KR20230061001A - Apparatus and method for correcting text - Google Patents

Abstract

Disclosed are an apparatus and a method for correcting a document. In accordance with one embodiment, the apparatus for correcting a document includes: a shape reproduction correction module correcting a sentence to be corrected based on visual characteristics of the sentence to be corrected; a pronunciation reproduction correction module correcting the sentence to be corrected based on auditory characteristics of the sentence to be corrected; a composite reproduction correction module correcting the sentence to be corrected based on statistical characteristics of a language to which the sentence to be corrected belongs; and a determination module receiving the sentence to be corrected, and determining whether to correct the sentence to be corrected using at least one of the shape reproduction correction module, the pronunciation reproduction correction module, and the composite reproduction correction module. Therefore, the present invention is capable of effectively correcting an error in a text without user intervention.

Description

Document correction device and method {APPARATUS AND METHOD FOR CORRECTING TEXT}

The disclosed embodiments relate to techniques for correcting errors in documents.

Sentences written by humans may include mistakes or intentional typos or indirectly expressed words using morphological and phonetic characteristics of characters. In addition, even when text is obtained by OCRing a document, incorrect text is often extracted for various reasons. Texts generated in daily life and texts generated as a result of OCR often contain typos or are not refined, requiring a lot of time and money for proofreading.

In particular, the demand for text correction is increasing as the use of portable terminals such as smart phones, etc., which are relatively prone to typos, are used to write articles or use chatbots. In particular, it is important for chatbots to understand the user's utterance intent, because if there is a typo, the chatbot process may proceed in a completely different direction or not start at all because the user's intention is not correctly grasped.

Republic of Korea Patent Publication No. 10-2021-0023288 (2021.03.04)

The disclosed embodiments are intended to provide technical means for correcting errors in text composed of one or more sentences.

According to an exemplary embodiment, a shape regeneration correction module correcting the target sentence based on visual characteristics of the target sentence; a pronunciation regeneration correction module for correcting the target sentence based on the auditory characteristics of the target sentence; a complex regeneration correction module for correcting the sentence to be proofread based on the statistical characteristics of the language to which the sentence to be proofread belongs; and a determination module that receives the sentence to be corrected and determines whether to correct the sentence to be corrected by using at least one of the shape regeneration correction module, the pronunciation regeneration correction module, and the complex regeneration correction module. An orthodontic device is provided.

The shape regeneration correction module may convert the sentence to be corrected into an image and then recognize text in the converted image using optical character recognition.

The shape regeneration correction module sets the recognized text as a corrected sentence when the number of noncomplete characters included in the recognized text is smaller than the number of noncomplete characters included in the correction target sentence. can

The shape regeneration correction module compares the recognized text with the sentence to be corrected to identify a changed area, and if the frequency of occurrence of the changed area is higher than the frequency of occurrence of the area before the change, the recognized text is converted to the corrected sentence. can be set to

The pronunciation regeneration correction module generates voice data from the sentence to be corrected using a speech synthesis unit (Text-to-Speech), and converts the voice data into text using a speech recognition unit (Speech-To-Text). can be converted to

The pronunciation regeneration correction module sets the converted text as a corrected sentence when the number of noncomplete characters included in the converted text is smaller than the number of noncomplete characters included in the sentence to be corrected. can

The pronunciation regeneration correction module compares the converted text with the sentence to be corrected to identify a changed region, and when the frequency of occurrence of the changed region is higher than the frequency of occurrence of the region before change, the converted text is converted to the corrected sentence. can be set to

The compound regeneration correction module may divide the proofread target sentence into a plurality of morphemes, select one or more proofreading tokens from among the plurality of proofreading tokens, and generate one or more proofreading candidate tokens for each proofreading token.

When the part-of-speech of the divided morpheme corresponds to a verb, the complex regeneration correction module may convert the corresponding morpheme into a basic form.

The compound regeneration correction module may select a morpheme that is not searched in a preset dictionary database among the plurality of morphemes as the correction target token.

The compound regeneration correction module inputs the sentence to be corrected into a preset language model and predicts a token corresponding to a position of a morpheme not searched in the dictionary database, but the language model is not searched in the dictionary database. If a token corresponding to is not predicted, a morpheme not searched in the dictionary database may be selected as the correction target token.

The compound regeneration correction module inputs the correction target sentence into a preset language model to predict a token corresponding to the location of the correction target token, and N number (N is a natural number of 1 or more) predicted by the language model. A prediction token may be set as the remediation candidate token for the remediation target token.

The complex regeneration calibration module may add a token whose editing distance is within K (K is a natural number equal to or greater than 1) from the primary calibration target token to the calibration candidate token.

The complex regeneration correction module, when the token to be proofreaded corresponds to a term, generates one or more candidate base types from the token to be proofread, and an editing distance from each of the one or more candidate base types is within K (where K is a natural number of 1 or greater). In tokens may be added to the calibration candidate tokens.

The compound regeneration calibration module may exclude tokens that are not searched in a preset dictionary or corpus from among the calibration candidate tokens.

The complex regeneration correction module may, for each of the one or more correction candidate tokens, at least one of a Levenstein distance with the correction target token, a part-of-speech similarity with the correction target token, a word frequency in a corpus, and an idiom frequency in the corpus. Calculate , and select a token to replace the calibration target token from among the one or more calibration candidate tokens based on the calculated value.

