KR102374405B1 - Apparatus and method for post-processing speech recognition in artificial intelligence interview - Google Patents

Abstract

The present invention relates to a method and a device for post-processing speech recognition in an artificial intelligence (AI) interview environment, which includes the following steps of: obtaining interview data on an AI interview target and generating a word vector space based on the interview data; obtaining a speech recognition result text on the AI interview target and verifying the non-fluency of the speech recognition result text based on the word vector space; defining the non-fluency part of the speech recognition result text as a correction section and generating a correction vocabulary candidate group corresponding to the correction section based on the word vector space; verifying the structural similarity and contextual suitability of words belonging to the correction vocabulary candidate group and selecting a final correction word candidate by calculating a similarity index through selective weighting therebetween; determining whether the selected final correction word candidate is suitable; and, if the selected final correction word candidate is suitable, replacing the correction section in the speech recognition result text with the final correction word candidate and outputting the candidate. The present invention is a technique developed through the "Commercialization of AI Recruitment Evaluation System for Non-face-to-face Recruitment Environment Improvement" of the 2020 AI Technology Commercialization Support Project (CY201042) of the Seoul Business Agency in Seoul.

Description

Speech recognition post-processing apparatus and method in AI interview environment

The present invention relates to a voice recognition post-processing method, and more particularly, to a voice recognition post-processing apparatus and method in an AI interview environment.

In general, in post-processing of speech recognition, detection of non-fluency is a technique used to correct or delete an erroneously recognized part of content uttered by a speaker.

As for the technology used for non-fluency detection, both the recently used data-based supervised learning method and the existing rule-based method can be used.

However, the data-based supervised learning technique had a problem that separate data for non-fluency detection had to be prepared and sufficient performance verification for Korean was not performed.

In addition, the rule-based method also had a problem in that it takes considerable human resources and time to define rules in consideration of both the complicated pronunciation and phonological fluctuations of Korean.

In addition, in the case of the existing technology, vocabulary correction was performed by separately generating a pronunciation dictionary database, etc. and measuring structural similarity between words based on this.

In this case, the speech recognition error correction result has a high dependence on the performance of the constructed pronunciation dictionary database, and the construction of the pronunciation dictionary corresponding to the number of structural changes due to the language characteristics of Korean is accompanied by a high construction cost. may exist.

In addition, Korean is a language that has quite disadvantageous conditions in the field of natural language processing due to its inherent linguistic characteristics (ambiguity, agglutinative language, etc.).

Overseas natural language processing-related technologies conducted for English are often difficult to apply to actual Korean and expect the same level of performance, and there is a need to selectively apply major natural language processing techniques.

Recently, the hiring culture is one of the things that have changed due to Corona, and it is often an untact interview where the interviewer and interviewer use a video conference program, etc., an online PT interview in which PT is recorded and submitted in advance, and the AI interviewer’s There is an AI interview with several questions and answers according to the procedure.

Among these, AI interview is decided by machine rather than human, and all interactions, including voice, between questions and answers are reflected in the interview results.

Therefore, it is an essential task to accurately grasp the voice of the interviewer in various environments in the AI interview, but it is necessary to supplement the voice recognition result in the AI interview environment due to the limitation of voice recognition.

Therefore, in the future, in the AI interview environment, the development of a voice recognition post-processing method capable of providing a voice recognition result with improved accuracy and reliability by supplementing the voice recognition result with minimum cost and minimum time is required.

The present invention is a technology developed through the Seoul Industry Promotion Agency's 2020 Artificial Intelligence (AI) Technology Commercialization Support Project (CY201042) "Commercialization of AI Recruitment Evaluation System for Non-face-to-face Recruitment Environment Improvement".

Republic of Korea Patent Publication No. 10-2018-0062003 (2018. 06. 08)

One object of the present invention to solve the above problems is to generate a word vector space based on interview data, and based on this, the structural similarity and context of the proofreading vocabulary candidate group corresponding to the proofreading section of the speech recognition result text Speech recognition in an AI interview environment that can provide speech recognition results with improved accuracy and reliability by supplementing speech recognition results with minimum cost and minimum time in an AI interview environment by verifying appropriate suitability and selecting final proofreading word candidates To provide a post-processing apparatus and method.

