JP2950823B1 - Speech recognition error correction device - Google Patents

Abstract

A speech recognition error correction device capable of determining the validity of a global speech recognition result as compared with a conventional example and performing error correction processing. SOLUTION: A structure analysis processing unit (2) performs a structure analysis process on a speech recognition result to generate an input analysis tree, and an error correction processing unit (3, 7, 8, 9, 4) uses an input analysis tree. A subtree pair that is phonologically close between each input subtree and each example subtree based on the input parse tree and the example parse tree to obtain a compatible structure between the input parse tree and the example parse tree Error correction is performed based on the error correction. Semantic distance calculation processing and output means (5,
6) calculates the global validity of the entire sentence by calculating the semantic distance calculated for the entire uttered speech sentence including the error correction result depending on the positional relationship of synonyms between words in the concept hierarchy. Judgment is performed, and when the calculated semantic distance is equal to or smaller than a predetermined threshold, an analytic tree of the uttered speech sentence including the error correction result is output.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition error correction device for recognizing a speech signal of an uttered speech sentence and performing an error correction process on the speech recognition result.

[0002]

2. Description of the Related Art For example, in a speech translation apparatus, a translation module needs to accept an utterance containing an error attributed to a speech recognition process. Therefore, the development of a strong error handling framework is extremely important in improving the performance of a speech translation system.

[0003] For example, in the prior art document 1 "H. Saito
et al. , "ParsingNoisy Sen
tences ", Proceedings of CO
LING '88, pp. 561-566, 1988 "
In, a correct hypothesis of an erroneous part is generated for a phoneme string, and the hypothesis is verified based on a result of the syntax analysis, but there is a problem that a correctable region is local.

[0004] In order to solve this problem, prior art document 2, "Yumi Wakita et al.," Introduction to post-processing correct part recognition method for speech recognition using semantic similarity and speech translation method ",
Technical Committee on Spoken Language Information Processing, Japan Society for Artificial Intelligence, 17-4, 19
On July 18, 1997, a method for identifying only a correct answer part from a result sentence after the speech recognition processing for the purpose of analyzing an erroneous sentence when the spoken word is speech-recognized has been proposed. This method considers that the grammatical description of spoken words is difficult and that the language processing for speech recognition restricts only the local part of the sentence. Using the dependencies between words determined from the closeness of the semantic distance, the correct answer is obtained by judging the naturalness of the expression in a broader range compared to the linguistic constraint range in speech recognition Identify the part. In this method, a language constituent boundary parser (Constituent Bound) disclosed in JP-A-6-274192 is disclosed.
ary parser; hereinafter, referred to as CB parser. ) Is used to translate only the reliable parts in the utterance determined according to the semantic distance calculated for each local subtree.

[0005]

However, in this method, the recall rate of the translated portion with respect to the entire utterance is low, and the translation may not be output. In addition, although the correct part is specified using the semantic distance, there is a problem that there is no framework for correcting the recognition error.

SUMMARY OF THE INVENTION An object of the present invention is to solve the above-mentioned problems, to make it possible to judge the validity of a general speech recognition result as compared with the conventional example, and to carry out an error correction process. It is to provide a device.

[0007]

According to a first aspect of the present invention, there is provided a speech recognition error correction apparatus for performing a speech recognition of a speech signal of an uttered speech sentence and performing an error correction process on the speech recognition result. In the error correction device, a storage means (22) for storing an example parse tree obtained by performing a predetermined structure analysis process on a predetermined example sentence, and performing a predetermined structure analysis process on the speech recognition result. (2) for generating and outputting an input parse tree by means of an input parse tree, an input parse tree output from the structural analytic process means (2), and an example parse tree stored in the storage means (22). Based on the input parse tree, a subtree pair that is phonologically close between each input subtree that is a subtree of the input analytic tree and each example subtree that is a subtree of the example analytic tree is searched. Based on the interchangeable structure between the tree and the example parse tree Error correction processing means (3, 7 outputs an error correction result of correction processing performed input parse tree Ri,
8, 9, 4) and the error correction processing means (3, 7, 8,
For the entire utterance sentence including the error correction result output from (9, 4), the semantic distance calculated depending on the positional relationship of the synonyms between words in the concept hierarchy is calculated, so that the overall sentence overall is calculated. Semantic distance calculation processing and output means (5, 6) for determining an appropriate validity and outputting a parse tree of an uttered speech sentence including an error correction result when the calculated semantic distance is equal to or less than a predetermined threshold value. ).

Further, the speech recognition error correction device according to the second aspect is the speech recognition error correction device according to the first aspect,
The error correction processing means (3, 7, 8, 9, 4) calculates a semantic distance for the input parse tree output from the structural analysis processing means (2), and calculates the calculated semantic distance. When the threshold value is exceeded, the input parse tree repair determining means (3) for determining that the input parse tree should be repaired and the input parse tree repair determining means (3) determine that the input parse tree should be repaired. Based on the input parse tree and the example parse tree stored in the storage means, a subtree pair that is phonologically close between each input subtree and each example subtree is searched, and each subtree is searched. An input subtree and an example subtree matching processing unit (7) for generating and outputting a pair of combinations of the above, and an input subtree and an example subtree output from the input subtree and example subtree matching processing unit (7) With reference to a predetermined word-phoneme conversion dictionary A subtree phoneme string conversion processing unit (8) that converts each of them into a phoneme string and outputs them; a phoneme string of the input subtree and a phoneme string of the example subtree output from the subtree phoneme string conversion processing unit (8) And a phonological distance calculation processing unit (9) for calculating a phonological distance between the input subtrees and outputting a pair of an input subtree and an example subtree in which the calculated phonological distance is equal to or smaller than another predetermined threshold value. For the input parse tree determined to be repaired by the input parse tree repair determining means (3), the input subtree output from the phonological distance calculation processing unit (9) and the example subtree are compared. An analysis tree restoration processing unit (4) that performs an error correction process by performing restoration of the input analysis tree using the pair and outputs an error correction result of the input analysis tree.

[0009]

Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing the configuration of a speech recognition error correction apparatus according to an embodiment of the present invention. The speech recognition error correction device according to the present embodiment includes (a) a speech recognition processing unit 1, (b) a structural analysis processing unit 2, (c) an input parse tree repair determination unit 3, and (d) a parse tree repair process. Part 4,
(E) a semantic distance calculation processing unit 5, (f) a correction analysis tree output unit 6, (g) an input subtree and example subtree collation processing unit 7, and (h) a subtree phoneme string conversion processing unit 8 (I) a phonological distance calculation processing unit 9, (j) a structure analysis processing unit 10,
It is characterized by having.

In FIG. 1, an uttered voice (consisting of a natural language sentence such as Japanese or English) to be subjected to voice recognition and error correction is input to a microphone 30 and converted into a voice signal. Is input to The speech recognition processing unit 1 is a known speech recognition device as disclosed in, for example, Japanese Patent Application Laid-Open No. 9-134192, and performs A / D conversion on an input speech signal, performs LPC analysis, and performs acoustic feature analysis. After extracting the parameters, a phoneme sequence is specified using a phoneme hidden Markov model, and a speech recognition result of a natural language sentence composed of a word sequence is obtained using a predetermined statistical language model, and output to the structure analysis processing unit 2. I do.

