JP2012018201A - Text correction and recognition method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a device which, when a recognized sentence as an output of voice recognition or the like includes an error, efficiently estimates a position of the recognized error to present a candidate for correcting the error, and preferentially corrects a serious error.SOLUTION: A device extracts a feature word list as an aggregate of feature words by feature word extraction processing from a recognized sentence; determines that words not included in the feature word list of words included in the recognized sentence are a noncontextualized word (hereinafter referred to as a peculiar word) and possibly have an error; compares each of the words determined as the peculiar word with each of the words included in the feature word list in terms of similarity therebetween; and determines a word having high similarity as a candidate for correcting the peculiar word. In this manner, the feature words are extracted in advance.

Description

  The present invention relates to a text correction method and a recognition method for correcting errors included in output text such as speech recognition and character input.

  A technique for receiving speech generated by humans and outputting a text sentence as a recognition result is called speech recognition. As a language model for determining a recognition word in speech recognition, what is called n-gram is widely used. The n-gram is a method of determining each word candidate when performing sequential speech recognition using the appearance probability of connection of n words. The appearance probability is calculated in advance from a large amount of sample sentences.

  However, in speech recognition using a single language model, it is inevitable that a recognition error is included in the recognition result. For example, since n-gram focuses only on a sequence of adjacent n words, a recognition result in consideration of a more global grammar and context cannot be obtained.

  Therefore, a technique has been proposed in which recognition accuracy is improved by combining a plurality of language models, or a misrecognition included in a recognition result of one language model is corrected by another language model.

  For example, in Patent Document 1, among words included in a recognition result of a first language model (for example, n-gram), a word that may be erroneous is recognized by a second model (example sentence model). A recognizing voice recognition device is disclosed. At this time, in the first language model, the effect of reducing the processing load due to the second model is shown by performing recognition with the second model only on the portion estimated to be an error.

  Patent Document 2 discloses an information processing apparatus that improves recognition performance by combining global context processing and local context processing in speech recognition processing.

JP-A-2005-84436 JP 2008-181537 A

  In a conventional apparatus that estimates an error included in a recognized sentence output by the first language model and corrects misrecognition using the second language model, as shown in Patent Document 1, each recognition sentence It is necessary to perform processing for searching for correction candidates from the recognition vocabulary set of the second language model for the words, and it is necessary to perform processing with high load for each word of the recognized sentence. Therefore, in Patent Document 1, the processing load by the second language model is reduced by performing correction processing by the second language model only for words that are likely to be erroneous among the recognition results by the first language model. However, there is a problem that an error included in a word that is not a target to be corrected cannot be corrected.

  In addition, in the related art, since each word of the recognized sentence is uniquely searched for the correction processing target regardless of the context of the recognized sentence, it is possible to perform processing preferentially for correction of an important error in the recognized sentence. could not.

  When a plurality of language models as shown in Patent Document 2 are combined and recognition processing is performed in real time, recognition processing using a plurality of language models needs to be performed for each word. Moreover, it is not possible to preferentially allocate computing resources for correcting important errors.

  The present invention has been made in view of the above problems, and corrects misrecognition using the second language model at high speed with respect to misrecognition of the output result by the first language model, or is important. It is an object of the present invention to provide a text correction method and a recognition method capable of preferentially allocating computer resources to various errors.

In order to achieve the above object, a text correction method according to the present invention includes:
(1) A text correction method for correcting an error word included in a text by error correction,
In the error correction, the text and a vocabulary set consisting of another set of words,
A feature word extraction step of extracting a set of feature words from the text and vocabulary set by comparing with the frequency of appearance of the words included in them,
Calculating a similarity between a word included in the text and a word included in the set of feature words;
A word included in the set of feature words determined to be similar to a word included in the text is output as a correction candidate word for the word included in the text.

  In this way, the amount of processing to be corrected can be reduced by extracting feature words in advance by the feature word extraction step.

  Further, in (1), (2) the error correction includes a singular word extraction step of extracting a word that is not included in the feature word set among words included in the text as a singular word set; Similarity calculation and correction candidate output may be performed for each word of the singular word set.

  In addition, in (1), (3) the error correction is performed in addition to the similarity between the word included in the text and the word included in the feature word set, as well as the word included in the feature word set. Whether or not a word included in the set of feature words is output as a correction candidate word may be determined based on the feature degree.

  In addition, the error correction may be a frequency at which words included in the set of feature words are included in the text in addition to the similarity between the words included in the text and the words included in the set of feature words. Thus, it may be determined whether or not a word included in the set of feature words is output as a correction candidate word.