Another object of the present invention is to provide a voice recognition post-processing apparatus and method for post-processing an interviewer's speech processing text, which has undergone speech recognition in an AI interview environment, using a word vector space generated by unsupervised learning-based technology, etc.

The problems to be solved by the present invention are not limited to the problems mentioned above, and other problems not mentioned will be clearly understood by those skilled in the art from the following description.

A voice recognition post-processing method according to an embodiment of the present invention for solving the above-described problems is a voice recognition post-processing method of a voice recognition post-processing apparatus in an artificial intelligence (AI) interview environment, and interview data of an AI interviewee generating a word vector space based on interview material data by acquiring defining the non-fluency part of the speech recognition result text as a proofreading section, generating a proofreading vocabulary candidate group corresponding to the proofreading section based on the word vector space; Selecting a final proofreading word candidate by verifying suitability and calculating a similarity index through selective weighting between them, determining whether the selected final proofreading word candidate is appropriate, and the selected final proofreading word candidate If appropriate, the correction section of the speech recognition result text is replaced with the final corrected word candidate and output.

In an embodiment, the generating of the word vector space is characterized in that the word vector space is generated based on self-introduction data among the interview data of the AI interviewee.

In an embodiment, the step of verifying the non-fluency of the speech recognition result text comprises: when the speech recognition result text is obtained, the speech recognition result text is divided into morpheme units, part-of-speech division is performed, and then, the speech recognition result text is stored in the word vector space. It is characterized in that a portion having non-fluency is detected in the speech recognition result text by mapping, and the corresponding portion is determined as a correction section.

In an embodiment, the generating of the proofreading vocabulary candidate group includes defining a non-fluent portion of the speech recognition result text as a proofreading section, dividing the proofreading section into morpheme units, and using the morpheme unit within the word vector space. It is characterized in that the group of words appearing together with other parts separated by .

In an embodiment, the verifying of the structural similarity of the words belonging to the proofreading vocabulary candidate group includes setting the minimum number of operations required for mutual conversion between the words in the proofreading section and the words belonging to the proofreading vocabulary candidate group as the editing distance between words. It is characterized by defining, calculating a minimum editing distance between phonemes and syllables, and verifying structural similarity of words belonging to the proofreading vocabulary candidate group based on the calculated minimum editing distance.

In an embodiment, the verifying of the contextual suitability of the words belonging to the proofreading vocabulary candidate group comprises dictating whether a contextually appropriate sentence is completed when words belonging to the proofreading vocabulary candidate group are replaced with a word in the proofreading section. And it is characterized in that it is verified from the viewpoint of the interviewer and the domain, including the technology.

In an embodiment, the step of selecting the final proofreading word candidate by calculating the similarity index may include quantifying the structural similarity and contextual relevance of the words of the proofreading vocabulary candidate group as values between 0 and 1, respectively, and based on this, the final proofreading word candidate is calculated. It is characterized in that a similarity index for extracting proofreading word candidates is calculated.

In an embodiment, the calculating of the similarity index comprises calculating the similarity index step by step while selectively adjusting each weight when the sum of the weights of the structural similarity and the contextual similarity is 1, It is characterized in that the TOP N words are extracted from the similarity indices calculated in each step.

In an embodiment, the calculating of the similarity index comprises calculating the similarity index by a formula consisting of similarity index = (contextual relevance × context similarity weight) + (structural similarity × structure similarity weight). .

In an embodiment, the step of calculating the similarity index is characterized in that words given a high similarity index in another step among the extracted top N words for each step are determined as final proofreading word candidates.

In an embodiment, the determining whether the selected final proofreading word candidate is suitable may include outputting an original sentence of the speech recognition result text if the selected final proofreading word candidate is not appropriate.

