the case of this embodiment, as shown in FIG. 6, the number of syllable candidates constituting the reduced notation syllable lattice is "1" for the first syllable, "2" for the second syllable, and "2" for the second syllable. Since the third syllable is “1”, the fourth syllable is “2”, and the fifth syllable is “2”, the number of reduced notation syllable candidate columns obtained from all combinations of reduced notation syllable lattices is “8 (= 1 × 2 × 1 × 2 ×
2) ", which is much smaller than the number of phrase candidate strings" 720 (= 4 × 4 × 3 × 3 × 5) ”obtained by expanding the syllable lattice shown in FIG. In combination with the fact that the number of headwords registered in the reduced notation dictionary 6 is reduced, the number of times of performing the reduced notation language processing on the reduced notation phrase candidate sequence can be greatly reduced. In addition, the number of phrase candidate columns that succeeded in the reduced notation language processing is “2”, which is smaller than the number of phrase candidate columns “9” that is not related to the reduced notation processing shown in FIG. make it easier.

FIG. 8 shows two reduced notation phrase candidates "OUIXO" and "OUIX" which succeeded in the above-described reduced notation language processing.
A combination of headwords of the original notation dictionary (see FIG. 2) 9 corresponding to "E" is shown. The clauses listed in parentheses are all the clauses obtained from the reduced notation clause described on the left side.

Next, a phrase candidate in the original notation is obtained from the two reduced phrase clause candidates obtained as described above. This process is performed as follows. That is, first, from the syllable candidates forming the syllable lattice shown in FIG. 3, the syllable candidates for which the reduced notation syllable candidates shown in FIG. 7 cannot be created are deleted by the candidate deleting unit 7. Then, the syllable lattice composed of the remaining syllable candidates is expanded to obtain syllable candidates. In other words,
Creating the reduced notation phrase candidate means performing preliminary selection of a phrase candidate sequence, and the reduced notation language processing unit 5 is the recognition candidate preliminary selection unit.

FIG. 9 shows the syllable lattice shown in FIG.
7 shows the result of deleting the syllable candidates that cannot be created (that is, cannot exist) after the reduced notation phrase processing after the reduced notation language processing shown in FIG. The deletion of this impossible syllable candidate is performed as follows. Since the first syllable candidate of the first syllable in the syllable lattice shown in FIG. 3 (FIG. 9) is / ko /, its reduced notation is / O /. On the other hand, the reduced notation of the first syllable in the reduced notation phrase candidate string shown in FIG. 7 is / O /, and both reduced notations match. Therefore, the syllable candidate / this / is adopted as a possible syllable candidate (in FIG. 9, a circle is added). Similarly, the reduced notation of all syllable candidates after the second syllable candidate of the first syllable and the reduced notation of the corresponding syllable in the reduced notation syllable candidate sequence are sequentially compared.

At this time, if the syllable candidates of the syllable lattice and the corresponding syllables of the reduced syllable syllable candidate sequence are different, the syllable candidate is rejected as an impossible syllable candidate (FIG. 9). Is added with x). However, in the case of the fifth syllable in the reduced notation phrase candidate sequence, two reduced notations, / O / and / E /, are allowed.
If the reduced notation of the fifth syllable in the syllable lattice is / O / or / E /, it is determined that there is a syllable candidate.

Here, as shown in FIG. 9, the two reduced notation phrase candidates “OUIXO” and “OUIX” obtained above are obtained.
The number of syllable candidates constituting the syllable lattice after deleting syllable candidates that do not contribute to "E" is "4" for the first syllable, "3" for the second syllable, "3" for the third syllable, and "4" for the fourth syllable. Since “1” and the fifth syllable are “5”, the number of phrase candidate strings obtained from all combinations of the syllable lattices shown in FIG. 9 is “180 (=
4 × 3 × 3 × 1 × 5) ", which is reduced to 1/4 of the number of phrase candidate strings“ 720 ”obtained by expanding the syllable lattice shown in FIG. The number of times of the original notation language processing performed by the notation language processing unit 8 can be greatly reduced.

