JP2021092713A - Correction candidate specification device, correction candidate specification method and correction candidate specification program - Google Patents

Abstract

To appropriately specify a correction candidate while a calculation load is suppressed.SOLUTION: A word extraction unit 22 extracts a word included in text data constituted of a plurality of sentences. A similarity calculation unit 25 sets the plurality of sentences as target sentences, sets at least one of sentences before and after the target sentence as an opponent sentence, and calculates similarity between a word extracted from the target sentence and a word extracted from the opponent sentence as time series similarity. A candidate specification unit 26 sets the plurality of sentences as the target sentences and specifies the target sentence as a correction candidate when the time series similarity calculated on the target sentence is lower than a first threshold.SELECTED DRAWING: Figure 2

Description

The present invention relates to a technique for estimating a position where text data composed of a plurality of sentences needs to be modified.

Conversations are converted into text data using voice recognition technology. For example, a plurality of voice collecting devices are provided in the conference room, and voice data of conversation in the conference room is collected and converted into text data. At this time, each voice collecting device is aggregated into one voice data by a voice mixer device, and then converted into text data by using voice recognition technology. As a result, the content spoken at the meeting is saved as text data.

However, the text data may contain errors. The errors include erroneous recognition of speech and erroneous conversion of the recognized speech into homonyms. Therefore, in order to obtain correct text data, it is necessary to manually check the text data and correct it correctly.

Patent Document 1 describes a technique for identifying correction candidates in text data recognized by the first language model and performing correction by the second language model.
In Patent Document 1, a set of characteristic words is extracted based on the frequency of appearance of words. Then, the similarity between the singular word not included in the feature word and the feature word is calculated, and the singular word having a low similarity is specified as a correction candidate.

Japanese Unexamined Patent Publication No. 2012-18201

In the technique described in Patent Document 1, the frequency of occurrence of each word included in the text data is calculated. Therefore, the calculation load is high.
An object of the present invention is to make it possible to appropriately identify correction candidates while suppressing a calculation load.

The modification candidate identification device according to the present invention is
A word extractor that extracts words contained in text data composed of multiple sentences, and a word extractor
The words extracted from the target sentence by the word extraction unit and the word extracted from the target sentence by the word extraction unit and the word extracted from the target sentence by the word extraction unit, with each of the plurality of sentences as the target sentence and at least one of the sentences before and after the target sentence as the partner sentence. A similarity calculation unit that calculates the similarity with words extracted from the other sentence as a time-series similarity,
A candidate identification unit that specifies the target sentence as a correction candidate when the time-series similarity calculated by the similarity calculation unit for the target sentence is lower than the first threshold value, with each of the plurality of sentences as the target sentence. And.

The similarity calculation unit uses sentences other than the correction candidates specified by the candidate identification unit as comparison sentences, and the similarity between the words extracted from the correction candidates and the words extracted from each of the comparison sentences. Calculate the degree as a combination similarity and
When the combination similarity of any word extracted from the comparative sentence is lower than the second threshold value, the candidate identification unit identifies the word extracted from the correction candidate as a correction target.

The text data was generated by synthesizing the converted partial text data of the voice data collected by each of the plurality of voice collecting devices.

The similarity calculation unit calculates the combination similarity with the sentences included in the partial text data other than the partial text data including the correction candidate as the comparison target.

The correction candidate identification device further
A vector generation unit for generating an expression vector representing the meaning of the target word is provided with the word extracted by the word extraction unit as the target word.
The similarity calculation unit calculates the similarity using the expression vector generated by the vector generation unit.

The similarity calculation unit calculates the inner product of the expression vector for the word extracted from one sentence and the expression vector for the word extracted from the other sentence as the similarity.

The method for identifying modification candidates according to the present invention is
The word extraction unit extracts words contained in text data composed of multiple sentences,
The similarity calculation unit extracts each of the plurality of sentences as a target sentence, sets at least one of the sentences before and after the target sentence as the other sentence, and extracts the words extracted from the target sentence and the other sentence. Calculate the similarity with the word as the time series similarity,
The candidate identification unit specifies each of the plurality of sentences as a target sentence, and when the time-series similarity calculated for the target sentence is lower than the first threshold value, the target sentence is specified as a correction candidate.