A speech recognition post-processing device according to an embodiment of the present invention is a speech recognition post-processing device in an AI (Artificial Intelligence) interview environment. A word vector space generating unit that generates, a non-fluency verification unit that obtains the speech recognition result text of the AI interviewee and verifies the non-fluency of the speech recognition result text based on the word vector space, a metaphor among the speech recognition result text A proofreading vocabulary candidate group generating unit that defines a generation part as a proofreading section and generates a proofreading vocabulary candidate group corresponding to the proofreading section based on the word vector space, and structural similarity verification that verifies the structural similarity of words belonging to the proofreading vocabulary candidate group A final proofreading word by calculating a similarity index through a part, a contextual relevance verification unit that verifies the contextual relevance of words belonging to the proofreading vocabulary candidate group, and structural similarity and selective weighting of the contextual relevance of words belonging to the proofreading vocabulary candidate group A final proofreading word candidate selection unit that selects a candidate, a determination unit determining whether the selected final proofreading word candidate is appropriate, and if the selected final proofreading word candidate is suitable, a correction section of the speech recognition result text is selected as the final proofreading word candidate It characterized in that it comprises a result output unit for outputting by replacing the candidate correction word.

A computer program providing a voice recognition post-processing method according to another embodiment of the present invention for solving the above-described problems is stored in a medium in order to perform any one of the methods described above in combination with a computer that is hardware.

In addition to this, another method for implementing the present invention, another system, and a computer-readable recording medium for recording a computer program for executing the method may be further provided.

As described above, according to the present invention, a word vector space is generated based on interview data data, and the structural similarity and contextual suitability of the proofreading vocabulary candidate group corresponding to the proofreading section of the speech recognition result text are verified based on this to verify the final proofreading word. By selecting candidates, it is possible to provide speech recognition results with improved accuracy and reliability by supplementing speech recognition results with minimum cost and minimum time in an AI interview environment.

In other words, the present invention suggests that the word vector space generated by unsupervised learning can be used to detect the non-fluency of speech recognition results, and applies it to a non-face-to-face interview environment that is in high demand since the post-corona era began. can do.

As such, the present invention is a method of post-processing the interviewer's speech processing text, which has undergone speech recognition in an AI interview environment, using a word vector space generated by unsupervised learning-based technology, and the like.

Effects of the present invention are not limited to the effects mentioned above, and other effects not mentioned will be clearly understood by those skilled in the art from the following description.

1 is a block diagram illustrating a voice recognition post-processing apparatus according to an embodiment of the present invention.
2 to 5 are flowcharts for explaining a voice recognition post-processing method of the voice recognition post-processing apparatus according to an embodiment of the present invention.
6 is a diagram for explaining a formula for calculating a similarity index.

Advantages and features of the present invention and methods of achieving them will become apparent with reference to the embodiments described below in detail in conjunction with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below, but may be implemented in various different forms, and only the present embodiments allow the disclosure of the present invention to be complete, and those of ordinary skill in the art to which the present invention pertains. It is provided to fully understand the scope of the present invention to those skilled in the art, and the present invention is only defined by the scope of the claims.

The terminology used herein is for the purpose of describing the embodiments and is not intended to limit the present invention. In this specification, the singular also includes the plural unless specifically stated otherwise in the phrase. As used herein, “comprises” and/or “comprising” does not exclude the presence or addition of one or more other components in addition to the stated components. Like reference numerals refer to like elements throughout, and "and/or" includes each and every combination of one or more of the recited elements. Although "first", "second", etc. are used to describe various elements, these elements are not limited by these terms, of course. These terms are only used to distinguish one component from another. Accordingly, it goes without saying that the first component mentioned below may be the second component within the spirit of the present invention.

Unless otherwise defined, all terms (including technical and scientific terms) used herein will have the meaning commonly understood by those of ordinary skill in the art to which this invention belongs. In addition, terms defined in a commonly used dictionary are not to be interpreted ideally or excessively unless specifically defined explicitly.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings.

Before the description, the meaning of the terms used in this specification will be briefly described. However, it should be noted that, since the description of the term is for the purpose of helping the understanding of the present specification, it is not used in the meaning of limiting the technical idea of the present invention unless explicitly described as limiting the present invention.