FIG. 10 shows an original notation language process performed by the original notation language processing unit 8 on the phrase candidate sequence obtained by expanding the syllable lattice shown in FIG. 9 using the original notation dictionary 9 shown in FIG. The results are shown including the morphological analysis results. Although not directly related to the present invention, detailed description is omitted, but the morphological analysis is also performed in a general kana-kanji conversion process (language process). The "//" part shown in the figure indicates the boundary between the prefix and the independent word in the prefix-independent word chain or the boundary between the independent word and the suffix in the independent word-suffix chain.

Thereafter, the phrase candidate strings obtained as shown in FIG. 10 are displayed in the order of shortest distance, and the target phrase candidate is designated and selected by the operator.

Next, the language processing time according to the present embodiment will be compared with the language processing time according to the conventional method. Here, assuming the language processing speed, the processing speed of one phrase candidate is set to 1 unit / candidate, while the processing speed of characters is set to 0.1 unit / character, and the processing time is expressed in "units". And
The language dictionaries used are the original notation dictionary 9 shown in FIG. 2 and the reduced notation dictionary 6 shown in FIG. 5, and the syllable lattice obtained as a result of recognition is the syllable lattice shown in FIG.

[Conventional Method (Language processing is performed by expanding all syllable lattices)] Linguistic processing = 1 unit / candidate × 720 candidate = 720 units Total = 720 units [Example] Reduced notation of syllable lattice = 0 .1 unit / character x 19 characters = 1.9 unit Reduced notation language processing = 1 unit / candidate x 8 candidates = 8 units Delete impossible syllable candidates from syllable lattice = 0.1 unit / character x 19 characters = 1 .9 units Language processing = 1 unit / candidate × 180 candidates = 180 units Total = 191.8 units That is, according to the present embodiment, the language processing time can be reduced to just over a quarter of the conventional method.

FIG. 11 is a flowchart of a voice recognition / language processing operation performed by the voice recognition device 10 and the language processing device 1 for recognition. Hereinafter, the speech recognition / language processing operation will be described step by step with reference to FIGS.

In step S 1, the speech signal input from the input / output unit 2 is subjected to recognition processing by the speech recognition device 10, and is recognized in syllable units. Step S
In 2, the syllable lattice as shown in FIG. 3 is created by the lattice creating unit 3 based on the recognition result in step S1.

In step S3, the syllabic lattice obtained in step S2 is subjected to reduced notation conversion by the reduced notation conversion unit 4, and a reduced notation syllable lattice as shown in FIG. 6 is created. In step S4, the reduced notation syllable lattice created in step S3 is expanded by the reduced notation language processing unit 5, and the obtained reduced notation phrase candidate sequence is subjected to reduced notation language processing by using the reduced notation dictionary 6. Will be implemented.

In step S5, the candidate deletion unit 7 refers to the reduced notation phrase candidate (reduced notation phrase candidate sequence marked with a circle in FIG. 7) that has been successfully processed in the reduced notation language processing in step S4. Then, impossible syllable candidates are deleted from the syllable lattice created in step S2. In step S6, the syllable lattice from which the syllable candidates that cannot exist in step S5 are deleted is expanded by the original notation language processing unit 8, and the obtained syllable candidate sequence is converted to the original notation language using the original notation dictionary 9 Processing is performed. In this way, a reliable phrase candidate sequence can be obtained in a short time from a phrase candidate sequence composed of many combinations of syllable candidates.

As described above, in the present embodiment, the syllabic lattice is converted into the reduced notation by the reduced notation conversion unit 4 and the reduced notation language processing unit 5 executes the reduced notation language processing unit 5 based on the obtained reduced notation syllable lattice. To obtain a reduced notation clause candidate sequence. As described above, the recognition result is once represented by the reduced notation by the reduced notation, and the language processing (reduced notation language processing) is performed in a short time. Then, with reference to the reduced notation syllable candidate sequence that has been successfully subjected to the reduced notation linguistic treatment, the candidate deleting unit 7 deletes impossible syllable candidates from the syllabic lattice in the original notation, and calculates the number of syllable candidates constituting the syllable lattice Reduce.

Therefore, thereafter, the expansion of the syllabic lattice and the processing of the original syllabic language performed by the original syllabic language processing section 8 can be executed in a short time by the syllable lattice composed of a small number of syllable candidates.