  In addition, in (1), (4) the error correction is performed by adding a word included in the feature word set in addition to the similarity between the word included in the text and the word included in the feature word set. Depending on the frequency included in the text, it may be determined whether or not a word included in the set of feature words is output as a correction candidate word.

  In addition, in (1), (5) the error correction is performed such that, in addition to the similarity between the word included in the text and the word included in the set of feature words, the word included in the text includes the text Whether or not to output a word included in the set of feature words as a correction candidate word may be determined based on the frequency included in the word.

  In (1), (6) the text is a text generated from a first recognized vocabulary set, and the error correction is included in the word included in the text and the feature word set. A word included in the set of feature words is corrected as a candidate word according to whether or not a word included in the set of feature words is included in the first recognition vocabulary set in addition to the similarity to the word. It may be determined whether to output as.

(7) In the recognition method according to the present invention, the error word included in the text output by the recognition step by the first recognition model using the non-text data as an input is converted into the error word by the error correction by the second recognition model. A recognition method for correcting
The recognition step generates a set of recognition words in time series from recognition words recognized by the first recognition model from non-text data in time series,
The error correction is performed by comparing the set of recognized words and the vocabulary set included in the second recognition model in terms of the frequency of appearance of the words included in the recognition word set and the vocabulary set. A feature word extraction step for extracting a set of words;
Calculating the similarity between the set of recognized words and the words included in the set of feature words;
A word included in the set of feature words determined to be similar to a word included in the set of recognized words is output as a correction candidate word for a word included in the set of recognized words.

  In this way, the amount of processing to be corrected can be reduced by extracting feature words in advance by the feature word extraction step.

  Further, in (7), (8) the error correction includes a singular word extraction step of extracting a word that is not included in the feature word set among words included in the text as a set of singular words, Similarity calculation and correction candidate output may be performed for each word of the singular word set.

  In addition, in (7), (9) the error correction is performed in addition to the similarity between the word included in the text and the word included in the feature word set, as well as the word included in the feature word set. Whether or not a word included in the set of feature words is output as a correction candidate word may be determined based on the feature degree.

  In addition, the error correction may be a frequency at which words included in the set of feature words are included in the text in addition to the similarity between the words included in the text and the words included in the set of feature words. Thus, it may be determined whether or not a word included in the set of feature words is output as a correction candidate word.

  Further, in (7), (10) the error correction is performed in such a manner that the word included in the feature word set is added to the similarity between the word included in the text and the word included in the feature word set. Depending on the frequency included in the text, it may be determined whether or not a word included in the set of feature words is output as a correction candidate word.

  Further, in (7), (11) the error correction is performed in such a manner that the word included in the text is added to the text in addition to the similarity between the word included in the text and the word included in the feature word set. Whether or not to output a word included in the set of feature words as a correction candidate word may be determined based on the frequency included in the word.

  In (7), (12) the text is a text generated from a first recognized vocabulary set, and the error correction is included in the word included in the text and the feature word set. A word included in the set of feature words is corrected as a candidate word according to whether or not a word included in the set of feature words is included in the first recognition vocabulary set in addition to the similarity to the word. It may be determined whether to output as.

  In (7), (13) the error correction may change a frequency of extracting the feature word from the vocabulary set.

  In (7), (14) the error correction may change a frequency of extracting the feature word from the vocabulary set based on an extraction time interval.

  In (7), (15) the error correction may change a frequency of extracting the feature word from the vocabulary set based on a processing load situation of a computer.

  According to the present invention, it is possible to perform estimation and replacement processing of a misrecognized portion at high speed with respect to a recognized sentence output by the first recognition model, by misrecognition correction processing using the second recognition model. Further, correction processing can be performed with priority given to important mistakes according to the load state of computer resources.

It is a figure which shows the structure of the mobile telephone which concerns on one Embodiment of this invention. It is a block diagram which shows the structure of the mobile telephone of FIG. It is an example sentence which shows the processing result of this invention. It is a figure which shows notionally the operation | movement of the mobile telephone of FIG. It is a flowchart which shows the procedure of a misrecognition correction process. It is a flowchart which shows the procedure of a feature word extraction process. It is a figure which shows an example of a word list. It is a figure which shows an example of a feature word list. It is a flowchart which shows the procedure of a correction | amendment determination process. It is an example of the result of a phoneme sequence conversion process. It is a flowchart which shows the process sequence of the text processing apparatus which concerns on one Embodiment of this invention. It is a figure which shows the structure of the system which concerns on one Embodiment of this invention. It is a figure which shows the structure of the mobile telephone of FIG. It is a figure which shows the structure of the server of FIG. It is a flowchart which shows the process sequence of the text processing apparatus which concerns on one Embodiment of this invention.