The apparatus may further include an additional correction module for additionally correcting the first correction sentence obtained by correcting the correction target sentence using at least one of the shape regeneration correction module, the pronunciation regeneration correction module, and the complex regeneration correction module. there is.

The additional correction module may receive the first corrected sentence, output a second corrected sentence obtained by additionally correcting the first corrected sentence, and learn to match the second corrected sentence with a preset correct answer sentence.

According to another exemplary embodiment, a method performed on a computer includes receiving a sentence to be corrected; A shape regeneration correction module for correcting the correction target sentence based on the visual characteristics of the correction target sentence, a pronunciation regeneration correction module for correcting the correction target sentence based on the auditory characteristics of the correction target sentence, and the correction target sentence determining whether or not to proofread the sentence to be proofread by using one or more complex regeneration proofreading modules for proofreading the sentence to be proofreaded based on statistical characteristics of the language to which it belongs; and correcting the target sentence based on the determination result.

The correcting may further include correcting the first correction sentence obtained by correcting the sentence to be corrected using at least one of the shape regeneration correction module, the pronunciation regeneration correction module, and the complex regeneration correction module. there is.

According to the disclosed embodiments, text errors can be effectively corrected without user intervention by applying an optimal error correction module according to the characteristics of errors appearing in the text.

1 is a block diagram illustrating a document proofreading apparatus 100 according to an exemplary embodiment.
2 is an exemplary diagram for explaining a learning process of a discrimination model 112 according to an embodiment
3 is an exemplary diagram for explaining a process of discriminating a sentence requiring correction from a document in a learned discrimination model 112 according to an embodiment.
4 is a flowchart for explaining a process of correcting sentences in the shape regeneration correction module 114 according to an embodiment
5 is a flowchart for explaining a process of correcting a sentence in the pronunciation regeneration correction module 116 according to an embodiment
6 is a flowchart for explaining a process of correcting sentences in the complex regeneration correction module 118 according to an embodiment.
7 is an exemplary diagram for explaining a process of learning an additional calibration model 120 according to an embodiment.
8 is a flowchart for explaining a document proofreading method 800 according to an embodiment.
9 is a block diagram illustrating and describing a computing environment including a computing device suitable for use in example embodiments.

Hereinafter, specific embodiments of the present invention will be described with reference to the drawings. The detailed descriptions that follow are provided to provide a comprehensive understanding of the methods, devices and/or systems described herein. However, this is only an example and the present invention is not limited thereto.

In describing the embodiments of the present invention, if it is determined that the detailed description of the known technology related to the present invention may unnecessarily obscure the subject matter of the present invention, the detailed description will be omitted. In addition, terms to be described later are terms defined in consideration of functions in the present invention, which may vary according to the intention or custom of a user or operator. Therefore, the definition should be made based on the contents throughout this specification. Terminology used in the detailed description is only for describing the embodiments of the present invention and should in no way be limiting. Unless expressly used otherwise, singular forms of expression include plural forms. In this description, expressions such as "comprising" or "comprising" are intended to indicate any characteristic, number, step, operation, element, portion or combination thereof, one or more other than those described. It should not be construed to exclude the existence or possibility of any other feature, number, step, operation, element, part or combination thereof.

1 is a block diagram illustrating a document proofreading apparatus 100 according to an exemplary embodiment. The document proofreading apparatus 100 according to an embodiment is a device for receiving a document composed of one or more sentences, determining sentences requiring proofreading among sentences in the document, and correcting the sentences. In the disclosed embodiment, documents to be corrected by the document proofreading apparatus 100 include all types of texts generated online or offline, such as Internet bulletin boards, SNS posts, chatting messages, etc., as well as official documents such as reports and articles, without limitation. include

As shown, the document proofreading apparatus 100 according to an embodiment includes a sentence separation module 102 , a judgment module 104 , and a proofreading module 106 .

The sentence separation module 102 separates the document to be proofread into sentence units. As described above, the document to be redacted may include all kinds of texts composed of one or more sentences. In one embodiment, the sentence separation module 102 may be configured to separate the document to be proofread into sentence units based on rules based on punctuation marks (periods, question marks, exclamation marks, etc.) or machine learning models.

In the disclosed embodiment, the document proofreading apparatus 100 performs proofreading for each sentence of the proofreading target document. At this time, since each sentence does not depend on each other, proofreading of a document can be performed very quickly when proofreading is performed in parallel for each sentence. In other words, the entire proofreading speed of the document proofreading device 100 is affected only by the number of iterative proofreadings for individual sentences, and is not affected by the number of whole sentences. In the following description, the sentence separated by the sentence separation module 102 will be referred to as a “sentence to be proofread”.

The determination module 104 determines whether correction is required for each correction target sentence generated in the sentence separation module 102 . The determination module 104 may determine whether to apply one or more correction modules among three correction modules in the correction module 106 to be described later to each sentence to be corrected. For example, the determination module 104 may determine that correction is not required according to the sentence or may determine that correction by one or more correction modules is required. In one embodiment, the discrimination module 104 may be configured to determine whether correction is required for each sentence to be corrected and the type of correction module to be used for correction by using the machine learning-based discrimination model 112 . Details of the discrimination model 112 will be described later.

The proofreading module 106 performs proofreading on sentences to be proofread which require proofreading by the determination module 104 . The correction module 106 includes a shape regeneration correction module 114 , a pronunciation regeneration correction module 116 and a complex regeneration correction module 118 .