The structure analysis processing unit 2 has a configuration shown in FIG. 2, and relates a relation between each word in the sentence to a natural language sentence composed of a word string of a speech recognition result output from the speech recognition processing unit 1. Analyze the receiving structure. Here, the language pattern is collated with respect to all parts of the input sentence, and the combination result of the collated language patterns is determined as a structure candidate of the input sentence. From the structure candidates selected in this way, the semantic distance between each word of the input sentence and the example word of the language pattern is determined, and the one having the smallest semantic distance is determined as the optimal structure and output. I do. As a method of calculating the semantic distance, a method described later in detail is used.

In FIG. 2, a natural language sentence composed of a character string of a predetermined language such as Japanese or English and composed of a plurality of words, which is speech-recognized by the speech recognition processing unit 1, is converted into a morpheme in the structure analysis processing unit 2. The morphological analysis unit 101 inputs the appearance form of each word of the input natural language sentence (hereinafter referred to as an input sentence) to the morphological analysis dictionary stored in the morphological analysis dictionary memory 111. See
The natural language sentence is divided into a plurality of words, and information such as a part of speech, a standard form or a standard expression form, an inflected form, and a synonym code is added to each of the appearance forms, and the information is output to the part of speech bigram insertion unit 102. . Table 1 shows a specific example of an English morphological analysis dictionary, and Table 2 shows a specific example of a Japanese morphological analysis dictionary.

[0014]

[Table 1] Example of English morphological analysis dictionary ―――――――――――――――――――――――――――――――― Appearance form Standard form Part-of-speech inflected form synonym code ―――――――――――――――――――――――――――――――――― the the article bus bus ordinary noun 123 leaves leave verb Present tense 234 Kyoto Kyoto Proper noun 333 at at Preposition eleven eleven number 400 a. ma. m. Suffix 555 ――――――――――――――――――――――――――――――――

[0015]

[Table 2] Specific examples of Japanese morphological analysis dictionary ―――――――――――――――――――――――――――― Appearance type Standard part of speech Participatory utilization Synonym code ―――――――――――――――――――――――――――― This is the pronoun 007 is a particle post office secretariat The secretariat is a common noun 700 Auxiliary verb final form ――― ―――――――――――――――――――――――――――

Next, the part-of-speech bigram insertion unit 102
Based on the part-of-speech information of the words of the input sentence obtained from the morphological analysis unit 101, the part-of-speech bigram as a language constituent boundary is inserted into the input sentence with reference to the part-of-speech bigram dictionary stored in the part-of-speech bigram dictionary memory 112. Then, it is output to the language pattern search unit 103 together with the input sentence. Here, the part-of-speech bigram refers to a combination of parts of speech of two words that may be adjacent in the input sentence. Table 3 shows a specific example of an English part-of-speech bigram dictionary, and Table 4 shows a specific example of a Japanese part-of-speech bigram dictionary. In the part-of-speech bigram insertion unit 102, the restrictive conditions shown in Table 5 are provided for the part-of-speech A and part-of-speech B of the part-of-speech bigram AB to be inserted, and the part of speech A and the part of speech B do not belong to the same language constituent. To insert a part-of-speech bigram. In the part-of-speech bigram dictionary stored in the part-of-speech bigram dictionary memory 112, the part-of-speech bigram may be stored under the restrictive conditions shown in Table 5 so that the part-of-speech A and the part-of-speech B are restricted so as not to belong to the same language constituent. Good.

[0017]

[Table 3] Specific examples of English part-of-speech bigram dictionaries (list of part-of-speech bigrams that can be language constituent boundaries in English) ―――――――――――――――――――――――― ―――――――――― noun-verb noun-auxv propn-verb verb-propn ………………………………… ―――――――――――――――― ――――――――――――――――――――

In Table 3 above, noun, verb,
propn and auxv are symbols representing common nouns, verbs, proper nouns, and auxiliary verbs, respectively.

[0019]

[Table 4] Specific examples of Japanese part-of-speech bigram dictionaries (lists of part-of-speech bigrams that can become language constituent boundaries in Japanese) ――――――――――――――――――――――― ――――――――――― pron-noun noun-verb ……………………………………………… ―――――――――――――――

In Table 4 above, pron, noun,
verb is a symbol representing a pronoun, a common noun, and a verb, respectively.

[0021]

[Table 5] Restrictions on A and B in part-of-speech bigram AB ―――――――――――――――――――――――――――――――――― (C1) Neither A nor B can be used as a preposition, particle, conjunction, etc. that associates the preceding and following language constituents. ―――――――――――――――――――――――――――――――――― (C2) Articles, adnominals, and prefixes that modify the language constituents after Words and English auxiliary verbs cannot be A. ―――――――――――――――――――――――――――――――― (C3) Suffix to modify the previous language constituent, Japanese The auxiliary verb cannot be B. ――――――――――――――――――――――――――――――――――

In the part-of-speech bigram insertion unit 102, for example, "t" in the English expression "the bus"
No part-of-speech bigram is inserted between "he" (article) and "bus" due to the restriction condition of (C2) above.

Next, the language pattern search unit 103 refers to the language pattern search table, which is stored in the language pattern memory 113 and describes the relationship between the input word and the collatable language pattern, and refers to each word of the input sentence and the language constituents. From the part-of-speech bigram inserted as a boundary, a language pattern that can be matched with the input sentence is searched, and the searched language pattern is output to the language pattern matching unit 104 together with the input sentence. Table 6 shows a specific example of the language pattern search table for English, and Table 7 shows a specific example of the language pattern search table for Japanese.

[0024]

[Table 6] English language pattern search table ―――――――――――――――――――――――――――――――――― Word Language pattern (linguistic Unit) ―――――――――――――――――――――――――――――――― the the X (compound) noun-verb X noun-verb Y (single sentence) verb-propn X verb-propn Y (verb phrase) at X at Y (verb phrase, noun phrase) a. m. X a. m. (Compound words) of X of Y (verb phrase, noun phrase) ……… ………… ―――――――――――――――――――――――――――― ――――――

[0025]

[Table 7] Language pattern search table for Japanese ―――――――――――――――――――――――――――――――――― Word Language pattern (language Unit) ―――――――――――――――――――――――――――――――― where X is Y (verb phrase) pron-noun X pron -Noun Y (verb phrase) It is X (single sentence) X of Y (noun phrase) ……… ……… ――――――――――――――――――――――― ―――――――――――

The language pattern search unit 103 searches for the language pattern “X of Y” from the English word “of”, and searches and outputs the language pattern “X of Y” from the Japanese word “no”, for example. I do.

Further, the language pattern matching unit 104 converts the plurality of language patterns retrieved by the language pattern retrieval unit 103 and the surface words contained in the input sentence from those having a relatively large linguistic unit to a relatively small one. Based on the combination that matches and matches the input sentence with the input sentence, and efficiently determines a structure candidate consisting of a combination of a language pattern and a surface phrase for the input sentence, and determines the maximum likelihood. Output to the structure determination unit 105.

The language pattern matching processing by the language pattern matching unit 104 is characterized in that a language pattern is matched with an input sentence from the largest linguistic unit to the smallest linguistic unit. If the matching is successful for all parts of the input sentence, the combination result of the checked language patterns is determined as a structure candidate of the input sentence. For example,
In the case of the Japanese sentence "Please send the guidebook", the single sentence-based language pattern "X te please" having the largest linguistic unit among the collatable language patterns is first collated with the entire input sentence. Next, the language pattern "X to Y" of the case relation unit is collated with "send a guidebook" which embodies the variable X of "X please". Further, a variable X of "X to Y",
The surface words “guide book” and “send” are matched and matched with Y, respectively, which means that the input sentence is completely matched with the combination of language patterns. As a result, "(((Guide) (send))") is obtained as a structure candidate for the input sentence.