<Second Embodiment> FIG. 12 is a block diagram of a recognition language processing apparatus according to this embodiment. The input / output unit, lattice creation unit, reduced notation conversion unit, reduced notation language processing unit, and reduced notation dictionary are the same as those of the recognition word post-processing device 1 shown in FIG. Will be described. The input / output unit 2 sends syllable candidates from the speech recognition device 10 to the lattice creation unit 3, and the lattice creation unit 3 creates a lattice from the input syllable candidates. The reduced notation conversion unit 4 converts the notation of the syllable lattice from the lattice creation unit 3 into reduced notation according to a predetermined rule. The reduced notation language processing unit 5 expands the reduced notation syllable lattice obtained by the reduced notation conversion unit 4 and performs reduced notation language processing.

The analysis unit 12 analyzes the structure of the phrase candidate sequence based on the reduced phrase candidate sequence from the reduced notation language processing unit 5 and creates a word chain using the reduced notation. The original notation conversion unit 13 converts the word chain based on the reduced notation created by the analysis unit 12 into the original notation word corresponding to the reduced notation headword. The top-down processing unit 14 calculates the syllable lattice created by the lattice creation unit 3 and the original notation word obtained by the original notation conversion unit 13,
Top-down evaluation is performed as described in detail below.

The score calculation unit 15 calculates the top-down evaluation result by the top-down processing unit 14 and the analysis unit 12
Then, a score is calculated for each phrase candidate based on the analysis result from. Then, sorting section 16 sorts the phrase candidate strings to which the scores from score calculation section 15 have been added in ascending order of the scores.

The phrase candidate strings arranged in ascending order of the scores are output via the input / output unit 2 to an external device (device not shown for obtaining a recognition result).

Next, the analysis unit 12, the original notation conversion unit 13,
Each process performed by the top-down processing unit 14 and the score calculation unit 15 will be described in detail. Note that the contents of the reduced notation dictionary 6, the contents of the syllable lattice created by the lattice creation unit 3, and the processes performed by the reduced notation conversion unit 4 and the reduced notation language processing unit 5 have already been described in the first embodiment. Are described in detail above, and a description thereof will be omitted.

FIG. 13 shows a result of the analysis unit 12 analyzing the reduced notation phrase candidate string obtained as a result of the reduced notation language processing in the reduced notation language processing unit 5 as shown in FIG.
FIG. 6 is a diagram also showing original notation word candidates converted by the original notation conversion unit 13. That is, the analysis unit 12 and the original notation conversion unit 13 constitute the recognition candidate generation unit.

FIGS. 14 and 15 show the results of the top-down evaluation performed by the top-down processing unit 14. FIG. FIG.
Is a top-down evaluation result for the part of the independent word "OUIX" shown in FIG. The top-down evaluation is to perform top-down linguistic processing on the original notation word candidates “Kokumin”, “Koumin”, and “Toumin” corresponding to, for example, the reduced notation “OUIX”. On the other hand, the linguistic processing conventionally performed is a bottom-up linguistic processing in which linguistic processing is performed on all syllable candidates obtained from all combinations of syllable candidates.

The above-mentioned top-down evaluation will be described in detail.
First, in the case of “Kokumin” shown in FIG. 14A, the first syllable is “Ko”, and the first syllable in the syllable lattice shown in FIG.
From the syllable candidates, the distance of the first syllable is determined to be “1.0”. Similarly, the distances between the second syllable, the third syllable, and the fourth syllable are obtained. Then, the distances of the first to fourth syllables thus obtained are added to obtain the distance “4.2” of the word candidate “Kokumin”. Similarly,
A top-down evaluation is performed on the word candidates “Komin” and “Tomin”. However, in the case of the word candidate “Komin” shown in FIG. 14B, the second syllable “U” does not exist in the syllable lattice shown in FIG. 3, and the word candidate “Tou” shown in FIG. In the case of "min", the first syllable "to" and the second syllable "u" do not exist in the syllable lattice. Therefore, since the distance cannot be calculated, it is rejected.

FIG. 15 shows a top-down evaluation result with respect to other parts in FIG. 13, and FIG. 15 (a) shows a prefix "O".
15 is a top-down evaluation result for the portion of FIG.
FIG. 15B shows a top-down evaluation result for the independent word “UX”, and FIG. 15C shows a top-down evaluation result for the suffixes “O” and “E”.