The shape regeneration correction module 114 corrects the sentence to be proofread based on the visual characteristics of the sentence to be proofread. In one embodiment, the shape regeneration correction module 114 converts the sentence to be corrected into a high-quality image by imaging and upscaling the sentence, and then recognizing the converted image through Optical Character Recognition (OCR). You can correct the sentence by doing this.

The pronunciation reproduction correction module 116 corrects the sentence to be corrected based on the auditory characteristics of the sentence to be corrected. In one embodiment, the pronunciation reproduction correction module 116 generates voice data from the correction target sentence using a predetermined speech synthesis means (Text-to-Speech), and then separate speech recognition means (Speech-To-Speech). Text) to correct the sentence by reconverting the voice data into text.

The complex regeneration correction module 118 corrects the sentence to be proofread based on the statistical characteristics of the language to which the sentence to be proofread belongs. In one embodiment, the complex regeneration correction module 118 divides a sentence to be corrected into morphemes using a morpheme analyzer and restores the original form of each word. Thereafter, the compound regeneration correction module 118 selects a morpheme not registered in the dictionary among the divided morphemes as a correction target token, and then corrects the sentence by substituting the correction target token with another word in consideration of the statistical characteristics of the corresponding language. do.

Meanwhile, the document proofreading apparatus 100 according to an embodiment may further include an additional proofreading module 108 and a judgment module 110 .

The additional correction module 108 adds a first correction sentence obtained by correcting the sentence to be corrected using one or more of the shape regeneration correction module 114, the pronunciation regeneration correction module 116, and the complex regeneration correction module 118 described above. correct In one embodiment, the supplemental calibration module 114 may further include a machine learning-based supplemental calibration model 120 . The additional correction model 120 receives the first correction sentence, outputs a second correction sentence obtained by additionally correcting the first correction sentence, and may learn to match the second correction sentence with a preset correct answer sentence.

The determination module 110 determines whether additional calibration is required for the calibration result according to the calibration module 106 or the additional calibration module 108 . In one embodiment, the determination module 110 may determine whether the additional calibration is required using the same determination model 112 as the determination module 104 . However, this is just an example, and the determination module 110 may determine whether the additional correction is necessary by using a separate determination model different from the determination module 104 .

As a result of the judgment by the discrimination model 112, if additional correction is required, the judgment module 110 re-inputs the sentence corrected by the correction module 106 or the additional correction module 108 to the correction module 106. However, if additional proofreading is not required otherwise, the decision module 110 outputs the corrected document. In addition, in one embodiment, the determination module 110 determines whether the number of calibrations by the calibration module 106 has reached a preset maximum number of calibrations, and if the maximum number of calibrations has been reached, the relationship between whether additional calibrations are required It can be configured to output a document that has been redacted without it.

As described above, the correction module 106 includes a shape regeneration correction module 114 , a pronunciation regeneration correction module 116 and a complex regeneration correction module 118 . Hereinafter, errors in sentences that can be corrected for each correction module will be described by way of example.

First, errors that can be corrected through the shape regeneration correction module 114 are as follows.

1. In case the word itself has no meaning, but is expressed indirectly by combining letters/vowels so that a specific text is associated when looking at the form for fun

(e.g. doggie -> doggie)

2. If the result of OCR of the document looks similar to the target text, but actually extracts incorrect text

(e.g. baby -> or71, story -> oloF71, gills -> or7rㅁㅣ, gills -> or7rzl)

3. If you make a typo by separating consonants and vowels while entering text

(Example: I ate rice. -> I ate rice.)

The above examples are cases in which errors were intentionally or unintentionally generated depending on the morphological characteristics of language. Thus, OCR can provide a corrective effect. Even when the OCR result is not good, an improved result can be expected by performing OCR again after improving the image quality.

Next, errors that can be corrected through the pronunciation regeneration correction module 116 are as follows.

1. When some words are replaced with other words or numbers with similar pronunciation using similarity of pronunciation

(Example: It's been a while ha2ha2~ 5ran! (input sentence) / Hi hi~ Long time no see! (intended sentence))

2. Did you actually pronounce it?? Replaced by another word pronounced with that word

(Example: Open person~ (input sentence) / Everyone~ (intended sentence))

3. In the case of writing a sentence by adding a harsh sound or a hard sound to the pronunciation or modifying the vowel. A sentence created by slightly changing consonants and vowels in the original sentence. When a native of the language sees it, it is possible to understand what it means, but it is not properly translated when a foreigner who does not know the language sees it or uses a translator

(Example: In a foreign restaurant review, the owner of the restaurant thinks the review is a good review and does not want to delete it, but it is written for the purpose of not recommending the restaurant to Koreans.

So delicious food and good atmosphere! I wanna visit again! dope?? Kachi Ma?施?! (input sentence)

So delicious food and good atmosphere! I wanna visit again! never go! (intended sentence))

The above examples are cases in which an error occurred using the similarity of pronunciation. Therefore, an awkward sentence is converted into voice data through TTS, an appropriate utterance is obtained as a result of correction, and the effect of correcting the utterance into text suitable for the entire context of the sentence can be expected through STT.

Finally, an error that can be corrected through the compound regeneration correction module 118 is a case where a typo in a general sense occurs, such as a slight mismatch of a consonant or a vowel at a word level.