Further, the maximum likelihood structure determining unit 105 obtains a semantic distance between each word of the input sentence and the example word of the language pattern from among the structure candidates obtained by the language pattern matching unit 104, A structure having a small sum of the semantic distances is determined as an optimal maximum likelihood structure and output. As a method of calculating the semantic distance, a known method is used as follows.

The semantic distance related to the language pattern is a value obtained by calculating the semantic distance between the input word and the example word for each variable of the language pattern, and multiplying the calculated semantic distance between words by a weight. . The semantic distance between words gives a value from 0 to 1 depending on the positional relationship of synonyms between words in the concept hierarchy. If two words belong to exactly the same concept, the semantic distance is 0, and if the concept is completely unrelated, it is 1. The semantic distance between words is calculated based on a synonym code indicating a position in the concept hierarchy. This synonym code is added by the morphological analysis unit 101 as described above.

Hereinafter, a specific example of a method of calculating the semantic distance between words will be described. The calculation of the semantic distance between the input word and the example word in the language pattern is performed for the central word that embodies the variable. Language pattern "X at Y"
(Verb phrase) is "leavesverb-propn K
yoto at eleven a. m. When matching with ", the character strings that embody the variables X and Y are
leaves verb-propn Kyoto ”
And "eleven am." In each language pattern, a part to be a central phrase is specified. For example, in "X verb-propn Y", X, "X
a. m. "In" a. m. Therefore, "leaves verb-propn Kyot
o "and" eleven a. m. The central word of "" is "le
aves "and" a. m. ", Which is an input for calculating the semantic distance in" Xat Y ".

Language pattern "X at Y" (verb phrase)
Is an example of embodying the variable terms X and Y as "X = pre
sent, Y = conference, "and" X = ar
live, Y = p. m. "Are stored in advance in, for example, a language pattern search table. At this time,
The distance between the input word (X = leaves Y = am.) And the example word (X = arrive Y = pm.) Is determined by setting the distance between leaves and arrive regarding the variable X to 0.
33, and a. m. And p. m. Assuming that the distance between the input and the variable is 0.00, the distance between the input and this example is 0.33 × 0.5 + 0.
00 × 0.5 = 0.17 can be calculated. Here, the distance between the input word and all the example words of the language pattern is calculated, and the minimum distance is set as the distance between the input word and the example word.

The maximum likelihood structure obtained as a result of the analysis by the structural analysis processing unit 2 is used as an input analysis tree in the subsequent processing. The parse tree is attached to the word sequence of the input sentence, with the language pattern matched at the time of the structural analysis, its attribute value, and the calculated semantic distance attached. The dependency by the language pattern is set as a node, and the word ends. It has a tree structure for symbols.

The input parse tree restoration judging section 3 judges whether or not to apply the analytic tree restoration process to the input parse tree data obtained from the structural analysis processing section 2. Specifically, a predetermined threshold value is set in advance for the sum of the semantic distances of the input parse tree, and the sum of the semantic distances is larger than this threshold value (that is, the sum is semantically distant). The analytic tree restoration process is determined to be necessary for the input parse tree, and the analytic tree restoration process is determined to be unnecessary for the input parse tree whose sum of semantic distances is small (ie, semantically close). Here, when it is determined that the parse tree restoration process is necessary, the input parse tree is output to the parse tree restoration processor 4 and the input subtree and example subtree collation processor 7, and when it is determined that the parse tree is unnecessary, the speech recognition error is determined. The entire processing of the correction device is completed, and the input parse tree is output as it is as a final output.

On the other hand, the example sentence data memory 21 stores, as an example sentence set, a natural language sentence set (text data) of a predetermined language that is the same as the language of the input sentence and does not include a speech recognition error. The structure analysis processing unit 10 has the same configuration as that of the structure analysis processing unit 2 in FIG. 2, and performs a structure for each example sentence in the example sentence data in the example sentence data memory 21 by the structure analysis processing in FIG. The analysis processing is performed, and the maximum likelihood structure obtained as a result of the analysis is used as an example analysis tree in the subsequent processing. This example parse tree, like the input parse tree, attaches to the word string of the input sentence the language pattern matched at the time of structural analysis, its attribute value, and the calculated semantic distance, and It has a tree structure with dependency as nodes and words as terminal symbols. That is, the example analysis tree is output from the structural analysis processing unit 10 and stored in the example analysis tree data memory 22.

As described above, the input parse tree is input to the input subtree and example subtree collation processing unit 7 when the input parse tree restoration determination unit 3 determines that the parse tree restoration process is necessary. Here, the parse tree is attached to the word string of the sentence, with the language pattern collated at the time of the structural analysis, its attribute value, and the calculated semantic distance. Has a terminal structure of a tree structure. The analytic tree includes a substructure in which each node in the analytic tree is a vertex and has a terminal symbol in a lower structure (base), that is, a closed subtree. Therefore, if the parse tree has N nodes including its own vertices, the number of closed subtrees included in the parse tree is also N. Here, the partial tree includes the entire parse tree itself. The input subtree and example subtree matching processing unit 7 fetches, from the input parse tree and the example parse tree in the example parse tree data memory 22, matching subtrees defined in this way, and performs input analysis. All possible combinations of the tree subtree and the example analysis tree subtree are generated and output to the subtree phoneme sequence conversion processing unit 8 as pairs of the input subtree and the example subtree. Here, in order to extract a matching subtree, for example, a subtree whose phonological distance is equal to or smaller than a predetermined threshold is searched for and extracted.

A word phoneme conversion dictionary memory 23 is connected to the partial tree phoneme string conversion processing section 8, and the word phoneme conversion dictionary memory 23 stores a plurality of data of those pairs for converting a word string into a phoneme string. Is done. The subtree phoneme string conversion processing unit 8 converts the word phoneme conversion dictionary in the word phoneme conversion dictionary memory 23 from the pair of the input subtree and the example subtree output from the input subtree and the example subtree matching processing unit 7. By reference, a corresponding pair of phoneme strings is generated. Here, the phoneme sequence for each subtree is obtained by extracting the word sequence of the terminal symbol from the subtree,
It is obtained by converting a word string into a phoneme string using the word phoneme conversion dictionary memory 23. According to this method, a phoneme sequence for each combination of subtrees of an input parse tree and an example analytic tree forming a subtree pair is obtained, and the pair is phonologically defined as a pair of a phoneme sequence of the input subtree and the example subtree. Output to the distance calculation processing unit 9. In the present embodiment, a subtree of the input parse tree is called an input subtree, and a subtree of the example analytic tree is called an example subtree.

The phonological distance calculation processing unit 9 compares a pair of the input subtree and the example subtree obtained by the subtree phoneme string conversion processing unit 8 with the input part of the phoneme sequence of the example subtree. A phonological distance to a phoneme sequence of the tree is calculated, and a pair of an input subtree and an example subtree whose phonological distance is equal to or less than a predetermined threshold value (that is, phonologically close or similar), and the phonology thereof. The target distance is output to the phonologically similar subtree buffer memory 24. Here, in the phonologically similar subtree buffer memory 24, the phonological distance is given to a pair of the input subtree and the example subtree obtained by the phonological distance calculation processing unit 9 and stored.