FIG. 16 is a diagram collectively showing the result of the structural analysis of the phrase candidate by the analysis unit 12, the result of the original notation conversion by the original notation conversion unit 13, and the result of the top-down evaluation by the top-down processing unit 14.

FIG. 17 shows a score calculated by the score calculation unit 15 by adding the distances of the word candidates constituting each phrase candidate created by combining the word candidates shown in FIG. And phrase candidates as a result of sorting in ascending order of each score.

Next, the language processing time according to the present embodiment will be compared with the language processing time according to the conventional method. Here, the assumption of the language processing speed is the same as that of the first embodiment.
Further, a language dictionary to be used is a reduced notation dictionary 6 shown in FIG.
And the syllable lattice obtained as a result of the recognition is the syllable lattice shown in FIG.

[Conventional method (to perform language processing by expanding all syllable lattices)] Language processing = 1 unit / candidate × 720 candidate = 720 units Total = 720 units [Example] Reduced notation of syllable lattice = 0 .1 unit / character x 19 characters = 1.9 unit Reduced notation language processing = 1 unit / candidate x 8 candidates = 8 units Top-down processing = 1 unit / candidate x 10 candidates = 10 units Score calculation = 1 unit / candidate x 9 candidates = 9 units Total = 28.9 units That is, according to the present embodiment, the language processing time can be reduced to a little more than 1/25 of the conventional method.

FIG. 18 is a flowchart of a speech recognition / language processing operation performed by the speech recognition device 10 and the recognition language processing device 11. Hereinafter, the speech recognition / language processing operation will be described step by step with reference to FIGS.

In step S11, the speech signal input from the input / output unit 2 is subjected to recognition processing by the speech recognition device 10 to be recognized in syllable units. Step S
At 12, based on the recognition result at step S11,
The syllable lattice as shown in FIG.

In step S13, the syllabic lattice obtained in step S12 is subjected to reduced-notation conversion by the reduced-notation conversion unit 4 to create a reduced-notation syllable lattice as shown in FIG. In step S14, the above step S13
The reduced notation syllable lattice created in step 2 is expanded by the reduced notation language processing unit 5, and the reduced notation phrase candidate sequence obtained is subjected to reduced notation language processing using the reduced notation dictionary 6.

At step S15, the analyzing unit 12 performs an analyzing process. That is, the parts having the same structure in the reduced notation phrase candidates that have been successfully processed in the reduced notation language processing in step S14 are put together and shared, and a word chain based on the reduced notation is obtained. In step S16, the original notation conversion unit 13 converts the word chain in reduced notation obtained in step S15 into a word chain in original notation. Note that this original notation conversion is performed using a restoration dictionary (not shown) in which word chains in reduced notation and original word word chains are associated with each other.

In step S17, the top-down processing unit 14 refers to the syllable lattice obtained in step S12, and performs top-down linguistic processing on each word in the original notation obtained in step S16. Is performed. Then, the distance between the established original written words is obtained. In step S18, the score of the original notation phrase candidate is calculated by the score calculation unit 15 based on the word chain based on the reduced notation obtained in step S15 and the distance between each of the original notation words obtained in step S17. You. In step S19, the sorting unit 16 rearranges the reduced notation phrase candidates to which the scores obtained in step S18 are added in ascending order of the scores, and obtains the final recognition result. In this manner, a correct recognition result can be obtained in a short time from a phrase candidate sequence composed of a combination of many syllable candidates.

As described above, in the present embodiment, the syllabic lattice is converted into the reduced notation by the reduced notation conversion unit 4, and the reduced notation language processing unit 5 executes the reduced notation language processing unit 5 based on the obtained reduced notation syllable lattice. To obtain a reduced notation clause candidate sequence. As described above, the recognition result is once represented by the reduced notation by the reduced notation, and the language processing (reduced notation language processing) is performed in a short time. Then, the analysis unit 12
In this way, the structure of the reduced notation phrase candidate sequence that has been successfully processed in the reduced notation language processing is analyzed, and this analysis result is used as the original notation conversion unit 13.
To the original notation. In this way, the original written word candidate string to be subjected to the language processing is specified in advance.

Therefore, after that, the language processing for the original written word candidate can be performed in a short time by the top-down processing section 14 with a small number of processes in a top-down manner.