(Example: Pista is a pendulum ? 記羚年实? The etiquette is too strong. (Input sentence)

The pasta was really good, but it's so hard to make a reservation. (intended sentence))

The example above is a typo at the word level. In existing proofreaders, typos were corrected by using a lot of editing distances or preset rules. However, in the disclosed embodiment, considering the dictionary, the deep learning language model, the word perturbator, and the characteristics of the corpus, it is configured to recommend a word suitable for the context of the entire sentence while matching the context of the corpus in which the sentence is used. .

In one embodiment, the correction module 106 may sequentially execute each of the sub-correction modules (shape regeneration correction module 114, pronunciation regeneration correction module 116, and complex regeneration correction module 118). For example, the shape regeneration correction module 114 and complex regeneration correction module 114 for a specific sentence in a state in which the execution order is set in the order of the shape regeneration correction module 114, the pronunciation regeneration correction module 116, and the complex regeneration correction module 118. When it is determined that the correction module 118 needs to be corrected through the correction module 106, the correction module 106 first performs the shape regeneration correction module 114 and inputs the result to the complex regeneration correction module 118 to generate the final correction result. can be configured to

2 is an exemplary diagram for explaining a learning process of the discrimination model 112 according to an embodiment. As shown, for learning of the discrimination model 112 , documents separated by sentences, indexes of sentences requiring correction, and module information required for correction are input to the discrimination model 112 .

For example, assume that the document to be studied consists of the following two sentences.

(index 0) I like dogs.

(Index 1) The dogs are too cute.

Also, assume that the following module numbers are assigned to each calibration module.

0: Calibration of the shape regeneration calibration module (114) is required

1: Pronunciation regeneration correction module 116 needs correction

2: Complex Regenerative Calibration Module (118) Needs Calibration

3: no calibration required

In the case of the learning document, correction is required only for index 1, and specifically, “dog” is changed to “dog” by the shape regeneration correction module 114, and “cute” is changed to “cute” by the complex regeneration correction module 118. Suppose we need to correct with ". Then, the pair of (index of the sentence requiring proofreading, module information required for proofreading) for the document is given as follows.

(1, 0) -> Correction of the shape regeneration correction module (114) at sentence index 1 is required

(1, 2) -> Complex regeneration correction module (118) correction required at sentence index 1

The discrimination model 112 may receive the above information and learn to output an index of a sentence requiring correction within a learning document and necessary module information.

3 is an exemplary diagram for explaining a process of discriminating a sentence requiring correction from a document in the learned discrimination model 112 according to an embodiment.

As shown, the discrimination model 112 learned through the learning process of FIG. 2 receives a document divided into sentence units and outputs a predicted value for each correction module for each sentence. For example, assume that the discrimination model 112 receives the following documents.

(index 0) I like dogs.

(Index 1) The dogs are too cute.

The discrimination model 112 may output the need for correction for each correction module for each sentence included in the input document with a probability between 0 and 1. For example, assume that the need for calibration for each module of (index 1) is output as follows (the module number is the same as in FIG. 2).

0: 0.5

1: 0.05

2:0.3

3: 0.15

When the output probability of the discrimination model 112 is greater than or equal to a preset threshold value, it may be determined that calibration of the corresponding module is required. For example, when the threshold value is 0.3, it may be determined that correction through the shape regeneration correction module 114 and the complex regeneration correction module 118 is necessary through the above output value. In this case, the discrimination model can output a pair of the index of the corresponding sentence and the module number requiring correction as follows.

(1, 0), (1, 2)

4 is a flowchart illustrating a process of correcting sentences in the shape regeneration correction module 114 according to an exemplary embodiment.

In step 402, the shape regeneration correction module 114 receives a sentence to be corrected.

In step 404, the shape regeneration correction module 114 converts the input sentence to be corrected into an image. In one embodiment, the shape regeneration correction module 114 may convert the text to have a size equal to or greater than a preset size so as to be easily identifiable with the naked eye when converted into an image. In addition, the shape regeneration correction module 114 may set the color of the image so that text and background color have clear contrast. For example, the text of the text may be set to black, and the background color may be set to white.

At step 406, the shape regeneration calibration module 114 performs one or more of upscaling or quality improvement on the transformed image. In general, when optical character recognition is performed on images with different image quality but the rest are the same, it is known that the higher the quality level, the higher the recognition rate. Accordingly, the shape regeneration calibration module 114 first performs upscaling of the image before performing optical character recognition on the image obtained in step 404 . In this case, as the upscaling method, a general upscaling method such as an interpolation method or a separate upscaling algorithm specialized for text may be variously used. In addition to upscaling, various preprocessing processes may be additionally performed to improve image quality.

At step 408, the shape regeneration correction module 114 recognizes text in the upscaled image using optical character recognition (OCR).

In step 410, the shape regeneration correction module 114 compares the input sentence with the text recognized in step 408, performs sentence correction, and outputs a correction result.

In one embodiment, the shape regeneration correction module 114 corrects the recognized text when the number of noncomplete characters included in the recognized text is smaller than the number of noncomplete characters included in the sentence to be corrected. can be set in a sentence. For example, if there are letters (incomplete letters) composed of only consonants or vowels in a sentence to be corrected, the correction is considered better as the number of incomplete letters decreases through OCR.

In another embodiment, the shape regeneration correction module 114 compares the recognized text with the sentence to be corrected to identify a changed area, and if the frequency of occurrence of the changed area is higher than the frequency of occurrence of the area before the change, the Recognized text can be set as a corrected sentence. In this case, the region is a part of a sentence and may be understood in various meanings such as a morpheme, a word, or a word.