The modification candidate identification program according to the present invention
Word extraction processing that extracts words contained in text data composed of multiple sentences, and
Each of the plurality of sentences is a target sentence, and at least one of the sentences before and after the target sentence is set as a partner sentence, and the word extracted from the target sentence by the word extraction process and the word extracted by the word extraction process are described. Similarity calculation processing that calculates the similarity with the word extracted from the other sentence as the time series similarity,
Candidate identification processing for specifying the target sentence as a correction candidate when the time-series similarity calculated by the similarity calculation process for the target sentence is lower than the first threshold value, with each of the plurality of sentences as the target sentence. Make the computer function as a correction candidate identification device.

In the present invention, each of a plurality of sentences is set as a target sentence, and at least one sentence before and after the target sentence is set as a partner sentence, and a correction candidate is specified by calculating as a time series similarity. This makes it possible to appropriately identify correction candidates while suppressing the calculation load.

The block diagram of the voice recognition system 1 which concerns on Embodiment 1. FIG. The block diagram of the correction candidate identification apparatus 10 which concerns on Embodiment 1. FIG. The flowchart which shows the operation of the correction candidate identification apparatus 10 which concerns on Embodiment 1. FIG. The explanatory view of the partial text data 51 which concerns on Embodiment 1. FIG. Explanatory drawing of Skip-Gram model. The explanatory view of the synthetic text data 54 which concerns on Embodiment 1. FIG. The explanatory view of the time series similarity calculation process which concerns on Embodiment 1. FIG. The explanatory view of the candidate identification process which concerns on Embodiment 1. FIG. The block diagram of the correction candidate identification apparatus 10 which concerns on modification 4. FIG. The flowchart which shows the operation of the correction candidate identification apparatus 10 which concerns on Embodiment 2. The explanatory view of the combination similarity calculation process which concerns on Embodiment 2.

Embodiment 1.
*** Explanation of configuration ***
The configuration of the voice recognition system 1 according to the first embodiment will be described with reference to FIG.
The voice recognition system 1 includes a correction candidate identification device 10, a plurality of voice collection devices 41, and a plurality of voice recognition devices 42.
The correction candidate identification device 10 is a computer that identifies correction candidates in voice-recognized text data. The correction candidates indicate the parts where there is a possibility that there is an error in speech recognition and it is presumed that correction is necessary. Here, the speech recognition error is at least one of a speech misrecognition and a misconversion of the recognized speech to a homonym. Each voice collecting device 41 is a device that collects voice data. Each voice collecting device 41 is, as a specific example, a microphone. Each voice recognition device 42 is a device that converts voice data into text data by voice recognition technology.

In FIG. 1, a voice recognition device 42 is provided for each voice collecting device 41. However, the present invention is not limited to this, and one voice recognition device 42 may be provided for each of the plurality of voice collection devices 41.

In the first embodiment, it is assumed that a voice collecting device 41 is provided for each person, and a person corresponding to each voice collecting device 41 is talking. This is assumed, for example, when a web conference is held and the participants of the conference are in different places. Further, it is assumed that a voice collecting device 41 is provided for each seat in the conference room.

The configuration of the modification candidate identification device 10 according to the first embodiment will be described with reference to FIG.
The modification candidate identification device 10 includes hardware for a processor 11, a memory 12, a storage 13, and a communication interface 14. The processor 11 is connected to other hardware via a signal line and controls these other hardware.

The processor 11 is an IC (Integrated Circuit) that performs processing. Specific examples of the processor 11 are a CPU (Central Processing Unit), a DSP (Digital Signal Processor), and a GPU (Graphics Processing Unit).

The memory 12 is a storage device that temporarily stores data. Specific examples of the memory 12 are SRAM (Static Random Access Memory) and DRAM (Dynamic Random Access Memory).

The storage 13 is a storage device for storing data. As a specific example, the storage 13 is an HDD (Hard Disk Drive). The storage 13 includes SD (registered trademark, Secure Digital) memory card, CF (CompactFlash, registered trademark), NAND flash, flexible disk, optical disk, compact disk, Blu-ray (registered trademark) disk, DVD (Digital Versaille Disk), and the like. It may be a portable recording medium.

The communication interface 14 is an interface for communicating with an external device. As a specific example, the communication interface 14 is a port of Ethernet (registered trademark), USB (Universal Serial Bus), HDMI (registered trademark, High-Definition Multimedia Interface).