As described above, the analytic tree restoration processing unit 4
When the input parse tree restoration determining unit 3 determines that the parse tree restoration process is necessary, the input parse tree is input. The parse tree restoration processing unit 4 restores the input parse tree for each pair of the input subtree and the example subtree in the phonologically similar subtree buffer memory 24. Here, first, the input subtree is compared with the input parse tree, the node of the input parse tree that matches the vertex node of the subtree is identified, and then all substructures of this node are replaced with the example subtree. As a result, a restoration analysis tree for the pair of the input subtree and the example subtree is obtained. If there is no pair of the input subtree and the example subtree in the phonologically similar subtree buffer memory 24 for the input sentence, the input parse tree repair determination unit 3 determines that the parse tree repair processing is unnecessary. As in the case where it is determined, the entire process of the speech recognition error correction device is terminated, and the input parse tree is output as it is as a final output.

The semantic distance calculation processing unit 5 recalculates the semantic distance for each repair parse tree obtained from the analytic tree repair processing unit 4. Here, for the repair analytic tree whose semantic distance has been recalculated, only the repair analytic tree whose sum of the semantic distances is equal to or less than a predetermined threshold is output as a correction analytic tree. Here, as the threshold value, the same value as the threshold value used for determining the necessity of the parse tree restoration process in the input parse tree restoration determination unit 3 is used. Further, the correction parse tree output unit 6
Among the correction analysis trees obtained from the semantic distance calculation processing unit 5, the correction analysis tree having the minimum phonological distance is output as an output analysis tree together with the speech recognition result.

FIG. 3 is a flowchart showing the overall processing flow of the speech recognition error correction processing in FIG.

In FIG. 3, first, in step S1, speech recognition processing of a speech input sentence is performed, and then in step S2, the structure of the speech recognition result is analyzed, and an input analysis tree is output. Next, in step S3, the value of the semantic distance of the input parse tree is compared with a threshold value. In step S4, it is determined whether or not the semantic distance is a value larger than the threshold value. Proceeding to S11, the parse tree of the input sentence is output. On the other hand, if YES, the matching between the input subtree and the example subtree and the calculation of the phonological distance are performed in step S5, and the phonological distance having a distance equal to or smaller than the threshold value is calculated. A similar subtree is output, and it is determined in step S6 whether a phonologically similar subtree candidate has been obtained. If NO, the process proceeds to step S11, while the YE
In the case of S, the process proceeds to step S7.

In step S7, the phonologically similar subtree of the input parse tree is replaced with an example subtree, and the input parse tree is restored. Next, the semantic distance of the restored input analytic tree is recalculated, and only the analytic tree having the semantic distance equal to or less than the threshold is output as the corrected analytic tree. It is determined in step S9 whether a candidate for a correction analysis tree has been obtained. If NO, the process proceeds to step S11. If YES, the process proceeds to step S11.
In step 10, among the candidates for the correction analysis tree, the analysis tree having the minimum phonological distance is output, and the speech recognition error correction processing ends.

Next, the structure analysis processing and the semantic distance will be described. In the present embodiment, the structural analysis processing units 2 and 10
Uses the above-mentioned conventional CB parser. C
The B parser applies a dependency relationship to an input sentence and creates a parse tree from the bottom up. The semantic distance of the parse tree is calculated from the semantic closeness between the example parse tree and the input parse tree. The smaller the value of the semantic distance, the higher the reliability of the validity of the dependency. A common threshold value for the semantic distance is set in the correction necessity determination by the input parse tree repair determination unit 3 and the structural recovery determination by the correction by the semantic distance calculation processing unit 5 in FIG. The necessity of input correction and the validity of correction are determined based on the information.

Further, the search processing and the phonological distance will be described. In the structure search processing by the input subtree and example subtree matching processing unit 7, a subtree having a morphological sequence phonetically close to each subtree of the input parse tree is searched. Then, the subtree phoneme sequence conversion processing unit 8 converts the partial structure of the input analytic tree and the morphological sequence of the subtree of the example in the database into phoneme sequences, respectively. The edit distance of the phoneme sequence with respect to the input phoneme sequence is calculated as a phonological distance, and in the subsequent search processing, a threshold value is set for the phonological distance, and a subtree having a smaller phonological distance is searched. Here, as is well known, the edit distance refers to the sum of the number of processing operations when deleting, inserting, or replacing phonemes in order to obtain an example phoneme sequence for an input phoneme sequence. In the output of the correction parse tree output unit 6, a correction candidate having the minimum phonological distance is output as a correction result.

In the above embodiment, the speech recognition processing unit 1, the structure analysis processing unit 2, and the input parse tree restoration determination unit 3
Analytic tree restoration processing unit 4 and semantic distance calculation processing unit 5
A correction parse tree output unit 6, an input subtree and example subtree collation processing unit 7, a subtree phoneme sequence conversion processing unit 8, a phonological distance calculation processing unit 9, and a structural analysis processing unit 10. The example analysis tree data memory 21, the example analysis tree data memory 22, the word phoneme conversion dictionary memory 23, and the phonologically similar subtree buffer memory 24 are configured by a central processing unit such as a digital computer. For example, it is configured by a storage device such as a hard disk memory.

In the above embodiment, the morphological analysis dictionary memory 111 includes synonym code information. However, the present invention is not limited to this. It may be configured to store synonym information describing a relationship between a word and a synonym code.

In the above embodiment, processing relating to Japanese and English is illustrated, but the present invention is not limited to these languages, and can be applied to various other natural languages.

[0049]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, an example of processing of each section of the speech recognition error correction apparatus according to the present embodiment will be described with reference to FIG. First, after the uttered voice of a natural language sentence is converted into a voice signal by the microphone 30, the voice signal is input to the voice recognition processing unit 1, and the voice recognition processing unit 1 converts the input voice signal of the natural language sentence into an appearance form of a word. Recognize voice. Here, the appearance form of a word is a form of a word that is obtained when a certain natural language sentence is divided into words and that includes an inflection and a sway in notation due to the use of words. On the other hand, the structural analysis processing unit 2 once converts the input sentence from the appearance form to a standard form that does not include the inflection or the sway of the notation due to the use, and then compares the language pattern. Table 8 shows a specific example of the output of the speech recognition processing unit 1.

[0050]

[Table 8] ―――――――――――――――――――――――――――――――― ORDER = 1 ――――――――― ――――――――――――――――――――――――― WORDS = UTT-START / That ’s right / Six + hours / About / Arrive / Available / I think + Masu / UTT-END ―――――――――――――――――――――――――――――――― wordsids = 5/20061/10046 + 10203/10344 / 10 104/10053/10589/10009 + 10090 + 10011/6 ―――――――――――――――――――――――――――――――― vars = 1/1 / 1/1/1/1/1/1/1 ―――――――――――――――――――――――――――――――――― ivs =-/ s + o + o + d + e + s + u + n + e + e +-/ r + o + k + u + zh + i / g + u + r + a + i / g + a +-/ t + o + o + ch + j + a + k + u / d + e + k + i + r + u + t--o + o + m ―――――――― times = 0.000000 / 0.190000 / 2.2800000 / 2.700000 / 3.0800000 / 3.5600000 / 3.910000 / 4.2.200000 / 5.2100000 ――― ――――――――――――――――――――――――――――――― score = -11881822.000000 ―――――――――――――― ――――――――――――――――――――― acoustic =- 65761822.0000000 ―――――――――――――――――――――――――――――― ngram = −533000.000000 ―――――――― ―――――――――――――――――――――――――――

Next, the syntax analysis processing unit 2 outputs the analysis result of the language dependent structure to the output of the speech recognition processing unit 1 as parse tree data using the above-mentioned CB parser. Table 9 shows the output result of the syntax analysis processing unit 2.