  Hereinafter, a speech recognition method according to a first embodiment of the present invention will be described with reference to the drawings.

  As shown in FIG. 1, a mobile phone 10 in which the text correction method of the present invention is implemented includes a microphone 101, a main memory 102, a CPU 103, a display unit 104, and a secondary storage device 105. The These are connected to the internal bus 105.

  The secondary storage device 105 stores a speech recognition program 107, a computer load acquisition program 108, a misrecognition correction program 109, a word list creation program 112, a feature word extraction program 115, and a document DB creation program 118. Are loaded into the main memory 102 as necessary, and the CPU 103 operates according to these programs, thereby executing a later-described erroneous recognition correction process.

  The erroneous recognition correction program 109 further includes partial programs of a phoneme string conversion program 110 and a correction determination program 111.

  The word list creation program 112 further includes partial programs such as a morpheme analysis program 113 and a frequency calculation program 114.

  The feature word extraction program 115 further includes partial programs such as a similar document search program 116 and a word feature degree calculation program 117.

  The secondary storage device 105 further stores a document DB 119 and a language model 120. The document DB 119 is a table in which, for a plurality of text documents, a table representing the appearance frequency of each word in the text document is managed for each text document. The plurality of text documents are, for example, conversation sentences on various topics, explanation sentences for each word in the dictionary, transmission / reception sentences of e-mails, and the like. The language model 120 is dictionary data used for speech recognition by the speech recognition program 107.

  FIG. 2 is a diagram conceptually showing the operation of the misrecognition correction process by the mobile phone 10 of the present embodiment.

  A human voice input from the microphone 101 is converted into a recognized sentence 212 by voice recognition processing. This recognition sentence 212 is considered to include misrecognition.

  By the word list creation process, the frequency of each word included in the recognized sentence 212 is calculated, and the word list 213 is created. A feature word list 215 is created from the created word list 213 by feature word extraction processing.

  The corrected sentence 216 is generated by comparing the similarity between the created feature word list 215 and the original word list 213 by the correction determination process, and is displayed to the user through the display unit 104.

  Specifically, the speech recognition processing, the word list creation processing, the feature word extraction processing, and the correction determination processing are respectively the speech recognition unit 202, the word list creation unit 203, the feature word extraction unit 206, and the error that are the functional units of the CPU 103. It is executed by the recognition correction unit 209.

  3 and 4 are conceptual diagrams showing an example of the erroneous recognition correction process of this embodiment. In this example, the recognition sentence 212 output by the speech recognition process for a human utterance sentence includes two words “planting tree” and “school song” as errors. A set of words included in the recognition sentence 212 is set as a set W. A set of feature words extracted by the feature word extraction process for the recognized sentence 212 is a feature word list 215. This is set K. Of the words in the set W, words that are not included in the set K are determined to be words that do not conform to this context (hereinafter referred to as singular words), and are determined to have a possibility of error. For each word determined to be a singular word, the degree of similarity is compared with each word included in the set K, and a word determined to have a high degree of similarity is taken as a candidate for correcting the singular word. In this example, correction candidates “meal” and “expensive” are selected for the above-mentioned “tree planting” and “school song”, and a correction sentence 216 is output.

  Hereinafter, the details of each of the processes executed by the mobile phone 10 will be described. FIG. 5 is a flowchart showing a procedure for erroneous recognition correction.

  The speech recognition unit 202 performs speech recognition from the speech input from the microphone 101, and generates a recognition sentence 212 from each word included in the language model 120 (step S501). The morphological analysis unit 204 of the word list creation unit 203 breaks down the recognized sentence 212 into words, and then the frequency calculation creation unit 205 counts the frequency of each word included in the recognized sentence 212 to create the word list 213. (Step S502). FIG. 7 shows an example of the word list 213.

  Next, the feature word extraction unit 206 generates a feature word list 215 from the word list 213 and the word list 213 of each document in the document DB 119 (step S503). FIG. 8 shows an example of the feature word list 215. Details of the processing in step S503 will be described later. The extracted feature word list 215 is data in which a list of characteristic words corresponding to the topic of the recognized sentence 212 is arranged in descending order of the feature degree.