The correction candidate identification device 10 includes a data acquisition unit 21, a word extraction unit 22, a vector generation unit 23, a text composition unit 24, a similarity calculation unit 25, and a candidate identification unit 26 as functional components. .. The functions of each functional component of the modification candidate identification device 10 are realized by software.
The storage 13 stores a program that realizes the functions of each functional component of the modification candidate identification device 10. This program is read into the memory 12 by the processor 11 and executed by the processor 11. As a result, the functions of each functional component of the modification candidate identification device 10 are realized.

In FIG. 2, only one processor 11 was shown. However, the number of processors 11 may be plural, and the plurality of processors 11 may execute programs that realize each function in cooperation with each other.

*** Explanation of operation ***
The operation of the modification candidate identification device 10 according to the first embodiment will be described with reference to FIGS. 3 to 8.
The operation procedure of the modification candidate identification device 10 according to the first embodiment corresponds to the modification candidate identification method according to the first embodiment. Further, the program that realizes the operation of the modification candidate identification device 10 according to the first embodiment corresponds to the modification candidate identification program according to the first embodiment.

(Step S1: Data acquisition process in FIG. 3)
The data acquisition unit 21 acquires partial text data 51 in which the voice data collected by each voice collecting device 41 is text data converted by each voice recognition device 42.
Specifically, each voice collecting device 41 collects voice data spoken by a person assigned to the voice collecting device 41. Then, the voice collected by each voice collecting device 41 is converted into voice data by the voice recognition device 42 corresponding to the voice collecting device 41, and the partial text data 51 is generated. The data acquisition unit 21 acquires the partial text data 51 generated by each voice recognition device 42.
As shown in FIG. 4, each partial text data 51 includes a plurality of sentences 52. When the utterance interval is longer than the reference time, it is recognized as another sentence 52 and the partial text data 51 is generated. In each sentence 52, the time when the utterance is started is set.

(Step S2 in FIG. 3: Word extraction process)
The word extraction unit 22 extracts the words included in the target partial text data 51 by using each partial text data 51 acquired in step S1 as the target partial text data 51.
Specifically, the word extraction unit 22 uses an existing word extraction technique to extract words included in the target partial text data 51. As a specific example, the word extraction unit 22 performs morphological analysis on the target partial text data 51 and extracts specific parts of speech such as nouns, verbs, and adjectives to obtain words included in the target partial text data 51. Extract.

(Step S3 in FIG. 3: Vector generation process)
The vector generation unit 23 uses each word extracted in step S2 as a target word to generate an expression vector 53 representing the meaning of the target word.
Specifically, the vector generation unit 23 generates the expression vector 53 for the target word by using a technique such as Skip-Gram model or CBoW (Continuous Bag-of-Words). In the first embodiment, the case where the Skip-Gram model is used will be described as an example.

A method of generating the representation vector 53 using the Skip-Gram model will be described.
As a preliminary preparation, the vector generation unit 23 provides the Skip-Gram model with text data of a large number of sentences including words expected to be included in the partial text data 51. For example, when the partial text data 51 in which the voice data of the conversation of the meeting is converted is acquired, it is conceivable to give the text data of the minutes of the past meeting to the Skip-Gram model.
As a result, for each word contained in the text data given to the Skip-Gram model, a vector representing a word that is predicted to appear around the word is generated. Specifically, as shown in FIG. 5, for each word, a vector having the probability that another word appears around the word is generated as an element. Note that FIG. 5 shows a vector having three elements for each word. However, in reality, for each word, a vector with many elements, such as hundreds, is generated.
Here, for similar words, the words appearing around the word are similar. Therefore, a vector with similar elements is generated for similar words.

The vector generation unit 23 stores the vector for each word in the storage 13. The vector generation unit 23 may store the vector for each word in an external storage device of the correction candidate identification device 10.
Then, when the word is extracted in step S2, the vector generation unit 23 searches the storage 13 for the vector of the target word which is the extracted word. The vector generation unit 23 sets the vector that hits the search in the expression vector 53 for the target word. As a result, the expression vector 53 for the target word is generated.