[0052]

[Table 9] ―――――――――――――――――――――――――――――――――― UTTERANCE = 62 ――――――――― ――――――――――――――――――――――――― (: S-SENT 62 ”If you are in a Japanese room, the service fee of 188,000 yen will be different ： ”： TAGGED-LIST (“ Japanese-style room ｜ Washitsu ｜ Japanese-style room ｜ Normal noun ｜｜ ”” ｜ No ｜｜ Conjunctive particle ｜｜ ”” Ho ｜ Ho | | Judgment | special | basic "" and | g | and | connected particle || "" one | ichi | one | numeral || "" million | man | million | numeral || "" 8 | ”| Thousand | sen | thousand | number | |“ ”circle | en | yen | common noun | |“ ”any | islet | any | pronoun | | "War | ｜ war ｜ ordinary noun ｜｜ ”” service charge ｜ service fee ｜ service charge ｜ ordinary noun ｜ ｜ ”｜ ha ｜ is ｜ adjunctive ｜｜” ”another | betsu | another | adjective | Ni | ni | case particle | | "” nar | nari | nar | main verb | five-stage la | continuous use ”mas | mass | mas | auxiliary verb | special sa | basic”)) ――――――――― ――――――――――――――――――――――――― (: S-TREE 62 (((0 16) (: PATTERN MARKER R-CN-PRON-SS: FORM) (? X <CN-PRON>? Y): HEAD ((: WORD "different": REG-EXP "different": POS adjective noun: ATTRIBUTE NIL: SEM-CODE (415 182 101): COMPOUND NIL) (: WORD "Becomes": REG-EX "Become": POS auxiliary verb: ATTRIBUTE ((: CONJ-FORM. Continuous use)): SEM-CODE NIL: COMPOUND NIL): DISTANCE 0.0: TOTAL-D ISTANCE 1.30000233: SUBSTRACT (SSUBSTRACTX) : PATTERN CONJ-DESUTO-NP: FORM (? X "is" and "? Y): HEAD ((: WORD" 18000 yen ": REG-EXP" N yen ": POS Common noun: ATTRIBUTE NIL: SEM-CODE (828 121): COMPOUND ((: WORD "18000": REG-EXP "00": POS Numeral: ATTRIBUTE NIL: SEM-CODE (120): COM POUND NIL) (: W RD "circle": REG-EXP "circle": POS common noun: ATTRIBUTE NIL: SEM-CODE NI L: COMPOUND NIL)))): DISTANCE 0.0: T OTAL-DISTANCE 0.0: SUBSTRUCTURES ((? X (: PATTERN SETSUBI-NOHOU-KANJI: FORM ("" "of? X")): HEAD ((: WORD "Japanese-style room": REG-TENIL: SEM-CODE (941): COMPOUND NIL)): DISTANCE 0.0: TOTAL-DISTANCE 0.0: SUBSTR ―――――――――――――――――――――――――――――――――― > ――――――――――――――――――――――――――――――――――

The input parse tree data is output to the input parse tree restoration determining unit 3. Here, if the value of the semantic distance of the input parse tree is smaller than the threshold, the subsequent correction processing is interrupted, the input parse tree is output as it is, and the processing is terminated. If the value of the semantic distance is larger than the threshold value, the input subtree / example subtree matching processing unit 7 performs collation on the input analysis tree and the example analysis tree subtree.

In the example of Table 9, the value of the semantic distance is 1.30.
Since it is 00233, if the threshold is set to 0.3 here, it means that subtree matching is performed. Here, the example parse tree data memory 22 stores the example sentence data memory 21 in advance.
It is generated from the example sentence inside using the syntax analysis processing unit 10. Table 10 shows data examples in the example sentence data memory 21.

[0055]

[Table 10] ―――――――――――――――――――――――――――――――― 140 ｜ 0150 ｜ 630 ｜ 1790 ｜ Which | Both | Pronouns | | | | ―――――――――――――――――――――――――――――――――― 140 | 0150 | 630 | 1800 | | Mo | mo | particles |||| ―――――――――――――――――――――――――――――――――― 140 ｜ 0150 ｜ 640 | 1810 | Tax | Zee | Tax | Normal noun |||| ―――――――――――――――――――――――――――――――――― 140 ｜ 0150 ｜ 650 ｜ 1820 ｜ Service charge ｜ Service charge ｜ Service charge ｜ Common noun ｜｜｜｜ ―――――――――――――――――――――――――――― ―――――― 140 ｜ 015 | 650 | 1830 | is | c | is | particle || | | ----------------------------------------------------------------------------------------- ― 140 ｜ 0150 ｜ 660 ｜ 1840 ｜ Separate | Bet | Separate | Adverb ｜｜｜｜ ―――――――――――――――――――――――――――――― ―――― 140 ｜ 0150 ｜ 670 ｜ 1850 ｜ Give me | Butterfly | Give me | Sa noun ｜｜｜｜ ―――――――――――――――――――――――――― ―――――――― 140 ｜ 0150 ｜ 670 ｜ 1860 ｜ Ishii ｜ Itashi ｜ Auxiliary verb ｜ Five-dan ｜ Continuous ｜｜ ―――――――――――――――――― ―――――――――――――――― 140 ｜ 0150 ｜ 670 ｜ 1870 ｜ Mas ｜ Mas ｜ Mas ｜ Auxiliary verb ｜ Special ｜ Basic ｜｜ ――――――――― ------------------------ 140 | 0150 | 670 | 1880 |. || │symbol ｜｜｜｜ ――――――――――――――――――――――――――――――――――

The data in the example parse tree data memory 22 uses exactly the same structure analysis processing as the syntax analysis processing unit 2 used for the dependency structure analysis of the output of the speech recognition processing unit 1. The obtained parse tree data has the same format as in Table 9. Table 11 shows an example of data in the example analysis tree data memory 22.