For example, the entire corpus of the language to which the sentence to be corrected belongs or a specific corpus selected by the user is spaced and a compound noun unit (for example, in the case of 'fruit box', two 'fruit' and 'box' Let's assume that we use a dataset that is divided into words) and stores the occurrence frequencies of the words and phrases.

If the optical character recognition is well performed, when the recognized text is compared with the sentence to be corrected, the frequency in the corpus of the changed area will be higher than that of the area before the change. However, conversely, if the optical character recognition is not good, unusual words that are not well searched in the corpus will come out. using these characteristics The shape regeneration correction module 114 compares the recognized text with the sentence to be corrected to identify a changed area, and if the frequency of occurrence of the changed area is higher than the frequency of occurrence of the area before the change, the recognized text is converted into the corrected sentence. can be set to

Through the above process, the shape regeneration correction module 114 outputs the text recognized in step 408 as a correction sentence or outputs the sentence input in step 402 as it is.

5 is a flowchart illustrating a process of correcting a sentence in the pronunciation regeneration correction module 116 according to an exemplary embodiment.

In step 502, the pronunciation regeneration correction module 116 receives a sentence to be corrected.

In step 504, the pronunciation reproduction correction module 116 generates voice data from the sentence to be corrected by using a text-to-speech. In an embodiment, the pronunciation regeneration correction module 116 may generate a plurality of different voice data for each pronunciation type of the corresponding number when there is a number in the sentence to be corrected. For example, the pronunciation regeneration correction module 116 may generate voice data for all cases in consideration of cases in which the number is read as cardinal numbers (one, two, etc.) or ordinal numbers (one, two, etc.). This is because the pronunciation can change depending on how the number is counted.

In step 506, the pronunciation reproduction correction module 116 converts the voice data into text using a speech recognition means (Speech-To-Text).

In step 508, the pronunciation regeneration correction module 116 compares the input sentence with the text converted in step 506, performs sentence correction, and outputs a correction result. At this time, the method for evaluating whether the correction of the sentence is properly performed is the same as in step 410 above. For example, in one embodiment, the pronunciation regeneration correction module 116, when the number of noncomplete characters included in the converted text is smaller than the number of noncomplete characters included in the target sentence for correction, the Converted text can be set as redacted sentences. In another embodiment, the pronunciation regeneration correction module 116 compares the converted text with the sentence to be corrected to identify a changed region, and if the frequency of occurrence of the changed region is higher than the frequency of occurrence of the original region, the converted text You can set text to redacted sentences.

If a plurality of voice data is generated in step 504, the pronunciation regeneration correction module 116 selects the text with the smallest number of incomplete characters or the highest frequency in the corpus or dictionary in the changed region and the corrected sentence. can be set to

Through the above process, the pronunciation regeneration correction module 116 outputs the text converted in step 506 (if a plurality of texts are converted) as a correction sentence, or outputs the sentence input in step 502 as it is will do

6 is a flowchart illustrating a process of proofreading sentences in the compound regeneration proofreading module 118 according to an embodiment.

In step 602, the complex regeneration proofreading module 118 receives a sentence to be proofread.

In step 604, the compound regeneration correction module 118 divides the input sentence to be proofread into a plurality of morphemes. If the part-of-speech of the divided morpheme corresponds to a verb, the compound regeneration correction module 118 may convert the corresponding morpheme into a basic form.

For example, suppose that the following correction target sentence is entered.

Sentence to be corrected: "Melon is ?記獵?."

Morphological division of the sentence to be corrected is as follows.

Melon (noun), ga (subjective), ?弩獵? (verb)

At this time, since "?弩獵?" is a verb, converting it into a basic form becomes "?年?".

In step 606, the complex regeneration calibration module 118 selects one or more of the morphemes converted to the division and base form as a calibration target token.

In one embodiment, the compound regeneration correction module 118 searches the extracted morpheme in a preset dictionary database, and if there is a morpheme that is not searched, it may select it as a correction target token. At this time, the compound regeneration correction module 118 can determine that the corresponding morpheme exists in advance only when both the headword and the part-of-speech match. In the case of the above example, since "?年?" is a word that is not looked up in a general dictionary, it may be selected as a token to be corrected.

In an additional embodiment, the compound regeneration correction module 118 inputs the correction target sentence into a preset language model to predict a token corresponding to a position of a morpheme not searched in the dictionary database, If a token corresponding to a morpheme not searched in the database is not predicted, a morpheme not searched in the dictionary database may be selected as the correction target token. That is, in this embodiment, a token to be corrected is selected through two stages of a dictionary and a language model. The compound regeneration correction module 118 may select the specific token as the correction target token only when the language model does not predict the morpheme. For example, let's assume that the sentence "You ate ?薦?" is entered into the language model and a token corresponding to the position of a morpheme (here, ??) not found in the dictionary database is predicted. Then, the language model may recommend tokens such as "bread", "rice", "lunch", and "breakfast" as tokens corresponding to the position of "??". In this case, since “??” is not included in the recommended token, the complex regeneration calibration module 118 may select “??” as a calibration target token. In addition, tokens such as "bread", "rice", "lunch", and "breakfast" are tokens that are the result of predicting the token where the token to be corrected "??" is located, so the token to be corrected is "??" can be set as correction candidate tokens for The calibration candidate token will be described in more detail in step 608 below.