(Step S4 in FIG. 3: Text composition processing)
The text composition unit 24 synthesizes each partial text data 51 acquired in step S1 to generate the composite text data 54.
Specifically, the text synthesizing unit 24 generates the composite text data 54 by arranging the sentences 52 included in each partial text data 51 in chronological order and synthesizing them. As a specific example, two partial text data 51 in which the voice data of the conversation between the two persons shown in FIG. 4 is converted is synthesized. Then, as shown in FIG. 6, synthetic text data 54 in which sentences 52 are arranged in the order of conversation is generated. Each sentence 52 of the composite text data 54 is given an original identifier 55 indicating the original partial text data 51. In FIG. 6, the partial text data 51 on the left side of FIG. 4 is given an original identifier of “01”, and the partial text data 51 on the right side of FIG. 4 is given an original identifier of “02”.

(Step S5 in FIG. 3: Time series similarity calculation process)
The similarity calculation unit 25 sets each of the plurality of sentences 52 included in the synthetic text data 54 generated in step S4 as the target sentence, and sets at least one of the sentences 52 before and after the target sentence as the partner sentence. .. In the first embodiment, when there are sentences 52 before and after the target sentence, the similarity calculation unit 25 sets the sentence 52 before and after the target sentence 52 as the other sentence, respectively.
Then, the similarity calculation unit 25 calculates the similarity between the word extracted from the target sentence in step S2 and the word extracted from the partner sentence in step S2 as the time series similarity. When a plurality of words are extracted from at least one of the target sentence and the other sentence, the similarity calculation unit 25 uses each combination of the words extracted from the target sentence and the words extracted from the other sentence. Calculate the similarity for. Then, the similarity calculation unit 25 calculates the average value or the median of the similarity for each combination as the time series similarity.

As shown in FIG. 7, it is assumed that the words in the synthetic text data 54 are extracted. In FIG. 7, it is assumed that the underlined word is extracted in step S2. Then, for example, when the sentence 52 of ID001 is the target sentence, the sentence 52 of ID002 is set as the partner sentence. When the sentence 52 of ID002 is the target sentence, the sentence 52 of ID001 and the sentence 52 of ID003 are set as the other sentence, respectively.
Then, when the sentence 52 of ID001 is the target sentence, the time series similarity with the sentence 52 of ID002 which is the partner sentence is calculated as follows. From the sentence 52 of ID001, "A" and "care" are extracted as words. Further, from the sentence 52 of ID002, "B" and "care" are extracted as words. Therefore, the combination of "A" and "B", the combination of "A" and "care", the combination of "care" and "B", and the combination of "care" and "care" The similarity for each of the four combinations is calculated. Then, the average value of the similarity for the four combinations is calculated as the time series similarity.

The similarity calculation unit 25 uses the expression vector 53 generated in step S3 for the word extracted from the target sentence and the expression vector 53 generated in step S3 for the word extracted from the other sentence, and the similarity calculation unit 25 uses the similarity degree. To calculate. Specifically, the similarity calculation unit 25 calculates the inner product of the expression vector 53 for the word extracted from the target sentence and the expression vector 53 for the word extracted from the partner sentence as the similarity.
As a specific example, a method of calculating the degree of similarity between the word "meeting" and the word "meeting" will be described. As shown in FIG. 5, the element of the expression vector 53 of the word “meeting” is {0.3, 0.2, 0.6}. The element of the expression vector 53 of the word "meeting" is {0.5, 0.1, 0.8}. Therefore, the similarity is 0.3 × 0.5 + 0.2 × 0.1 + 0.6 × 0.8 = 0.65.

(Step S6 in FIG. 3: Candidate identification process)
The candidate identification unit 26 sets each of the plurality of sentences 52 included in the synthetic text data 54 generated in step S4 as the target sentence, and whether the time series similarity calculated in step S5 for the target sentence is lower than the first threshold value. Judge whether or not. The candidate identification unit 26 identifies the target sentence as a correction candidate when the time series similarity is lower than the first threshold value. When the candidate identification unit 26 has sentences 52 before and after the target sentence, both the time-series similarity with the previous sentence 52 and the time-series similarity with the subsequent sentence 52 are the first. If it is lower than the threshold value, the target sentence is specified as a correction candidate.
As shown in FIG. 8, the time series similarity is calculated for each sentence 52, and when the first threshold value is 0.6, the sentence 52 of ID004 is specified as a correction candidate.

*** Effect of Embodiment 1 ***
As described above, the modification candidate identification device 10 according to the first embodiment calculates the similarity between sentences 52 in chronological order, and identifies the sentence 52 having a low similarity as a modification candidate. Thereby, the sentence 52 as a correction candidate can be specified without obtaining the appearance frequency of all the words included in the synthetic text data 54. As a result, it becomes possible to appropriately identify correction candidates while suppressing the calculation load.