[0057]

[Table 11] ―――――――――――――――――――――――――――――――― （： S-SENT 1 ”Thank you New York Cityホ テ ル で ホ テ ル ―――――――――――――――――――――――――――――――――― (: S-TREE 1 (((0 4) (: PATTERN MARKER- INTERJ-IS: FORM (? X <INTERJ->? Y): H EAD ((: WORD "New York City Hotel": REG-EXP "New York City Hotel": POS Common noun: ATTRIBUTENIL : SEM-CODE (729): COMPOUND NIL)): DISTANCE 1.0e-5: TOTAL-DISTANCE 1.0e-5: SUBSTRUCTURES (? X (: PATTERN: STRING: FORM ((: WORD "Thank you"): REG-EXP "Thank you": POS Inspirational verb: ATTRI BUTE NIL: SEM-CODE (696 476): COMPU ND NIL ): HEAD (: WORD "Thank you"): REG-EXP "Thank you": POS Inflection: ATTRI BUTE NIL: SEM-CODE (696 476): COMPOUN ND NIL): DISTANCE 0.0: TOTAL-DISTANCE E 0.0: SUBSTRUCTURES NIL)) (? Y (: PAT TERN DEGOZAIMASU: FORM (? X "is")): HEAD ((: WORD "Newyo City Hotel ": REG-EXP" New York City Hotel ": POS Common Noun: ATTRI BUTE NIL: SEM-CODE (729): COMPOUND N IL): DISTANCE 0.0: TOTAL-DISTANCE 0 0.0: SUBSTRUT? X (: PATTERN: ST RING: FORM ((: WORD "New York City Hotel": REG-EXP "New York City Hotel"): POS Common Noun: AT TRIBUTE NIL: SEM-CODE (729): COMPOUND DNIL): HEAD (: WORD “New York City Hotel”: REG-EXP “New York City Hotel”: POS Common Noun: ATTRIBUTE NIL: SEM- ODE (729): COMP OUND NIL): DISTANCE 0.0: TOTAL-DISTANCE NCE 0.0: SUBSTRUCTURES NIL)))))))))))))) ―――――――――――――――――――

FIGS. 4 and 5 are diagrams showing an example of the processing results of the embodiment in the speech recognition error correction device of FIG. Here, the parse tree is obtained by the structure analysis processing units 2 and 10 as a structure including a combination of a language pattern and a surface phrase for an input sentence. In the parse trees in FIGS. 4 and 5, a solid line indicates a reference to a variable that does not mean a specific linguistic unit, and a dotted line indicates a reference to a linguistic component boundary. For example, as for “_________________________” of the input analysis tree in FIG. These indicate that the structure is made by combining the variable words X and Y of the pattern “X Y” with the surface words “Japanese room” and “kata”, respectively. Further, "" connected to the node by a dotted line indicates that the language component boundary of the pattern is "".

The input parse tree in FIG. 4 is the same as in Table 9;
This is conceptually represented for explanation. For example, when the threshold of the semantic distance is 0.3 and the threshold of the phonological distance is 0.3, the value of the semantic distance of the input parse tree is 1.30.
Since it is 0, the input parse tree repair determination unit determines that a repair process is necessary. Here, the example sentence data memory 2
Assuming that the example sentence “Every one is different from the tax service charge” is included, the example parse tree data memory 22 stores “any_tax_service_charge becomes different_” Is included. Here, in the input subtree and the example subtree matching processing unit, as an input subtree and an example subtree, a subtree having an apex of a node indicated by a black circle (input subtree) as an input analysis tree and an example analysis tree, respectively. Eventually _ war _ service fee _ will be _ another _ _
And the example subtrees “all_tax_service_charges_are_different_”), the phonological distance between the subtrees is set to 0, as shown in FIG. 078
Value. This value is a threshold value for the phonological distance of 0.
Since the value is smaller than 3, the pair of the input subtree and the example subtree collated here is stored in the phonologically similar subtree buffer memory 24. Further, the analytic tree restoration processing unit 4 replaces the subtree (substructure) having the node indicated by the black circle of the input analytic tree at the apex with an example subtree to obtain an analytic tree that has been repaired. The value of the semantic distance of the analytic tree is recalculated in the semantic distance calculation processing unit 5 to obtain a value of 0.0.
The value of the semantic distance of the modified parse tree is equal to the threshold value of 0.
Since it is smaller than 3, it is output to the correction analysis output unit 6. Here, if the phonological distance is the smallest of all candidates,
The final output. Table 12 shows the phonological distance calculation processing unit 9
Is shown (when an example different from the examples of FIGS. 4 and 5 is used).

[0060]

[Table 12] ―――――――――――――――――――――――――――――――― Input partial tree = (((((“Japanese-style room”)) "" "" "" "" "" "" "" "" "" "" "" "" ("" ")" <">" "(" "") Fee “)))” becomes “((“ different ”)” ”))” masu ”)) ――――――――――――――――――――――――― ――――――――― Example partial tree = ((((“Japanese room”) ”or“ <SUBP-> ((“Twin room”) ”” of “(((“ one person ”)“ like ”)) <PERSONPNOM-SN> ("Use")))) "" "" "" "" ("" "" "" "" "" "" "" "" "" ")" --- " --------------------- for Sentence = (: S-SENT 1900 "Can be taken by a single person in a Japanese-style room or a twin room.") ――――――――――――――――――――――― ――――――――――― Input subtree word string = Japanese-style room,,,, 18,000 yen,,, battle, service charge, is different,, ―――――――――――――――――――――――――――――――― Example subtree word sequence = Japanese-style room, or twin room, one person, Sama, use, desu, if you can, take it, you can ―――――――――――――――――――――――――――――――――― Input subtree phoneme sequence = washitsunokasadastuto18000english wabsununi-narumasusu ―――――――――――――――――――――――――――――――――― -Example partial tree phoneme sequence = wasitsukatsutsuuingruumuooohonorisamasashiodaisitaitai o-toridoekirumamasu ―――――――――――――――――――――――――――――――― Phonological distance = 0.65551724137931 ――――――――――――――――――――――――――――――――――

Table 13 shows an example of the output of the semantic distance calculation processing section 5. This parse tree is obtained by replacing the structure of the input subtree with the example subtree in the analytic tree restoration processing unit 4 and recalculating the semantic distance in the semantic distance calculation processing unit 5. is there. Here, as a result of the calculation, the value of the semantic distance is 0.0.

[0062]

[Table 13] ―――――――――――――――――――――――――――――――― (： S-SENT 62 ” The service fee for every 188,000 yen will be different. ”) ―――――――――――――――――――――――――――――――――― (: S-TREE 62 (((0 16) (: PATTERN MASU: FORM (? X "Masu")): HEAD ((: WORD "Otori": REG-EXP "Take": POS Main verb: ATTRIBUTE (( : CONJ-FORM. Continuous use)): SEM-CODE (413 390 383 379 373 372 305 225): COMPOUND ((: WORD "O": REG-EXP "O": POS Prefix: ATTRIBUN NEL DENIL) : COMPOUND D NIL) (: WORD "take": REG-EXP "take"): POS Main verb: ATTRIBUTE ((:: CONJ-FORM. Continuous use)): SEM-CODE NIL: COMPOUND NIL))): DISTAN 0.0: TOTAL-DISTANCE 0.0: SUBS TRUCTURES ((? X (: PATTERN CONJ-DESHIT ARA-NP: FORM (? X ")"? ""? Y): HEAD ((: WORD " ": REG-EXP" "take": POS main verb: ATTRIBUTE ((: CONJ-FORM. Continuous use)): SEM-CODE (413 390 383 379 373 373 372 305 225): COMPOUND ((: WORD "O" REG-E XP "O": POS Prefix: ATTRIBUTE NIL: SEM-CODE NIL: COMPOUND NIL) (: WORD "take": RE G-EXP "take": POS Main verb: ATTRIBUTE ((: CONJ-FOR )): SEM-CODE NIL: COMPO UND NIL)))): DISTANCE 0.0: TOTAL-DIS TANCE 0.0: SUBSTRUCTURES ((? X (: PATT ERN KA-SUBP-S + N:? X ?? "<SUBP->? Y): HEAD ((: WORD "Use": REG-EXP "Use"): POS suffix noun: ATTRIBUTE NIL: SEM-CODE (381 363): COMPOUND NIL): DISTANCE E0.0: TOTAL-DI 0.0: SUBSTRUCT URES ((? X (: PATTERN: STRING: FORM ((: WORD "Japanese-style room"): REG-EXP "Japanese-style room": POS Common noun: ATTRIBUTE NIL: SEM-CODE (941) UPON ): HEAD (: WORD "Japanese-style room": REG-EXP "Japanese-style room": POS Common noun: ATTRIBUTE NIL: SEM-CODE (941): COMPOUND NIL): DISTANCE 0.0: TO AL-DISTANCE 0.0: SUBSTRUCTURE ES NIL)) (? Y (: PATTERN KAKUJO-NO-N + N: FORM (? X's ?? Y)): HEAD ((: WORD "Use": REG-EXP "Use") : POS service ―――――――――――――――――――――――――――――――― ―――――――――――――――――――――――――――