1 is a block diagram illustrating a voice recognition post-processing apparatus according to an embodiment of the present invention.

1, the speech recognition post-processing apparatus in an AI interview environment according to the present invention includes a word vector space generation unit 110, a non-fluency verification unit 120, a corrected vocabulary candidate group generation unit 130, It may include a structural similarity verification unit 140 , a contextual suitability verification unit 150 , a final proofread word candidate selection unit 160 , a determination unit 170 , and a result output unit 180 .

Here, the word vector space generating unit 110 may acquire interview data of the AI interviewee and generate a word vector space based on the interview data data.

As an example, the word vector space generator 110 may generate a word vector space based on self-introduction data among interview materials of the AI interviewee, but this is only an example, and is not limited thereto.

In addition, the word vector space generating unit 110 may pre-process portions and stopwords including personal information in the interview data data, and tokenize the pre-processed interview data data to generate a word vector space.

Here, when tokenizing, the word vector space generating unit 110 may tokenize the interview data data that has been pre-processed through morpheme separation and part-of-speech classification.

Next, the non-fluency verification unit 120 may obtain the speech recognition result text of the AI interviewee and verify the non-fluency of the speech recognition result text based on the word vector space.

As an example, the non-fluency verification unit 120, upon obtaining the speech recognition result text, separates the speech recognition result text into morpheme units, performs part-of-speech classification, and maps the speech recognition result text to the word vector space to determine the non-fluency level in the speech recognition result text. It is possible to detect a portion having a portion and determine that portion as a calibration section.

Here, when the non-fluency verification unit 120 maps the speech recognition result text, which has undergone morpheme separation and part-of-speech classification, to the word vector space, the morpheme incorrectly recognized as a speech recognition error is not generally arranged in the vector space or By taking advantage of the fact that it has an Out-of-Vocabulary (OOV) problem, it is possible to detect parts with inflexibility.

Next, the proofreading vocabulary candidate group generating unit 130 may define a non-fluent part of the speech recognition result text as a proofreading section and generate a proofreading vocabulary candidate group corresponding to the proofreading section based on the word vector space.

Here, the proofreading vocabulary candidate group generating unit 130 defines a non-fluent part of the speech recognition result text as a calibration section, separates the proofreading section into morpheme units, and together with other parts separated by morpheme units in the word vector space. It is possible to generate a proofreading vocabulary candidate group by preferentially extracting the appearing word group and limiting the candidate group to the expected part-of-speech group of the proofreading section.

Next, the structural similarity verification unit 140 may verify the structural similarity of words belonging to the proofreading vocabulary candidate group.

Here, the structural similarity verification unit 140 defines the minimum number of operations required for mutual conversion between the words in the proofing section and the words belonging to the proofreading vocabulary candidate group as the editing distance between words, and the minimum editing distance between phonemes and syllables. and, based on the calculated minimum editing distance, it is possible to verify the structural similarity of words belonging to the proofreading vocabulary candidate group.

As an example, the structural similarity verification unit 140 performs correction based on the assumption that the editing distance between the words in the proofreading section having the same number of syllables and the words in the proofreading vocabulary candidate group and the structural similarity of the two words are inversely proportional to each other in consideration of the interview environment. Structural similarity of words belonging to the lexical candidate group can be verified.

In this case, the structural similarity verifying unit 140 may verify the structural similarity of words belonging to the proofreading vocabulary candidate group except for words including communicators, profanity, and abbreviations in consideration of the interview environment.

In addition, the contextual relevance verification unit 150 may verify the contextual relevance of words belonging to the proofreading vocabulary candidate group.

Here, the contextual relevance verification unit 150 verifies whether or not a contextually appropriate sentence is completed when words belonging to the proofreading vocabulary candidate group are replaced with words in the proofreading section from the interviewer's point of view and domain point of view, including oral and descriptive. can

As an example, the contextual relevance verification unit 150 calculates a contextual relevance index by assigning a selective weight to each verification item and performing verification through topic modeling and emotion analysis when verifying from the interviewer's point of view and the domain point of view. and the calculated contextual relevance index can be reflected.