When a plurality of people have a conversation, it is assumed that they speak the content related to the sentence 52 before or after. Therefore, by calculating the similarity between the sentences 52 in chronological order and specifying the sentence 52 having a low similarity as the correction candidate, the correction candidate can be appropriately specified.

In the first embodiment, the voice collecting device 41 is provided for each person. In this case, it is difficult to record the voices of a plurality of people in an overlapping manner, so that the accuracy of voice recognition is high. On the other hand, since the characters are converted without considering the contents spoken by other people, erroneous conversion to homonyms is likely to occur. For example, the sentence 52 of ID003 in FIG. 8 and the sentence 52 included in the partial text data 51 different from the sentence 52 of ID004. Therefore, when the sentence 52 of ID004 is generated, the character conversion is performed without considering the sentence 52 of ID003. As a result, erroneous conversion of "candy" instead of "rain" is likely to occur.
The correction candidate identification device 10 according to the first embodiment calculates the similarity between the sentences 52 in time series, and the correction candidates are specified from the relationship with the preceding and following sentences 52 in the other partial text data 51. Therefore, it is particularly effective when the voice collecting device 41 is provided for each person.

*** Other configurations ***
<Modification example 1>
In the first embodiment, the similarity calculation unit 25 calculates the inner product of the expression vector 53 as the time series similarity. However, the similarity calculation unit 25 may calculate the expression vector 53 for the word extracted from the target sentence, the expression vector 53 for the word extracted from the other sentence, and the cosine similarity as the time series similarity. ..

<Modification 2>
In step S3 of FIG. 3, the vector for some words is not stored, and the expression vector 53 may not be generated. When there is a word for which the expression vector 53 cannot be generated, the sentence 52 including the word may be specified as the sentence 52 for which the expression vector 53 cannot be generated.
A word for which the expression vector 53 cannot be generated may be a misrecognized word. Therefore, the candidate to be appropriately corrected can be specified by specifying the word 52 in which the expression vector 53 cannot be generated as the sentence 52.

At this time, the correction candidate identified in step S6 of FIG. 3 and the sentence 52 containing a word for which the expression vector 53 cannot be generated may be presented separately.

<Modification example 3>
In the first embodiment, at least one sentence 52 before and after the target sentence is set as the partner sentence. As the partner sentence, not only the sentence 52 immediately before or after the target sentence but also a plurality of sentences 52 may be set as the partner sentence. For example, a sentence 52 that precedes the target sentence and has a source identifier 55 different from the target sentence may be used as the partner sentence. Further, a sentence 52 that is after the target sentence and has a source identifier 55 different from the target sentence may be used as the partner sentence.

<Modification example 4>
In the first embodiment, each functional component is realized by software. However, as a modification 4, each functional component may be realized by hardware. The difference between the modified example 4 and the first embodiment will be described.

With reference to FIG. 9, the configuration of the modification candidate identification device 10 according to the modification 4 will be described.
When each functional component is realized by hardware, the modification candidate identification device 10 includes an electronic circuit 15 instead of the processor 11, the memory 12, and the storage 13. The electronic circuit 15 is a dedicated circuit that realizes the functions of each functional component, the memory 12, and the storage 13.

Examples of the electronic circuit 15 include a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, a logic IC, a GA (Gate Array), an ASIC (Application Specific Integrated Circuit), and an FPGA (Field-Programmable Gate Array). is assumed.
Each functional component may be realized by one electronic circuit 15, or each functional component may be distributed and realized by a plurality of electronic circuits 15.

<Modification 5>
As a modification 5, some functional components may be realized by hardware, and other functional components may be realized by software.

The processor 11, the memory 12, the storage 13, and the electronic circuit 15 are referred to as processing circuits. That is, the function of each functional component is realized by the processing circuit.

Embodiment 2.
The second embodiment is different from the first embodiment in that the parts to be corrected are further narrowed down from the correction candidates. In the second embodiment, these different points will be described, and the same points will be omitted.