Table 14 shows an example of data in the word phoneme conversion dictionary memory 23. The phoneme sequence for the word expression on the left is
It is described on the right.

[0064]

[Table 14] ―――――――――――――――――――――――――――――――― 10340 [And] so shiite @ | −｝ # And | Society | And | Connective |||| ―――――――――――――――――――――――――――――――― 10341 [ Both days] rjoozhitsu ｛|-｝ # Both days | Ryojitsu | Both days | Ordinary noun ||| | ――――――――――――――――――――――― ――――――――――― 10342 [Tomo] tomo ｛｜ −｝ # Both | Tomo | Both | Common nouns |｜｜｜ ―――――――――――――――― ―――――――――――――――――――― 10343 [Please] tekudasaq # please | techdasasa | please | auxiliary verb | special ra | ｜ ―――― ―――――――――――――――――――――――――――――― 10344 [about] g urai |-｛# about | gray | about | Adjunct |||| ―――――――――――――――――――――――――――――――― 10345 [Written] m a r i ｛ ｜ −｝ # worship ｜ miri ｜ worship ｜ main verb ｜ five steps ｜ continuous use ｜｜ ――――――――――――――――――――――――――――― ――――― 10346 [Kyoto Kanko Hotel] kjo ootka kang kohohoteru |-｝ # Kyoto Kanko Hotel | Kyoto Kanko Hotel | Kyoto Kanko Hotel | Kyoto Noun | | │ ｜ ―――――――――――――――――――――――――――――――― 10347 [Front] hur ng to o ｛|- ｝ # Flow To | front | front | common noun | | | | ―――――――――――――――――――――――――――――――――― 10348 [Sunday] nichijoobobi ｛|-｝ # Sunday | Nichijobi | Sunday | Ordinary noun | | | | <partial> ―――――――――――――――――――― ――――――――――――――

<Evaluation Experiment> The present inventor conducted an evaluation experiment as follows using the speech recognition error correction apparatus of FIG. 1 configured as described above. Travel utterance data 310 utterances (Japanese) learned by the structural analysis were input. The example parse tree database in the example parse tree 22 was constructed from error-free 14,111 utterances including the input 310 utterances. Table 1
As shown in FIG. 5, in the correction necessity determination process, 147
The utterance was determined to be corrected.

[0066]

[Table 15] Judgment necessary for correction (310 utterances) ―――――――――――――――――――――――――――――――― Correction judgment Error utterance Correct utterance ―――――――――――――――――――――――――――――――― Required 147 (γ _sem ≤ Γ _sem ) 141 6 _Not required 163 ( γ _sem > Γ _sem ) 71 92 ――――――――――――――――――――――――――――――――――

Here, γ _sem is the semantic distance to the input parse tree, and Γ _sem is its threshold. here,
0.3 was used as the threshold Γ _sem of the semantic distance. The evaluation criterion shown in Table 16 was used to evaluate the semantic recovery of the correction result in units of speech.

[0068]

[Table 16] Evaluation criteria ―――――――――――――――――――――――――――――――― Level 5: Completely agree with the correct answer (complete correct answer) ). Level 4: Semantically almost equivalent to the correct answer. Level 3: Partial recovery, slightly different from the correct answer. Level 2: no recovery, improper correction of error part. Level 1: improper correction of the correct recognition part. ――――――――――――――――――――――――――――――――――

Next, Table 17 shows the evaluation results for the correction target.

[0070]

[Table 17] Semantic recovery by correction (147 utterances) ―――――――――――――――――――――――――――――――― No distinction Only perfect answer Other than perfect answer ―――――――――――――――――――――――――――――――― Level5 43% (63) 33% (49) 14% (20) Level4 5% (8) 0% (0) 6% (9) Level3 3% (4) 1% (1) 4% (6) Level2 1% (2) 0% (0) 2% (3) Level1 0% (0) 0% (0) 0% (0) ―――――――――――――――――――――――――――― ―――――― NIL 48% (70) 66% (97) 74% (109) ―――――――――――――――――――――――――――― ――――――

Since the input correct sentence (complete correct answer) is included in the database in the example analysis tree data memory 22, the correction is made by all the examples including the correct answer (no distinction), and the correction is made only by the correct answer ( Only the correct answer) and corrections other than the correct sentence (other than the correct answer) were evaluated separately. here,
The threshold for the phonological distance was 0.3. In addition, a correction candidate having the minimum phonological distance for each utterance was determined as a correction result.

As described above, the evaluation experiments show that the method is a high-precision correction method with few erroneous corrections of the input, and effective correction is possible from examples other than the correct sentence of the input. It became clear.

<Effects of Embodiment> As described above,
According to the present embodiment, an example in which the input sentence and the phonological feature are close to each other is selected, so that erroneous correction is reduced. The correction result is
Since the validity is determined and output based on the semantic distance, the erroneous correction is reduced because the validity is guaranteed. Here, since the validity of the entire input sentence is determined based on the semantic distance, the validity of the global speech recognition result can be determined. Therefore, a more accurate error correction process can be performed as compared with the conventional example.

[0074]

As described above in detail, according to the speech recognition error correction apparatus of the first aspect of the present invention, the speech signal of the uttered speech sentence is recognized, and the speech recognition result is subjected to error correction processing. A speech recognition error correction device for performing a predetermined structure analysis process on a predetermined example sentence, a storage unit (22) for storing an example parse tree, and a predetermined structure for the voice recognition result. Structural analysis processing means (2) for performing analysis processing to generate and output an input analytic tree, input analytic tree output from the structural analytic processing means (2), and stored in the storage means (22) Based on the example parse tree, a subtree pair that is phonologically close is searched between each input subtree that is a subtree of the input parse tree and each example subtree that is a subtree of the example analytic tree. , Interchangeable structure between input parse tree and example parse tree Error correction processing means (3, 7, 8, 9, 4) for performing error correction processing based on the error and outputting an error correction result of the input parse tree, and error correction processing means (3, 7, 8, 9, 9, 4) ), The overall validity of the entire sentence is calculated by calculating the semantic distance calculated depending on the positional relationship in the concept hierarchy of synonyms between words for the entire uttered speech sentence including the error correction result output from The semantic distance calculation processing and output means (5, 6) for outputting a parse tree of an uttered speech sentence including an error correction result when the calculated semantic distance is equal to or less than a predetermined threshold value. Prepare. Therefore, according to the present invention, since an example whose input sentence is close to the phonological feature is selected, erroneous correction is reduced. The correction result is
Since the validity is determined and output based on the semantic distance, the erroneous correction is reduced because the validity is guaranteed. Here, since the validity of the entire input sentence is determined based on the semantic distance, the validity of the global speech recognition result can be determined. Therefore, it is possible to perform a more accurate error correction process than in the conventional example.