For example, the verification item is the first verification item that verifies whether the content flow is similar to other sentences not included in the correction target among the matters described by the interviewer voice, A second verification item that verifies whether it proceeds with an emotional flow similar to other sentences not included, a third verification item that verifies whether the flow of content is similar to the self-introduction submitted by the interviewer, and emotional flow similar to the self-introduction submitted by the interviewer It may include a fourth verification item verifying whether the flow proceeds or not, and a fifth verification item verifying whether observation of a similar contextual flow within the self-introduction data is possible, which is only an example, but is not limited thereto.

Next, the final proofreading word candidate selection unit 160 may select the final proofreading word candidate by calculating a similarity index through selective weighting of structural similarity and contextual relevance of words belonging to the proofreading vocabulary candidate group.

Here, the final proofreading word candidate selection unit 160 quantifies the structural similarity and contextual relevance of the words in the proofreading vocabulary candidate group as values between 0 and 1, respectively, and calculates a similarity index for extracting the final proofreading word candidate based on this value can do.

For example, when calculating the similarity index, the final proofreading word candidate selection unit 160 selectively adjusts each weight when the sum of the structural similarity weight and the contextual similarity weight is 1, and the similarity index step by step , and top N words can be extracted from the similarity indices calculated step by step.

For example, when calculating the similarity index, the final proofreading word candidate selection unit 160 calculates the similarity index by using a formula consisting of similarity index = (contextual relevance × context similarity weight) + (structural similarity × structure similarity weight). can be calculated.

In addition, the final proofreading word candidate selection unit 160 may determine, as final proofreading word candidates, words given a high similarity index in another step among the extracted top N words for each step.

Next, the determination unit 170 may determine whether the selected final proofreading word candidate is appropriate.

Here, the determination unit 170 may determine whether the final proofreading word candidate selected based on the proofreading decision index for determining whether to proofread is appropriate.

Then, if the selected final proofreading word candidate is suitable, the result output unit 180 may replace the proofreading section of the speech recognition result text with the final proofreading word candidate and output the same.

Also, the result output unit 180 may output the original sentence of the speech recognition result text if the selected final proofreading word candidate is not suitable.

As described above, the present invention generates a word vector space based on interview data, and verifies the structural similarity and contextual suitability of the proofreading vocabulary candidate group corresponding to the proofreading section of the speech recognition result text based on the result of the final proofreading word candidate. By selecting , it is possible to provide speech recognition results with improved accuracy and reliability by supplementing speech recognition results with minimum cost and minimum time in an AI interview environment.

In other words, the present invention suggests that the word vector space generated by unsupervised learning can be used to detect the non-fluency of speech recognition results, and applies it to a non-face-to-face interview environment that is in high demand since the post-corona era began. can do.

As such, the present invention is a method of post-processing the interviewer's speech processing text, which has undergone speech recognition in an AI interview environment, using a word vector space generated by unsupervised learning-based technology, and the like.

2 to 5 are flowcharts for explaining a voice recognition post-processing method of a voice recognition post-processing apparatus according to an embodiment of the present invention, and FIG. 6 is a diagram for explaining a formula for calculating a similarity index.

2 to 5 , the present invention may generate a word vector space based on the interview data by acquiring interview data of an AI interviewee (S10).

Here, the present invention can create a word vector space based on self-introduction data among the interview data of the AI interviewee. In the interview data data, parts and stop words containing personal information are pre-processed, and the pre-processing is completed. A word vector space can be created by tokenizing data through morpheme separation and part-of-speech classification.

Next, according to the present invention, it is possible to obtain the speech recognition result text of the AI interviewee and verify the non-fluency of the speech recognition result text based on the word vector space (S20).

Here, as shown in FIG. 3 , in the present invention, when the speech recognition result text is obtained (S22), the speech recognition result text is separated into morpheme units and part-of-speech classification is performed (S24), and then mapped to a word vector space for speech recognition A portion having non-fluency in the resulting text may be detected (S26), and the corresponding portion may be determined as a correction section (S28).