*** Explanation of operation ***
The operation of the modification candidate identification device 10 according to the second embodiment will be described with reference to FIGS. 10 and 11.
The operation procedure of the modification candidate identification device 10 according to the second embodiment corresponds to the modification candidate identification method according to the second embodiment. Further, the program that realizes the operation of the modification candidate identification device 10 according to the second embodiment corresponds to the modification candidate identification program according to the second embodiment.

The processing of steps S1 to S6 is the same as that of the first embodiment.

(Step S7 in FIG. 10: Combination similarity calculation process)
The similarity calculation unit 25 sets each correction candidate specified in step S6 as a target correction candidate. The similarity calculation unit 25 uses a sentence 52 other than the target correction candidate as a comparison sentence, and combines the similarity between the words extracted from the target correction candidates and each word extracted from each of the comparison sentences, and the similarity degree. Calculate as. When a plurality of words are extracted from the target correction candidates, the similarity calculation unit 25 combines the similarity between each word extracted from the target correction candidates and each word extracted from each of the comparative sentences. Calculate as similarity.
As shown in FIG. 11, it is assumed that the sentence 52 of ID004 is specified as a correction candidate. In this case, the similarity calculation unit 25 sets each sentence 52 of ID001, ID002, ID003, ID005, and ID006 as a comparison sentence. Then, for each of the words "recent" and "candy" extracted from the sentence 52 of ID004, the similarity with each word extracted from the comparative sentence is calculated as a combination similarity. As in step S5, the similarity calculation unit 25 calculates the similarity using the expression vector 53 for the words extracted from the correction candidates and the expression vector 53 for the words extracted from the comparison target. To do.

(Step S8 of FIG. 10: Correction target identification process)
The candidate identification unit 26 sets each word extracted from the correction candidates as a target word. The candidate identification unit 26 determines whether or not the combination similarity with the target word calculated in step S7 is lower than the second threshold value for each word extracted from the comparative sentence. The candidate identification unit 26 specifies the word extracted from the correction candidates as the correction target when the combination similarity with the target word is lower than the second threshold value for all the words extracted from the comparative sentence.

In the case of FIG. 11, “recent” has a higher similarity to “tomorrow” in the sentence 52 of ID003 than the second threshold value. On the other hand, "candy" has a similarity with all words lower than the second threshold value. As a result, the "candy" in the sentence 52 of ID004 is specified as a correction target.

*** Effect of Embodiment 2 ***
As described above, the correction candidate identification device 10 according to the second embodiment further narrows down the parts to be corrected from the correction candidates. As a result, it is possible to reduce the labor involved in the correction work as compared with the first embodiment.

For example, when the sentence 52 having the content related to the sentence 52 before the immediately preceding sentence 52 is a correction candidate, the word included in the correction candidate is the word of the sentence 52 before the immediately preceding sentence 52. May be similar to. In the second embodiment, the combination similarity for all the words is calculated with respect to the correction candidate sentence 52, using all the other sentences 52 as comparative sentences.
As described above, the correction candidate identification device 10 according to the second embodiment calculates the combination similarity with the entire sentence 52 for the correction candidate sentence 52, and then identifies the correction candidate with higher accuracy. It also has the effect of being able to.

*** Other configurations ***
<Modification 6>
In step S8 of FIG. 10, the candidate identification unit 26 specified, for all the words extracted from the comparative sentence, the words whose combination similarity with the target words was lower than the second threshold value as the correction target. When the candidate identification unit 26 calculates the combination similarity with the target word for all the words extracted from the comparative sentence, the correction candidate sets the word whose combination similarity with any word is higher than the second threshold value. It may act to exclude. In this case, the words that are not excluded are candidates for correction.
Further, the candidate identification unit 26 calculates the combination similarity with the target word for all the words extracted from the comparative sentence, and specifies the maximum value of the combination similarity for each target word. It may operate so as to specify the one having the lowest maximum value of the combination similarity as a correction candidate. Alternatively, it may operate so as to exclude a certain amount from the correction candidates in descending order of the maximum value of the combination similarity.