In the first and second embodiments, the syllable lattice created by the lattice creation unit 3 based on the speech recognition result by the speech recognition device 10 is reduced by the reduced notation syllable by the reduced notation conversion unit 4. I try to convert it to lattice. Here, converting into the reduced notation means converting the original notation into a notation having a smaller number of categories than the original notation.

Therefore, as a modified example, a rough recognition result is obtained first by performing rough recognition (that is, a recognition result with a small number of categories such as reduced notation).
By performing language processing on the result, the same effect as the reduced notation language processing by the reduced notation language processing unit 5 can be obtained. In this case, it is necessary to perform a detailed recognition process thereafter, and a two-stage recognition method is used.

<Third Embodiment> In this embodiment, if there is a word candidate to be rejected as a result of the top-down evaluation in the second embodiment, this word candidate is prevented from being rejected. .

In the above-described second embodiment, when the top-down processing section 14 performs the top-down processing on the word candidate "Komin", the second syllable "U"
Does not exist in the syllable lattice shown in. Therefore, in the present embodiment, as shown in FIG. 19, a syllable candidate "U" is registered with a predetermined distance "3.0" added to the second syllable in the syllable lattice. In this way, the phrase candidate including the word candidate “Komin” is prevented from leaking out of the recognition candidate.

As the predetermined distance to be added to the new syllable candidate, a large value is simply given, or the distance between similar syllables when all syllables are the first syllable candidate. Is prepared in advance, and the distance between the first syllable candidate at the corresponding syllable position of the syllable lattice and the syllable candidate to be newly added is obtained from the distance table.

<Fourth Embodiment> In the present embodiment, when converting into the reduced notation, syllables that are likely to appear as recognition candidates for the same syllable are converted into the reduced notation so as to have the same reduced notation. The processing time of the language processing is reduced.

For example, as shown in FIG.
Line, “Za” line and “Cha” line have reduced notation / S /
In the original notation other than the original notation converted into the reduced notation / S /, the reduced notation / A * / is given to the “A” stage, the reduced notation / I * / is given to the “I” stage, and "U" is given a reduced notation / U * /, "E" is given a reduced notation / E * /, "O" is given a reduced notation / O * /, and "n" is a reduced notation The reduced notation assignment rule is set so as to give / X /. Such a reduced notation assignment rule is a reduced notation assignment rule created from the tendency of misrecognition that syllable candidates obtained when recognizing syllables / sa / often include / shi / ya / su /. .

By doing so, for example, the syllable candidates of a certain syllable obtained as a result of speech recognition are, as shown in FIG. 21 (a), / sa /, / za /, / su /, / set /, / so / , / Cha / and / da /, the first to third embodiments as shown in FIG.
According to the reduced notation conversion based on the reduced notation in the embodiment, as shown in FIG. 21C, it is converted into four reduced notations / A /, / U /, / E / and / O /. On the other hand, according to the reduced notation conversion by the reduced notation in the present embodiment, FIG.
As shown in (b), it can be reduced to two reduced notations / S / and / A * /. Thus, according to the present embodiment, efficient reduced notation conversion is possible.

Incidentally, as the preprocessing conventionally performed,
When developing a recognition result lattice to generate a phrase candidate or a word candidate, a transition probability between two syllables or three syllables that has been obtained in advance is used to remove a combination of almost impossible recognition candidates. There is processing. There is no problem if such a pretreatment is used in the present invention.

The algorithm of the speech recognition / language processing operation in the present invention is not limited to the algorithm shown in FIG. 11 or FIG. In each of the above embodiments, the linguistic processing is performed in units of phrase candidates. However, the present invention is not limited to this, and there is no problem even in units of words or sentences. Further, in each of the above embodiments, the case where the audio signal is recognized in units of syllables has been described as an example, but the present invention is not limited to this. For example, recognition may be performed in phoneme units using phoneme recognition using an HMM (Hidden Markov Model).

In each of the above embodiments, the case where a speech signal is recognized to obtain a phrase candidate sequence is described as an example. However, the present invention can also be applied to a case where a phrase candidate sequence or a word candidate sequence is created by recognizing an image signal in units of characters / symbols.