  Subsequently, the misrecognition correction unit 209 performs processing for each entry in the word list 213 (steps S504 to S515). The misrecognition correction unit 209 checks, for each entry in the word list 213, whether or not there is the same word in the feature word list 215, and if it is not determined that this word is a singular word, in step S505, The process proceeds to the next entry process (Yes in step S506). If it is not in the feature word list 215 (No in step S506), the process proceeds to the correction determination process after step S507.

  The current load of the computer is acquired from the computer load acquisition unit 201, and a comparison number m ′ with a feature word list described later is determined according to the load (step S 507), and the correction determination unit 211 of the misrecognition correction unit 209. However, similarities of the top m ′ of the feature word list 215 are compared in descending order of feature (steps S508 to S514).

  The correction determination unit 211 compares the similarity between Wi and Kj, and calculates a determination value as to whether or not to be a correction candidate (step S511). Here, Wi represents the i-th word in the word list 213, and Kj represents the j-th word in the feature word list 215. Details of step S511 will be described later.

If the determination value calculated in step S511 exceeds the threshold value (Yes in step S512), it is determined to be misrecognition, and Kj is stored as a misrecognition correction candidate for Wi (step S513). If it is below the threshold (No in step S512), it is determined that Kj is not a Wi correction candidate, and the process returns to the next feature word (step S514).
When the comparison is completed with all m ′ feature words (Yes in step S509), i is added in step S510, and then the process returns to step S505 to be reflected in the processing of the next input word (step S505). When the processing of the word list of all the recognized sentences is finished (Yes in step S505), among the words in the word list 213, the corresponding word of the recognized sentence 212 is found for the word for which the correction candidate is found in step S513. Then, it is replaced with a correction candidate word and output (step S515). In this case, when there are a plurality of correction candidates, the correction candidate determined to have the highest similarity in step S511 is adopted.

  FIG. 6 is a flowchart showing the procedure of the feature word extraction process in step S503.

  The similar document search unit 207 of the feature word extraction unit 206 performs an inner product operation of vectors between the word list 213 created from the recognized sentence 212 and the word list 213 of each document stored in the document DB 119 (step S601). Here, each element of the vector is an appearance frequency of each word. Therefore, the number of words is different from the number of dimensions of a vector. Then, a fixed number (assumed to be α) is extracted as the similar document 214 with respect to the recognized sentence 212 in order from the smallest inner product value (step S602).

  The word feature degree calculation unit 208 calculates the feature degree for each appearing word for each of the extracted similar documents 214 by a method called tf-idf (step S603). tf-idf is a calculation method that is widely used as a method of calculating the word feature. When the appearance frequency of the word w in the document d is tf and the number of appearing documents is df in all documents (in this example, all documents in the document DB 109), the tf-idf value in the document d of the word w is In this embodiment, the calculation is performed using the following calculation formula. However, this is an example of a method for calculating the feature degree, and is not limited thereto.

Tf-idf value of word w = tf / idf
Then, an average between the extracted similar documents 214 is taken for the calculated if-idf value of each word (step S604). Then, m items in descending order of the average value are generated as the feature word list 215 (step S605).

FIG. 9 is a flowchart illustrating the procedure of the correction determination process in step S511.
First, the phoneme string conversion unit 210 of the misrecognition correction unit 209 converts the word Wi and the word Kj into kana and then converts them into phoneme string notation (step S901). FIG. 10 shows an example of conversion to phoneme string notation for each of Wi = “tree planting” and Kj = “meal”.

  Next, the edit distance between each Romaji notation is calculated (step S902). The edit distance is a value obtained by calculating how many times insertion / deletion / replacement of word A can be performed to change to word B. There is a method of changing the weight of insertion / deletion / replacement, such as increasing the weight of replacement. The value calculated in this way is the edit distance, and is one of the indices of similarity between word A and word B (determined that the smaller the value, the higher the similarity).

  In this embodiment, Wi and Kj are changed to phoneme strings, but this can be changed depending on the field of application of the present invention. As another example, a method of converting to Roman character conversion in character input is conceivable.

  Next, the number t of Kj appearing in the word list 213 of the recognized sentence 212 is calculated (step S903). Finally, a final determination value is calculated from the edit distance and the number of appearances. For example, it calculates with the following formulas. In this example, the higher the number of times Kj appears in the recognized sentence 212, the higher the degree of similarity. That is, when it appears in a portion other than the correction determination target, it is a method for increasing the similarity of the correction candidate.