<Modification 7>
In the second embodiment, the similarity calculation unit 25 uses sentences 52 other than the target correction candidates as comparative sentences. However, the similarity calculation unit 25 may use only the sentence 52 having the original identifier 55 different from the target correction candidate among the sentences 52 other than the target correction candidate as the comparison sentence.
As a result, only the sentence 52 generated from the voice data uttered by a person other than the person who uttered the voice data that is the source of the target correction candidate becomes the comparison sentence. Some people have a habit of intonation of words. Due to this, the sentence 52 generated from the voice data spoken by the same person may contain the same error. For example, when a person speaks "Ame (rain)", when a person speaks, it may be mistakenly recognized as "candy" instead of "rain".
Therefore, if the other sentence 52 of the person who uttered the voice data that is the source of the target correction candidate is used as the comparison sentence, the same error is included in the comparison sentence, and there is a possibility that the correction target cannot be properly specified. There is. On the other hand, if the method according to the modification 7 is used, the correction target can be appropriately specified.

The embodiments and modifications of the present invention have been described above. Some of these embodiments and modifications may be combined and carried out. In addition, any one or several may be partially carried out. The present invention is not limited to the above embodiments and modifications, and various modifications can be made as needed.

10 Correction candidate identification device, 11 Processor, 12 Memory, 13 Storage, 14 Communication interface, 15 Electronic circuit, 21 Data acquisition unit, 22 Word extraction unit, 23 Vector generation unit, 24 Text synthesis unit, 25 Similarity calculation unit, 26 Candidate identification unit, 41 voice collector, 42 voice recognition device, 51 partial text data, 52 sentences, 53 expression vectors, 54 synthetic text data, 55 original identifiers.

Claims (8)

A word extractor that extracts words contained in text data composed of multiple sentences, and a word extractor
The words extracted from the target sentence by the word extraction unit and the word extracted from the target sentence by the word extraction unit and the word extracted from the target sentence by the word extraction unit, with each of the plurality of sentences as the target sentence and at least one of the sentences before and after the target sentence as the partner sentence. A similarity calculation unit that calculates the similarity with words extracted from the other sentence as a time-series similarity,
A candidate identification unit that specifies the target sentence as a correction candidate when the time-series similarity calculated by the similarity calculation unit for the target sentence is lower than the first threshold value, with each of the plurality of sentences as the target sentence. Correction candidate identification device including. The similarity calculation unit uses sentences other than the correction candidates specified by the candidate identification unit as comparison sentences, and the similarity between the words extracted from the correction candidates and the words extracted from each of the comparison sentences. Calculate the degree as a combination similarity and
The candidate identification unit according to claim 1, wherein when the combination similarity of any word extracted from the comparative sentence is lower than the second threshold value, the word extracted from the correction candidate is specified as a correction target. Correction candidate identification device. The modification candidate identification device according to claim 2, wherein the text data is generated by synthesizing partial text data obtained by converting voice data collected by each of a plurality of voice collecting devices. The correction candidate identification device according to claim 3, wherein the similarity calculation unit calculates the combination similarity using a sentence included in the partial text data other than the partial text data including the correction candidate as the comparison sentence. The correction candidate identification device further
A vector generation unit for generating an expression vector representing the meaning of the target word is provided with the word extracted by the word extraction unit as the target word.
The modification candidate identification device according to any one of claims 1 to 4, wherein the similarity calculation unit calculates the similarity using the expression vector generated by the vector generation unit. The similarity calculation unit according to claim 5 calculates the inner product of the expression vector for a word extracted from one sentence and the expression vector for a word extracted from the other sentence as the similarity. Modification candidate identification device. The word extraction unit extracts words contained in text data composed of multiple sentences,
The similarity calculation unit extracts each of the plurality of sentences as a target sentence, sets at least one of the sentences before and after the target sentence as the other sentence, and extracts the words extracted from the target sentence and the other sentence. Calculate the similarity with the word as the time series similarity,
A method for identifying a correction candidate in which the candidate identification unit specifies the target sentence as a correction candidate when the time-series similarity calculated for the target sentence is lower than the first threshold value, with each of the plurality of sentences as the target sentence. .. Word extraction processing that extracts words contained in text data composed of multiple sentences, and
Each of the plurality of sentences is a target sentence, and at least one of the sentences before and after the target sentence is set as a partner sentence, and the word extracted from the target sentence by the word extraction process and the word extracted by the word extraction process are described. Similarity calculation processing that calculates the similarity with the word extracted from the other sentence as the time series similarity,
Candidate identification processing for specifying the target sentence as a correction candidate when the time-series similarity calculated by the similarity calculation process for the target sentence is lower than the first threshold value, with each of the plurality of sentences as the target sentence. A modification candidate identification program that makes the computer function as a modification candidate identification device.

Images