As an example, according to the present invention, when a speech recognition result text subjected to morpheme separation and part-of-speech classification is mapped to a word vector space, a morpheme incorrectly recognized as a speech recognition error is not generally arranged in the vector space or OOV (Out-of-Speech) By taking advantage of the fact that it has the of-Vocabulary problem, it is possible to detect parts with non-fluency.

Next, according to the present invention, the non-fluent part of the speech recognition result text may be defined as a proofreading section, and a proofreading vocabulary candidate group corresponding to the proofreading section may be generated based on the word vector space ( S30 ).

Here, as shown in FIG. 4, in the present invention, the non-fluent portion of the speech recognition result text is defined as a correction section (S32), the correction section is separated by morpheme units (S34), and separated into morpheme units in the word vector space. A group of words appearing together with other parts of the subject may be preferentially extracted (S36), and the candidate group may be limited to an expected part-of-speech group of the proofreading section to generate a proofreading vocabulary candidate group (S38).

And, according to the present invention, the final proofreading word candidate can be selected by verifying structural similarity and contextual suitability of words belonging to the proofreading vocabulary candidate group, and calculating a similarity index through selective weighting between them (S40).

That is, in the present invention, as shown in FIG. 5 , the structural similarity of words belonging to the proofreading vocabulary candidate group is verified (S42), the contextual suitability of the words belonging to the proofreading vocabulary candidate group is verified (S44), and selective weighting is applied therebetween. A similarity index is calculated through (S46), and a final proofreading word candidate can be selected based on the similarity radix (S48).

For example, in the present invention, when calculating the similarity index, when the sum of the weights of structural similarity and contextual similarity is 1, the similarity index is calculated step by step while selectively adjusting each weight, and calculated step by step Top N words can be extracted from the obtained similarity indices.

Here, as shown in FIG. 6 , in the present invention, the similarity index can be calculated by a formula consisting of similarity index = (contextual relevance × context similarity weight) + (structural similarity × structure similarity weight).

In addition, in the present invention, when verifying the structural similarity of words belonging to the proofreading vocabulary candidate group, the minimum number of operations required for mutual conversion between the words in the proofreading section and the words in the proofreading vocabulary candidate group is defined as the editing distance between words, The minimum editing distance between phonemes and syllables is calculated, and the structural similarity of words belonging to the proofreading vocabulary candidate group can be verified based on the calculated minimum editing distance.

At this time, in the present invention, in consideration of the interview environment, the editing distance of words in the proofreading section having the same number of syllables and words belonging to the proofreading vocabulary candidate group except for words including communicators, profanity, and abbreviations are inversely proportional to the structural similarity of the two words Structural similarity of words belonging to the proofreading vocabulary candidate group can be verified based on the assumption that

In addition, when verifying the contextual suitability of words belonging to the proofreading vocabulary candidate group, the present invention provides dictation and description of whether a contextually appropriate sentence is completed when words belonging to the proofreading vocabulary candidate group are replaced with a word in the proofreading section. It can be verified from the viewpoint of the interviewer including the domain and the viewpoint of the domain.

In addition, the present invention performs verification through topic modeling and emotion analysis when verifying from the interviewer's point of view and from the domain point of view, assigning an optional weight to each verification item to calculate a contextual relevance index, and calculated contextual relevance index can be reflected.

In addition, the present invention quantifies the structural similarity and contextual relevance of the words of the proofreading vocabulary candidate group as values between 0 and 1, respectively, when selecting the final proofreading word candidate by calculating the similarity index, and based on this, the final proofreading word candidate A similarity index for extraction can be calculated.

Here, in the present invention, when calculating the similarity index, when the sum of the weights of structural similarity and contextual similarity is 1, the similarity index is calculated step by step while selectively adjusting each weight, and the calculated similarity index is calculated step by step. Top N words are extracted from the similarity indices, and words given a high similarity index in another step among the extracted top N words for each stage may be determined as final proofreading word candidates.