Judgment value = (Edit distance / (Number of appearances + 1))
Further, as another calculation method, a method of changing the weight depending on whether or not Kj is a word included in the language model 120 is also included in the present invention. That is, since a word that is not in the language model 120 does not appear in the recognized sentence 212, this is a method of increasing the probability of being a correction candidate.

  Further, the present invention includes a method of using the characteristic degree of Kj as a weight. This is a method of increasing the probability that a word having a high Kj feature level, that is, a high importance level, is a correction candidate.

  The determination value calculation process illustrated here is an example, and any method may be used as long as it is a method for determining replacement between words.

  In the text correction method according to the first embodiment, for all the words in the recognized sentence, the language model for correction is searched only once, and a list of feature words that are correction candidates is extracted in advance. Thus, it is possible to efficiently perform correction determination processing for all words in the recognized sentence.

  Hereinafter, the speech recognition method according to the second embodiment of the present invention will be described with reference to the drawings. In the second embodiment, an example of a speech recognition method in which correction processing for erroneous recognition is performed in parallel while performing speech recognition processing for an uttered speech in real time will be described. Note that description of the same processing as in the first embodiment will be omitted. Details of the correction processing will be described with reference to the flowchart of FIG.

  In the correction processing in this embodiment, a task for performing speech recognition and a task for performing feature word extraction for performing correction processing operate in parallel. The task of performing speech recognition repeatedly performs speech recognition processing and outputs sequentially recognized recognition words 1101 (same as step S501 in the first embodiment) while there is speech input (Yes in step S1101). Next, a misrecognition determination process is implemented with respect to the recognition word 1101 (similar to step S504-step S514 of 1st embodiment). If it is determined that the word is an incorrect word as a result of the erroneous recognition determination, the correction candidate word is added to the recognized sentence 1102. If it is determined that there is no mistake, the original recognized word 1101 is added to the recognized sentence 1102.

  On the other hand, the task that performs feature word extraction performs feature word extraction processing as necessary while the recognition task is operating (No in step S1104). First, a feature word extraction necessity determination is performed. Here, the timing for performing the process of extracting feature words is determined. Examples of determination criteria include whether a certain time has passed since the previous determination, whether the number of words included in the recognized sentence has increased by a certain number, or whether the CPU load on the computer is below a certain level (step S1105). If it is determined that feature word extraction is necessary (Yes in step S1106), the feature word list 215 is extracted from the recognized sentence 212 (similar to step S503 in the first embodiment). If it is determined that the feature word is unnecessary (No in step S1106), the process is stopped for a certain time (step S1107).

  In the present embodiment, the feature word extraction processing step S503 is not performed on the recognition word 1101 output in step S501, but the feature word list 215 extracted intermittently based on the determination criterion in step S1105 is used. In step S1102, erroneous recognition is determined. Since the feature word extraction process with a high processing load is not performed each time the word recognition is performed for speech recognition, and the erroneous recognition correction in step S1103 is performed for each word, the second recognition model for correction is searched ( It is possible to perform correction processing on all words of the recognition result by the first recognition model in real time while suppressing a calculation load necessary for searching the document DB 119 in this example.

  Hereinafter, a voice recognition method according to a third embodiment of the present invention will be described with reference to the drawings.

  12, 13 and 14 are diagrams showing the system configuration of the embodiment. In the present embodiment, a system is shown in which the feature word extraction process shown in the first and second embodiments is executed by another computer connected by a network.

  The word list 213 created by the mobile phone 10 is transmitted to the server 20 (steps S1301 and S1302), and the server extracts the feature word list 215 based on the received word list 213 (step S503). The server transmits the extracted feature word list 215 to the mobile phone (steps S1303 and S1304), and the mobile phone performs a misrecognition correction process to generate a correction sentence 216 (step S1102).

  In this example, an example is shown in which processing with a large amount of data and a large amount of calculation is performed on the server side.

  According to the present invention, it is possible to efficiently execute estimation / correction candidate presentation of a misrecognized portion with respect to a recognized sentence including an error, and it is possible to preferentially perform correction processing for an important error. Therefore, it can be applied to all corrections of recognition techniques such as voice recognition, language input system in PC, voice command recognition, etc., and can be easily applied to mobile phones and navigation systems with limited computer resources.