[0077]

As is clear from the above, in the recognition language processing apparatus of the first invention, the notation conversion unit converts the notation of the recognition result from the recognition apparatus into a notation of a smaller number of categories than the number of categories to which the notation belongs. Is converted according to a predetermined rule, and the recognition candidate preliminary selection unit converts the notation converted by the notation conversion unit into a headword of a language processing dictionary used when performing language processing. The language processing is performed using the compressed dictionary converted according to the rules of the above, to obtain the pre-selection recognition candidates, and the pre-selection recognition candidates are reconstructed into the chain of the headwords of the language processing dictionary by the recognition candidate generation unit for recognition. A candidate is generated, and a top-down processing unit calculates the weight of the generated recognition candidate in a top-down manner with reference to the recognition result, thereby performing final recognition. Since to obtain a candidate can be performed in a short time by converting the language processing section representation of a small number of categories using the compression dictionary. Furthermore, since the weights of the recognition candidates generated based on the preliminary selection recognition candidates can be calculated top-down, bottom-up language processing is not required, and the final recognition candidates are obtained in a very short time. be able to.

Therefore, according to the present invention, it is possible to provide a recognition language processing apparatus capable of performing high-speed language processing on a recognition result.

Further, in the recognition language processing apparatus according to the second invention, the recognition result from the recognition apparatus is composed of recognition unit groups each weighted, and the top-down processing unit calculates the weight of the recognition candidate. When there is a recognition unit not included in the recognition result among the plurality of recognition units constituting the recognition candidate, the recognition result adding unit adds a predetermined weight to the recognition unit and automatically adds the recognition unit to the recognition result. Since the addition is performed, correct recognition candidates are not rejected, and the weights of all the recognition candidates generated by the recognition candidate generation unit can be calculated from the top down.

Therefore, according to the present invention, it is possible to provide a recognition language processing apparatus capable of performing correct language processing on a recognition result at higher speed.

Further, in the recognition language processing apparatus according to a third aspect of the present invention, the predetermined rule used when the notation conversion unit converts the notation of the recognition result from the recognizing apparatus includes the following: Since the notation of the recognition result applicable to the existing pattern of erroneous recognition is a rule that is converted into a notation that belongs to one category, the recognition result has a smaller number of categories according to the tendency of recognition errors by the recognition device. It is efficiently converted to notation.

Therefore, according to the present invention, it is possible to provide a recognition language processing apparatus capable of performing language processing on a recognition result at higher speed.

[Brief description of the drawings]

FIG. 1 is a block diagram of a recognition language processing apparatus according to the present invention.

FIG. 2 is an explanatory diagram of contents of an original notation dictionary in FIG. 1;

FIG. 3 is a diagram illustrating an example of a syllable lattice created by a lattice creating unit in FIG. 1;

FIG. 4 is a diagram showing an example of a syllable candidate sequence obtained when the syllable candidate sequence obtained by expanding the syllable lattice shown in FIG. 3 is subjected to the original notation language processing using the original notation dictionary shown in FIG. 2; It is.

FIG. 5 is an explanatory diagram of the contents of a reduced notation dictionary in FIG. 1;

6 is a diagram showing an example of a reduced syllable lattice obtained by performing a reduced notation conversion on the syllable lattice shown in FIG. 3;

7 is a reduced notation phrase candidate obtained by performing reduced notation language processing using the reduced notation dictionary shown in FIG. 5 on a reduced notation phrase candidate string obtained by expanding the reduced notation syllable lattice shown in FIG. 6; It is a figure showing an example of a column.

FIG. 8 is a correspondence diagram between reduced notation phrase candidates obtained as a result of reduced notation language processing and original notation phrase candidates.

FIG. 9 is an explanatory diagram of the result of removing impossible syllable candidates from the syllable lattice shown in FIG. 3;

10 is a syllable candidate sequence obtained by expanding a syllable lattice shown in FIG. 9 and performing a morphological analysis by performing an original notation language process using the original notation dictionary shown in FIG. It is a figure showing an example of a column.

FIG. 11 is a flowchart of a voice recognition / language processing operation performed by the voice recognition device and the recognition language processing device in FIG. 1;

FIG. 12 is a block diagram of a recognition language processing device different from FIG. 1;

FIG. 13 is a diagram illustrating a reduced notation phrase candidate sequence shown in FIG. 7;
FIG. 6 is a diagram illustrating an analysis result by an analysis unit in FIG.