In general, a pre-trained language model such as BERT is a language model that can perform various natural language processing tasks through fine-tuning. learning is carried out in a way Accordingly, if the language model cannot infer the corresponding word, it may be determined that the corresponding word is highly likely to correspond to the token to be corrected. For example, the compound regeneration correction module 118 generates M guessed words (M is a natural number greater than or equal to 1) for the deleted morpheme from the language model, and corrects the corresponding morpheme when the corresponding morpheme is not included therein. It can be finally selected as the target token.

In step 608, the complex regeneration calibration module 118 generates one or more reclamation candidate tokens for each of the reclamation target tokens.

In one embodiment, the compound regeneration correction module 118 inputs a sentence to be proofread into a preset language model to predict a token corresponding to the position of the token to be proofread, and the language model predicts N (N is a natural number of 1 or greater) may be set as the calibration candidate token for the calibration target token.

In the above example, the compound regeneration correction module 118 utilizes a pretrained language model such as BERT as a proofreading candidate token for "?磯?" Five words of "delicious", "ripe", and "expensive" can be set as correction candidate tokens.

Meanwhile, the complex regeneration calibration module 118 may additionally generate a calibration candidate token using a word perturbator.

If the proofreading target token is not a verb, the complex regeneration proofreading module 118 determines that the editing distance from the proofreading token (first proofreading token) derived from the language model is within K (where K is a natural number greater than or equal to 1) (1 to 1). K) tokens may be added to the calibration candidate tokens.

In contrast, if the proofreading target token corresponds to a term, the complex regeneration correction module 118 generates one or more candidate basic types from the proofreading target token, and the editing distance between each of the one or more candidate basic types is K (K is greater than or equal to 1). natural number) may be added to the calibration candidate tokens. This is in consideration of the probability that the stem of the token to be corrected generated through morpheme analysis is incorrectly analyzed. In the process of generating the candidate basic type, it is assumed that L consecutive letters from the first letter of the token to be corrected are the stem, and the candidate basic type is created by attaching suffixes such as 'had' or 'da'.

For example, since "?磯?" is a verb, a process of regenerating the candidate basic type is required.

First, if the stem is assumed to be '??', the prototypes are '?磯?', '?饑求?', and the following correction candidate tokens can be additionally generated through the word perturbator.

?磯? -> Taste, every, roll...

?饑求? -> Talk...

If the stem is '?gi?', the original form becomes '?gi獵?', '?記例求?', and the following correction candidate tokens can be additionally generated through the word perturbator.

?Record? -> Delicious... / ?記例求? -> ...

The complex regeneration calibration module 118 may exclude tokens that are not searched in a preset dictionary or corpus from among the calibration candidate tokens.

In step 610, the complex regeneration calibration module 118 replaces the token to be reclaimed with one of the reclamation candidate tokens. In one embodiment, the compound regeneration correction module 118 may, for each of the one or more proofreading candidate tokens, determine the Levenstein distance with the proofreading token, the part-of-speech similarity with the proofreading token, the frequency of words in the corpus, and the corpus. One or more of the idiom frequencies within may be calculated, and a token to replace the proofreading target token may be selected from among the one or more proofreading candidate tokens based on the calculated value. Each of the features for selecting the final calibration candidate token is described as follows.

1. Levenstein distance: An algorithm that calculates the editing distance between the proofreading target token and the proofreading candidate token. The smaller the editing distance, the more morphologically similar words are considered. For example, tangerine and tangerine are the same word with an edit distance of 0, and tangerine and text are different words with an edit distance of 1.

2. Part-of-speech similarity: The sentence to be corrected and the sentence in which the correction candidate token is replaced are run through a morpheme analyzer to determine the part-speech, and the higher the score, the more similar the corresponding part-of-speech is. For example, if each part of speech appears as a verb and an adjective, high scores can be given because they are similar in nature. The degree of similarity between parts of speech may be set in consideration of linguistic characteristics of each part of speech.

3. Word frequency in the corpus: After recording the frequency of each word in a general corpus containing a large amount of prose data or a specific corpus designated by the user using a morpheme analyzer, the frequency of proofreading candidate tokens is compared with each other. At this time, in the case of a verb, it is converted into a prototype and the frequency is calculated, and a higher score is given to a correction candidate token with a higher frequency.

4. Idiom Recommendation in the corpus: Idiom candidate group is created with the frequency of N grams recorded in the corpus. The token to be corrected in the sentence to be corrected is replaced with the token to be corrected, and the frequencies of N-grams including the token are compared in the idiom candidate group. In the case of verbs, they are converted into prototypes and counted in the same way as word frequencies, and the higher the frequency, the higher the score.

The method of generating the idiom candidate group is as follows.

If the example sentence is "I ate rice deliciously", simple N-grams are extracted from it. For example, in the case of bigrams, (I, rice), (rice, delicious), (delicious, eat) are extracted (converted to the basic form in the case of verbs).

Next, pairs of subject (S) and object (O), subject (S) and verb (V), and object (O) and verb (V) are extracted using the SVO Extractor. In the example above, three pairs of (I eat rice), (I eat rice), and (I eat rice) are extracted.

Next, a dependency parser is used to extract and count dependent pairs. In the example above, the three pairs of (I, eat), (Rice, eat), and (Rice, delicious) are extracted as pairs with dependencies.