According to the second aspect of the present invention, in the first aspect of the present invention, the error correction processing means (3,7,8,9,4).
Calculates the semantic distance for the input parse tree output from the structural analysis processing means (2), and restores the input parse tree when the calculated semantic distance exceeds the threshold. Input parse tree restoration determination means (3)
Based on the input parse tree determined to be repaired by the input parse tree repair determination means (3) and the example parse tree stored in the storage means, And an input subtree and example subtree matching processing unit (7) for generating and outputting pairs of combinations of subtrees that are phonologically close to each other. A subtree phoneme that converts a pair of a combination of an input subtree and an example subtree output from the example subtree matching processing unit (7) into a phoneme sequence with reference to a predetermined word phoneme conversion dictionary and outputs the phoneme sequence Calculating a phoneme distance between a phoneme sequence of the input subtree and a phoneme sequence of the example subtree output from the sequence conversion processing unit (8) and the subtree phoneme sequence conversion processing unit (8); An input subtree whose calculated phonological distance is less than or equal to another predetermined threshold A phonological distance calculation processing unit (9) for outputting a pair with the example subtree, and the phonological distance calculation for the input parse tree determined to be repaired by the input parse tree repair determination means (3). An analysis tree that outputs an error correction result by performing an error correction process by repairing an input analysis tree using a pair of an input subtree and an example subtree output from the processing unit (9). A restoration processing unit (4). Therefore, according to the present invention, since an example whose input sentence is close to the phonological feature is selected, erroneous correction is reduced. Also,
Correction results are output based on the determination of validity based on the semantic distance. Therefore, since the validity is guaranteed, erroneous corrections are reduced. Here, since the validity of the entire input sentence is determined based on the semantic distance, the validity of the global speech recognition result can be determined.
Therefore, it is possible to perform a more accurate error correction process than in the conventional example.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a configuration of a speech recognition error correction device according to an embodiment of the present invention.

FIG. 2 is a block diagram showing a configuration of structural analysis processing units 2 and 13 of FIG.

FIG. 3 is a flowchart showing the flow of the entire speech recognition error correction process in FIG. 1;

FIGS. 4A and 4B are diagrams showing a first part of a processing result of the embodiment in the speech recognition error correction device of FIG. 1, wherein FIG. 4A shows a speech recognition result, FIG. (C)
Shows an input subtree, (d) shows an example sentence, (e) shows an example parse tree, and (f) shows an example subtree.

5 is a diagram showing a second part of the processing result of the embodiment in the speech recognition error correction device of FIG. 1, (g) shows the calculation of the phonological distance between the subtrees, and (h) shows A morpheme sequence for the input subtree and its phoneme sequence are shown, (i) shows a morpheme sequence for the example subtree and its phoneme sequence, and (j) is a repaired input parse tree for which the recalculation of the semantic distance has been performed. 3 shows a parse tree.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 ... Voice recognition processing part, 2 ... Structural analysis processing part, 3 ... Input analytic tree restoration determination part, 4 ... Analysis tree restoration processing part, 5 ... Semantic distance calculation processing part, 6 ... Correction analytic tree output part, 7 ... Input subtree and example subtree matching processing unit, 8: partial tree phoneme sequence conversion processing unit, 9: phonological distance calculation processing unit, 10: structural analysis processing unit, 21: example sentence data memory, 22: example analysis tree data Memory 23: Word phoneme conversion dictionary memory 24: Phonemic similar partial tree buffer memory 30: Microphone 101: Morphological analysis unit 102: Part of speech bigram insertion unit 103: Language pattern search unit 104: Language pattern matching unit Reference numeral 105: Maximum likelihood structure determination unit 111: Morphological analysis dictionary memory 112: Part-of-speech bigram dictionary memory 113: Language pattern memory

────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Jin Iida Kyoto, Soraku-gun, Seika-cho, 5th, Inaniya, 5th, Sanpira-ya ATI Ron Co., Ltd. Voice Translation and Communication Research Laboratories (56) References JP-A-2-2 214990 (JP, A) JP-A-6-274192 (JP, A) JP-A-10-232693 (JP, A) (58) Fields investigated (Int. Cl. ⁶ , DB name) G10L 3 / 00-9 / 20 G06F 15/38 JICST file (JOIS)

Claims (2)

(57) [Claims] 1. A speech recognition error correction device for recognizing a speech signal of an uttered speech sentence and performing an error correction process on the speech recognition result, wherein a predetermined structural analysis process is performed on a predetermined example sentence. Storage means (22) for storing the obtained example analysis tree; and structural analysis processing means (2) for performing a predetermined structural analysis process on the speech recognition result to generate and output an input analysis tree.
Based on the input parse tree output from the structural analysis processing means (2) and the example parse tree stored in the storage means (22), ,
A subtree pair that is phonologically close to each example subtree, which is a subtree of the example analysis tree, is searched, and error correction processing is performed based on a compatible structure between the input analysis tree and the example analysis tree. Error correction processing means (3, 7, 8, 9, 4) for performing error correction and output of error correction results of the input parse tree, and output from the error correction processing means (3, 7, 8, 9, 4) For the entire vocal sentence including the error correction result, the overall validity of the entire sentence is determined by calculating the semantic distance calculated depending on the positional relationship in the concept hierarchy of synonyms between words, A semantic distance calculation process and an output means for outputting a parse tree of an uttered speech sentence including an error correction result when the calculated semantic distance is equal to or smaller than a predetermined threshold value;
6) A speech recognition error correction device characterized by comprising: 2. The speech recognition error correction device according to claim 1, wherein said error correction processing means (3, 7, 8, 9, 4) includes an input analysis tree output from said structure analysis processing means (2). Input parse tree restoration determining means (3) for calculating a semantic distance for the input parse tree and determining that the input parse tree should be repaired when the calculated semantic distance exceeds the threshold value; Based on the input parse tree determined to be repaired by the input parse tree repair determining means (3) and the example parse tree stored in the storage means, a difference between each input subtree and each example subtree is determined. An input subtree and an example subtree matching processing unit (7) for searching for subtree pairs that are phonologically close to each other and generating and outputting pairs of combinations of the subtrees; Input subtree output from the matching processing unit (7) and examples Pairs of combination of partial tree by referring to the predetermined word phoneme conversion dictionary, each subtree phoneme string conversion processing unit for converting the phoneme string (8)
And calculating a phonological distance between the phoneme sequence of the input subtree output from the subtree phoneme sequence conversion processing unit (8) and the phoneme sequence of the example subtree, and calculating the calculated phonological distance. A phonological distance calculation processing unit (9) that outputs a pair of an input subtree and an example subtree that is equal to or smaller than another predetermined threshold value, and the input analytic tree restoration determination unit (3) determines that restoration is to be performed. Error correction processing is performed on the input parse tree by repairing the input parse tree using the pair of the input subtree and the example subtree output from the phonological distance calculation processing unit (9). And a parse tree repair processing unit (4) for outputting an error correction result of the input parse tree.