Next, the present invention may determine whether the selected final proofreading word candidate is appropriate (S50).

Next, according to the present invention, if the selected final proofreading word candidate is suitable, the calibration section of the speech recognition result text may be replaced with the final proofreading word candidate and outputted (S60).

Also, in the present invention, if the selected final proofreading word candidate is not suitable, the original sentence of the speech recognition result text may be output (S70).

As described above, the present invention generates a word vector space based on interview data, and verifies the structural similarity and contextual suitability of the proofreading vocabulary candidate group corresponding to the proofreading section of the speech recognition result text based on the result of the final proofreading word candidate. By selecting , it is possible to provide speech recognition results with improved accuracy and reliability by supplementing speech recognition results with minimum cost and minimum time in an AI interview environment.

In other words, the present invention suggests that the word vector space generated by unsupervised learning can be used to detect the non-fluency of speech recognition results, and applies it to a non-face-to-face interview environment that is in high demand since the post-corona era began. can do.

As such, the present invention is a method of post-processing the interviewer's speech processing text, which has undergone speech recognition in an AI interview environment, using a word vector space generated by unsupervised learning-based technology, and the like.

The method according to an embodiment of the present invention described above may be implemented as a program (or application) to be executed in combination with a server, which is hardware, and stored in a medium.

The above-described program is C, C++, JAVA, machine language, etc. that a processor (CPU) of the computer can read through a device interface of the computer in order for the computer to read the program and execute the methods implemented as a program It may include code (Code) coded in the computer language of Such code may include functional code related to a function defining functions necessary for executing the methods, etc., and includes an execution procedure related control code necessary for the processor of the computer to execute the functions according to a predetermined procedure. can do. In addition, the code may further include additional information necessary for the processor of the computer to execute the functions or code related to memory reference for which location (address address) in the internal or external memory of the computer to be referenced. there is. In addition, when the processor of the computer needs to communicate with any other computer or server located remotely in order to execute the above functions, the code uses the communication module of the computer to determine how to communicate with any other computer or server remotely. It may further include a communication-related code for whether to communicate and what information or media to transmit and receive during communication.

The storage medium is not a medium that stores data for a short moment, such as a register, a cache, a memory, etc., but a medium that stores data semi-permanently and can be read by a device. Specifically, examples of the storage medium include, but are not limited to, ROM, RAM, CD-ROM, magnetic tape, floppy disk, and an optical data storage device. That is, the program may be stored in various recording media on various servers accessible by the computer or in various recording media on the computer of the user. In addition, the medium may be distributed in a computer system connected to a network, and a computer-readable code may be stored in a distributed manner.

The steps of a method or algorithm described in relation to an embodiment of the present invention may be implemented directly in hardware, as a software module executed by hardware, or by a combination thereof. A software module may contain random access memory (RAM), read only memory (ROM), erasable programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, hard disk, removable disk, CD-ROM, or It may reside in any type of computer-readable recording medium well known in the art to which the present invention pertains.

As mentioned above, although embodiments of the present invention have been described with reference to the accompanying drawings, those skilled in the art to which the present invention pertains know that the present invention may be embodied in other specific forms without changing the technical spirit or essential features thereof. you will be able to understand Therefore, it should be understood that the embodiments described above are illustrative in all respects and not restrictive.

Claims (10)