In the case of the disclosed embodiment, tuples with high co-occurrence like idioms can be obtained from the corpus using the SVO extractor and the dependency parser.

The complex regeneration calibration module 118 calculates a score by obtaining features 1 to 4 above for a plurality of calibration candidate tokens and weighted averaging their values. In this case, the weighted average is performed so that the final score is higher as the Levenstein distance is lower and as the other features are higher. The degree of each weight used in the weighted average can be set by the user or optimized through model learning. Depending on embodiments, some of the features may be excluded or only one specific feature may be used.

The complex regeneration correction module 118 selects a correction candidate token having the highest score as a final correction candidate token, and replaces the correction target token with the selected final correction candidate token, thereby correcting the correction target sentence. If the token to be corrected is a verb, the final correction candidate token is appropriately transformed considering the use of the verb.

In step 612, the complex regeneration calibration module 118 outputs the calibration result.

7 is an exemplary diagram for explaining a process of learning an additional calibration model 120 according to an embodiment.

The additional correction model 120 according to an embodiment is used to additionally correct sentence corrections that are not completely performed in the shape regeneration correction module 114, the pronunciation regeneration correction module 116, and the complex regeneration correction module 118. can be used To this end, the additional correction model 120 receives sentences that have passed through the shape regeneration correction module 114, the pronunciation regeneration correction module 116, and the complex regeneration correction module 118, outputs additionally corrected sentences, and outputs additionally corrected sentences. The sentence may be learned to be identical to the correct answer. As an additional calibration model, language models such as seq2seq and transformer may be used. In this way, when the additional correction model 120 is used, it is possible to correct detailed typos that may remain in the corrected sentence.

8 is a flowchart illustrating a document proofreading method 800 according to an exemplary embodiment. The illustrated method may be performed in a computing device having one or more processors and a memory storing one or more programs executed by the one or more processors, such as the document proofreading device 100 described above. In the illustrated flowchart, the method or process is divided into a plurality of steps, but at least some of the steps are performed in reverse order, combined with other steps, performed together, omitted, divided into detailed steps, or shown. One or more steps not yet performed may be added and performed.

In step 802, the document proofreading apparatus 100 receives a sentence to be proofread.

In step 804, the document correction apparatus 100 determines whether to correct the sentence to be corrected using at least one of the shape regeneration correction module 114, the pronunciation regeneration correction module 116, and the complex regeneration correction module 118. judge

In step 806, the document correction apparatus 100 corrects the correction target sentence based on the determination result. In this case, in step 806, the first correction sentence corrected by using at least one of the shape regeneration correction module 114, the pronunciation regeneration correction module 116, and the complex regeneration correction module 118 is additionally corrected. Further steps may be included.

9 is a block diagram illustrating and describing a computing environment including a computing device suitable for use in example embodiments. In the illustrated embodiment, each component may have different functions and capabilities other than those described below, and may include additional components other than those not described below.

The illustrated computing environment 10 includes a computing device 12 . In one embodiment, the computing device 12 may be the document proofreading device 100 described above.

Computing device 12 includes at least one processor 14 , a computer readable storage medium 16 and a communication bus 18 . Processor 14 may cause computing device 12 to operate according to the above-mentioned example embodiments. For example, processor 14 may execute one or more programs stored on computer readable storage medium 16 . The one or more programs may include one or more computer-executable instructions, which when executed by processor 14 are configured to cause computing device 12 to perform operations in accordance with an illustrative embodiment. It can be.

Computer-readable storage medium 16 is configured to store computer-executable instructions or program code, program data, and/or other suitable form of information. Program 20 stored on computer readable storage medium 16 includes a set of instructions executable by processor 14 . In one embodiment, computer readable storage medium 16 includes memory (volatile memory such as random access memory, non-volatile memory, or a suitable combination thereof), one or more magnetic disk storage devices, optical disk storage devices, flash memory devices, other forms of storage media that can be accessed by computing device 12 and store desired information, or any suitable combination thereof.

Communications bus 18 interconnects various other components of computing device 12, including processor 14 and computer-readable storage medium 16.

Computing device 12 may also include one or more input/output interfaces 22 and one or more network communication interfaces 26 that provide interfaces for one or more input/output devices 24 . An input/output interface 22 and a network communication interface 26 are connected to the communication bus 18 . Input/output device 24 may be coupled to other components of computing device 12 via input/output interface 22 . Exemplary input/output devices 24 include a pointing device (such as a mouse or trackpad), a keyboard, a touch input device (such as a touchpad or touchscreen), a voice or sound input device, various types of sensor devices, and/or a photographing device. input devices, and/or output devices such as display devices, printers, speakers, and/or network cards. The exemplary input/output device 24 may be included inside the computing device 12 as a component constituting the computing device 12, or may be connected to the computing device 12 as a separate device distinct from the computing device 12. may be

Although the present invention has been described in detail through representative examples above, those skilled in the art can make various modifications to the above-described embodiments without departing from the scope of the present invention. will understand Therefore, the scope of the present invention should not be limited to the described embodiments and should not be defined, and should be defined by not only the claims to be described later, but also those equivalent to these claims.