FIG. 14 is a diagram illustrating an example of a top-down evaluation result of the original notation word candidate by the top-down processing unit in FIG. 12;

FIG. 15 is a diagram illustrating an example of a top-down evaluation result different from that of FIG. 14;

FIG. 16 is a diagram collectively showing a result of a structural analysis of a phrase candidate, a result of an original notation conversion, and a result of a top-down evaluation.

17 is a diagram illustrating an example of a phrase candidate sequence in which scores are sorted by a sorting unit after scores are calculated by a score calculating unit in FIG. 12;

FIG. 18 is a flowchart of a voice recognition / language processing operation performed by the voice recognition device and the recognition language processing device in FIG. 12;

19 is an explanatory diagram of registering a new syllable candidate when no syllable candidate exists in the syllable lattice at the time of the top-down evaluation by the top-down processing unit in FIG. 12;

FIG. 20 is an explanatory diagram of two different reduced notation provision rules.

FIG. 21 is an explanatory diagram of a reduced notation conversion result based on the two reduced notation provision rules shown in FIG. 20;

[Explanation of symbols]

Reference numerals 1, 11: recognition language processing unit, 3: lattice creation unit, 4: reduced notation conversion unit, 5: reduced notation language processing unit, 6: reduced notation dictionary, 7: candidate deletion unit, 8: original notation language processing unit Reference numeral 9: Original notation dictionary, 10: Voice recognition device, 12: Analysis unit, 13: Original notation conversion unit, 14: Top-down processing unit, 15: Score calculation unit, 16: Sorting unit.

──────────────────────────────────────────────────続 き Continuation of the front page (56) References JP-A-2-297598 (JP, A) JP-A-58-87599 (JP, A) JP-A-63-265299 (JP, A) JP-A-2-297 186400 (JP, A) JP-A-51-136208 (JP, A) Patent 2815579 (JP, B2) JP-B 5-67040 (JP, B2) JP-B 7-66276 (JP, B2) The Institute of Electronics, Information and Communication Engineers Transactions Vol. J 67-D No. 8 “Vocabulary reduction effect by specifying phoneme sequences in words” p. 869-876 (1984/8) (58) Field surveyed (Int. Cl. ⁷ , DB name) G10L 3/00-9/20 JICST file (JOIS)

Claims (3)

(57) [Claims] 1. A recognition language processing apparatus for obtaining a recognition candidate by performing language processing on a recognition result from a recognition apparatus for recognizing speech or characters based on an input speech signal or image signal, comprising the steps of: A notation conversion unit that converts the notation of the recognition result into a notation having a smaller number of categories than the number of categories to which the notation of the recognition result belongs according to a predetermined rule; and a headword of a language processing dictionary used when performing the language processing. The notation is converted according to the predetermined rule, the number of categories to which the notation of the headword belongs is compressed, and the compression dictionary is used for the recognition result whose notation is converted by the notation conversion unit. A recognition candidate preliminary selection section for performing language processing using the recognition candidate preliminary selection section, and a preliminary selection section obtained by the recognition candidate preliminary selection section. A recognition candidate generation unit that obtains the recognition candidate by restoring the recognition candidate into a chain of headwords of the language processing dictionary; and a recognition candidate obtained by the recognition candidate generation unit with reference to a recognition result from the recognition device. A top-down processing unit that calculates the weight of the top-down and obtains the final recognition candidate. 2. The recognition language processing device according to claim 1, wherein the recognition result from the recognition device includes a recognition unit group each weighted, and a distance of the recognition candidate is determined by the top-down processing unit. When calculating, if there is a recognition unit not included in the recognition result among a plurality of recognition units constituting the recognition candidate,
A recognition language processing apparatus comprising: a recognition result adding unit that adds a predetermined weight to the recognition unit and automatically adds the recognition result to the recognition result. 3. The recognition language processing device according to claim 1, wherein the predetermined rule used when the notation conversion unit converts a notation of a recognition result from the recognition device includes: If there is an erroneous recognition pattern in the recognition by the recognition device,
A language processing device for recognition, characterized in that the rule is converted into a notation that belongs to one category when the notation of a recognition result that fits the pattern is converted.