In the voice recognition post-processing method of a voice recognition post-processing device in an AI (Artificial Intelligence) interview environment,
acquiring the interview data of the AI interviewee and generating a word vector space based on the interview data;
obtaining the speech recognition result text of the AI interviewee and verifying the non-fluency of the speech recognition result text based on the word vector space;
defining a non-fluent part of the speech recognition result text as a correction section, and generating a proofreading vocabulary candidate group corresponding to the correction section based on the word vector space;
selecting a final proofreading word candidate by verifying structural similarity and contextual suitability of words belonging to the proofreading vocabulary candidate group, and calculating a similarity index through selective weighting between them;
determining whether the selected final proofreading word candidate is appropriate; and
If the selected final proofreading word candidate is suitable, replacing the calibration section of the speech recognition result text with the final proofreading word candidate and outputting it;
The step of calculating the similarity index and selecting a final proofreading word candidate includes:
A method for post-processing speech recognition, characterized in that the structural similarity and contextual relevance of the words of the proofreading vocabulary candidate group are digitized as values between 0 and 1, respectively, and a similarity index for extracting the final proofreading word candidate is calculated based on this value. According to claim 1,
The step of generating the word vector space comprises:
Speech recognition post-processing method, characterized in that generating a word vector space based on self-introduction data among the interview data of the AI interviewee. 3. The method of claim 2,
The step of verifying the non-fluency of the speech recognition result text includes:
When the speech recognition result text is obtained, the speech recognition result text is separated into morpheme units, part-of-speech division is performed, and then mapped to the word vector space to detect a portion having non-fluency in the speech recognition result text, and the corresponding part Speech recognition post-processing method, characterized in that it is determined as a calibration section. 4. The method of claim 3,
The step of generating the proofreading vocabulary candidate group includes:
A non-fluent part of the speech recognition result text is defined as a correction section, the correction section is divided into morpheme units, and a group of words appearing together with other parts separated by the morpheme unit in the word vector space is preferentially extracted. and limiting the candidate group to the expected part-of-speech group of the proofreading section to generate the proofreading vocabulary candidate group. According to claim 1,
The step of verifying the structural similarity of words belonging to the proofreading vocabulary candidate group includes:
The minimum number of operations required for mutual conversion between the word in the proofreading section and the words belonging to the proofreading vocabulary candidate group is defined as the editing distance between words, the minimum editing distance between phonemes and syllables is calculated, and the calculated minimum editing distance based on the structural similarity of words belonging to the proofreading vocabulary candidate group. 6. The method of claim 5,
The step of verifying the contextual relevance of words belonging to the proofreading vocabulary candidate group includes:
Speech recognition post-processing method, characterized in that when words belonging to the proofreading vocabulary candidate group are replaced with words in the proofreading section, whether a contextually appropriate sentence is completed from the point of view of the interviewer including the oral and description and the domain point of view. delete According to claim 1,
The step of determining the suitability of the selected final proofreading word candidate includes:
If the selected final proofreading word candidate is not suitable, the speech recognition post-processing method according to claim 1 , wherein the original sentence of the speech recognition result text is output. It is combined with a computer that is hardware, and stored in a computer-readable recording medium to execute the method of any one of claims 1, 2, 3, 4, 5, 6 and 8. program. In an AI (Artificial Intelligence) interview environment, a post-processing apparatus for speech recognition,
a word vector space generator that acquires interview data of an AI interviewee and generates a word vector space based on the interview data;
a non-fluency verification unit that obtains the speech recognition result text of the AI interviewee and verifies the non-fluency of the speech recognition result text based on the word vector space;
a proofreading vocabulary candidate group generating unit defining a non-fluent part of the speech recognition result text as a proofreading section and generating a proofreading vocabulary candidate group corresponding to the proofreading section based on the word vector space;
a structural similarity verification unit that verifies the structural similarity of words belonging to the proofreading vocabulary candidate group;
a contextual relevance verification unit for verifying the contextual relevance of words belonging to the proofreading vocabulary candidate group;
a final proofreading word candidate selector selecting a final proofreading word candidate by calculating a similarity index through selective weighting of structural similarity and contextual relevance of words belonging to the proofreading word candidate group;
a determination unit for determining whether the selected final proofreading word candidate is appropriate; And,
and a result output unit for replacing the correction section in the speech recognition result text with the final correction word candidate and outputting the selected final correction word candidate if the selected final correction word candidate is suitable;
The final proofreading word candidate selection unit,
The speech recognition post-processing apparatus of claim 1, wherein structural similarity and contextual relevance of the words of the proofreading vocabulary candidate group are digitized as values between 0 and 1, respectively, and a similarity index for extracting the final proofreading word candidate is calculated based on the numerical values.

Images