  DESCRIPTION OF SYMBOLS 10 ... Mobile phone, 101 ... Microphone, 102 ... Main memory, 103 ... CPU, 104 ... Display part, 105 ... Internal bus, 106 ... Secondary storage device, 107 ... Voice recognition program, 109 ... Error recognition correction program, 112 ... Word list creation program, 115 ... feature word extraction program, 119 ... document DB, 120 ... language model, 116 ... scene determination unit, 211 ... keypad, 202 ... voice recognition unit, 203 ... word list creation unit, 206 ... feature word Extraction unit, 209 ... erroneous recognition correction unit, 212 ... recognition sentence, 213 ... word list, 215 ... feature word list, 216 ... correction sentence

Claims (15)

A text correction device that corrects an error word included in a text by an error correction unit,
The error correction unit includes the text and a vocabulary set composed of another set of words.
A feature word extraction unit that extracts a set of feature words from the text and the vocabulary set by comparing the appearance frequencies of the words included in them;
Calculating a similarity between a word included in the text and a word included in the set of feature words;
A text correction apparatus that outputs a word included in the set of feature words determined to be similar to a word included in the text as a correction candidate word for the word included in the text. The error correction unit includes a singular word extraction unit that extracts words included in the text that are not included in the feature word set as a singular word set. On the other hand, calculate the similarity and output correction candidates.
The text correction apparatus according to claim 1. The error correction unit is configured to determine the set of feature words based on the degree of feature of the words included in the set of feature words in addition to the similarity between the words included in the text and the words included in the set of feature words. To determine whether to output the words included in as correction candidate words,
The text correction apparatus according to claim 1. In addition to the similarity between the word included in the text and the word included in the feature word set, the error correction unit may determine whether the word included in the feature word set is included in the text. Determining whether to output words included in the set of feature words as correction candidate words;
The text correction apparatus according to claim 1. In addition to the similarity between the words included in the text and the words included in the set of feature words, the error correction unit determines the characteristics according to the frequency that the words included in the text are included in the text. Determine whether to output words included in the set of words as correction candidate words;
The text correction apparatus according to claim 1. The text is a text generated from a first recognized vocabulary set, and the error correction unit, in addition to the similarity between a word included in the text and a word included in the set of feature words, Determining whether or not to output a word included in the set of feature words as a correction candidate word depending on whether or not a word included in the set of feature words is included in the first recognition vocabulary set;
The text correction apparatus according to claim 1. A recognition device for correcting an error word included in a text output by a recognition unit having a first recognition model with non-text data as an input, and correcting the error word by an error correction unit having a second recognition model,
The recognition unit generates a set of recognition words in time series from recognition words recognized by the first recognition model from non-text data in time series,
The error correction unit compares the set of recognized words and the vocabulary set included in the second recognition model with the appearance frequency of the words included in them, and from the set of recognized words and the vocabulary set, A feature word extraction unit for extracting a set of feature words;
Calculating the similarity between the set of recognized words and the words included in the set of feature words;
A recognition apparatus that outputs a word included in the set of feature words determined to be similar to a word included in the set of recognized words as a correction candidate word for a word included in the set of recognized words . The error correction unit includes a singular word extraction unit that extracts words included in the text that are not included in the feature word set as a singular word set. On the other hand, calculate the similarity and output correction candidates.
The recognition apparatus according to claim 7. The error correction unit is configured to determine the set of feature words based on the degree of feature of the words included in the set of feature words in addition to the similarity between the words included in the text and the words included in the set of feature words. To determine whether to output the words included in as correction candidate words,
The recognition apparatus according to claim 7. In addition to the similarity between the word included in the text and the word included in the feature word set, the error correction unit may determine whether the word included in the feature word set is included in the text. Determining whether to output words included in the set of feature words as correction candidate words;
The recognition apparatus according to claim 7. In addition to the similarity between the words included in the text and the words included in the set of feature words, the error correction unit determines the characteristics according to the frequency that the words included in the text are included in the text. Determine whether to output words included in the set of words as correction candidate words;
The recognition apparatus according to claim 7. The text is a text generated from a first recognized vocabulary set, and the error correction unit, in addition to the similarity between a word included in the text and a word included in the set of feature words, Determining whether or not to output a word included in the set of feature words as a correction candidate word depending on whether or not a word included in the set of feature words is included in the first recognition vocabulary set;
The recognition apparatus according to claim 7. The error correction unit changes a frequency of extracting the feature word from the vocabulary set;
The recognition apparatus according to claim 7. The error correction unit changes a frequency of extracting the feature word from the vocabulary set based on an extraction time interval;
The recognition apparatus according to claim 7. The error correction unit changes a frequency of extracting the feature word from the vocabulary set based on a processing load situation of a computer;
The recognition apparatus according to claim 7.

Images