100: document proofing device
102: sentence separation module
104: discrimination module
106 Calibration module
108: additional calibration module
110: judgment module
112: discriminant model
114: shape regeneration correction module
116: Pronunciation regeneration correction module

Claims (20)

a shape regeneration correction module for correcting the target sentence based on the visual characteristics of the target sentence;
a pronunciation regeneration correction module for correcting the target sentence based on the auditory characteristics of the target sentence;
a complex regeneration correction module for correcting the sentence to be proofread based on the statistical characteristics of the language to which the sentence to be proofread belongs; and
Document proofreading comprising a determination module receiving the sentence to be corrected and determining whether to correct the sentence to be corrected using at least one of the shape regeneration correction module, the pronunciation regeneration correction module, and the complex regeneration correction module. Device. The method of claim 1,
The shape regeneration correction module,
The document proofreading device converting the sentence to be corrected into an image and then recognizing text in the converted image using optical character recognition. The method of claim 2,
The shape regeneration correction module,
and setting the recognized text as a corrected sentence when the number of noncomplete characters included in the recognized text is smaller than the number of noncomplete characters included in the sentence to be corrected. The method of claim 2,
The shape regeneration correction module,
A document proofreading device that compares the recognized text with the sentence to be corrected to identify a changed area, and sets the recognized text as a corrected sentence when the frequency of occurrence of the changed area is higher than the frequency of occurrence of the area before the change. . The method of claim 1,
The pronunciation regeneration correction module,
A document proofreading device that generates voice data from the sentence to be corrected using a speech synthesis unit (Text-to-Speech) and then converts the voice data into text using a speech recognition unit (Speech-To-Text). . The method of claim 5,
The pronunciation regeneration correction module,
and setting the converted text as a corrected sentence when the number of noncomplete characters included in the converted text is smaller than the number of noncomplete characters included in the sentence to be corrected. The method of claim 5,
The pronunciation regeneration correction module,
A document proofreading device that compares the converted text with the sentence to be corrected to identify a changed region, and sets the converted text as a corrected sentence when the occurrence frequency of the changed region is higher than the occurrence frequency of the original region. . The method of claim 1,
The complex regeneration correction module,
Dividing the sentence to be corrected into a plurality of morphemes;
Selecting one or more correction target tokens from among the plurality of morphemes;
A document proofreading device that generates one or more proofreading candidate tokens for each proofreading target token. The method of claim 8,
The complex regeneration correction module,
If the part-of-speech of the divided morpheme corresponds to a predicate, converting the corresponding morpheme into a basic form, the document proofreading device. The method of claim 8,
The complex regeneration correction module,
A document proofreading device that selects a morpheme that is not searched in a preset dictionary database among the plurality of morphemes as the proofread target token. The method of claim 10,
The complex regeneration correction module,
The sentence to be corrected is input into a preset language model to predict a token corresponding to a position of a morpheme not searched in the dictionary database, but the language model does not predict a token corresponding to a morpheme not searched in the dictionary database. If not, select a morpheme that is not searched in the dictionary database as the proofreading target token. The method of claim 8,
The complex regeneration correction module,
The sentence to be corrected is input into a preset language model to predict a token corresponding to the position of the target token to be corrected, and N prediction tokens (N is a natural number equal to or greater than 1) predicted by the language model are the tokens to be corrected. set as the redaction candidate token for the document redaction device. The method of claim 12,
The complex regeneration correction module,
The document proofreading apparatus of claim 1 , wherein a token whose edit distance is within K (K is a natural number greater than or equal to 1) from the primary proofreading target token is added to the proofreading candidate token. The method of claim 12,
The complex regeneration correction module,
If the token to be corrected corresponds to a term, one or more candidate basic forms are generated from the token to be corrected;
and adding a token whose editing distance from each of the one or more candidate primitives is within K, where K is a natural number equal to or greater than 1, to the proofreading candidate tokens. The method according to any one of claims 12 to 14,
The complex regeneration correction module,
A document proofreading device that excludes tokens that are not looked up in a predetermined dictionary or corpus from among the proofreading candidate tokens. The method of claim 8,
The complex regeneration correction module,
For each of the one or more calibration candidate tokens,
Calculate at least one of a Levenstein distance with the token to be corrected, a similarity of parts of speech with the token to be corrected, a frequency of a word in a corpus, and a frequency of an idiom in the corpus, and based on the calculated values, among the one or more candidate tokens for correction A document proofreading apparatus for selecting a token to replace the proofreading target token. The method of claim 1,
Further comprising an additional proofreading module for additional proofreading of a first proofread sentence obtained by proofreading the sentence to be proofread by using at least one of the shape regeneration correction module, the pronunciation regeneration correction module, and the complex regeneration correction module. The method of claim 17
wherein the additional correction module receives the first correction sentence, outputs a second correction sentence obtained by further correcting it, and learns the second correction sentence to match a preset correct answer sentence. As a method performed on a computer,
receiving a sentence to be corrected;
A shape regeneration correction module for correcting the sentence to be corrected based on the visual characteristics of the sentence to be corrected;
A pronunciation regeneration correction module for correcting the sentence to be corrected based on the auditory characteristics of the sentence to be corrected; and
determining whether or not to proofread the sentence to be proofread by using one or more of complex regeneration proofreading modules for correcting the sentence to be proofreaded based on statistical characteristics of the language to which the subject sentence to be proofread belongs; and
and correcting the target sentence based on the determination result. The method of claim 19
The correcting step is
Further comprising the step of additionally proofreading the first proofread sentence obtained by proofreading the sentence to be proofread by using at least one of the shape regeneration correction module, the pronunciation regeneration correction module, and the complex regeneration correction module.

Images