KR20180113849A - Method for semantic rules generation and semantic error correction based on mass data, and error correction system implementing the method - Google Patents

Abstract

The present invention relates to a method for generating a semantic error correction rule of an error correction system operating by at least one processor. The method includes a step of extracting error data pairs from mass data, a step of generating a semantic error correction rule including correction content to be corrected when a context is detected based on a correction target where a semantic error occurs, the context, a direction of searching the context for from the correction target, and correction content for correcting the correction target if the context is detected, and a step of storing the semantic error correction rule in a semantic error rule dictionary. The error data pairs are data pairs having the same length in the left and right context but different lengths of phonemes. It is possible to efficiently search for semantic error rules.

Description

TECHNICAL FIELD [0001] The present invention relates to a method and apparatus for generating a semantic error correction rule based on a large amount of data, and a semantic error correction method, and an error correction system implementing the semantic error based correction rule,

The present invention relates to a Korean spelling error correction technique.

As the level of user 's requirement for Korean spelling error correction increases with the development of technology, there is an increasing need to detect and correct corrective errors. Conventionally, mean error correction methods have been studied, but error correction methods based on probability statistics have been used. Probabilistic statistical error correction techniques have higher recall but lower accuracy than rule - based semantic error correction techniques. On the other hand, rule - based semantic error correction techniques are highly accurate, but the surrounding word lexicals must be precisely matched before they can be created. Therefore, if large amounts of data can not be accessed, it is difficult to create rules because it is difficult to collect data pairs required for rule generation.

The problem to be solved by the present invention is to create semantic error correcting rules based on a large amount of data, to extend semantic error correcting rules by using parts of speech information, and to use semantic error rules And to provide a method and system for efficiently searching for errors and correcting errors.

A method for generating a semantic error correction rule for an error correction system operating by at least one processor, the method comprising: extracting error data pairs from a large amount of data; Generating a semantic error correction rule including a correction target, a context, a direction in which a context is searched for from the correction target, and a correction content for correcting the correction target when the context is detected, Wherein the error data pair is a data pair having the same left and right context but with a certain length of phoneme difference.

Each of the correction subject, the correction content, and the context of the meaning error correction rule is described by at least one of a word, morpheme, and part of speech information, and the part of speech information may be one of a partly part representative tag and a part of speech tag.

The direction in which the context is searched includes a first direction for instructing to detect a context existing before the object to be corrected, a second direction for instructing to detect a context existing after the object to be corrected, And a third direction for instructing to detect an existing context.

Wherein the semantic error correcting rule further includes information indicating whether to extend the rule, information indicating whether the rule is extended is first information for instructing the rule to be applied when the rule is explicitly specified, And second information indicating to apply the rule to the specified dependency relation.

The semantic error correction rule generation method comprises the steps of extracting a specific calibration object and a specific calibration content included in the generated specific semantic error correction rule by a calibration pair and extracting an extended error data pair having the same left and right contexts of the calibration pair from a large amount of data And extracting a context of the extended error data pair to generate a new semantic error correcting rule for the corrected pair.

The step of extracting the erroneous data pair comprises: purifying a large amount of data by using an n-gram pattern; calculating a difference between the erroneous data and the erroneous data, Extracting the pair of error data, or extracting the error data pair from the dependency pair obtained by segmenting and analyzing a large amount of data.

A semantic error correction method for an error correction system operated by at least one processor according to another embodiment, comprising the steps of: classifying input statements into phrases, morphemes, and parts of speech information; Extracting a specific meaning error correction rule corresponding to the context included in the calibration object and the input statement in the first rule dictionary, extracting a specific meaning error correction rule corresponding to the context included in the input statement, And calibrating the calibration object. Wherein the first rule dictionary stores a plurality of semantic error correction rules including the specific semantic error correction rule and the specific semantic error correction rule is a rule that the correction subject is displayed together with the context, Indicates the rule to correct by content. The context is represented by at least one of a word, morpheme, and part of speech information.

Wherein the step of extracting the semantic error correcting rule comprises generating at least one candidate semantic error rule that can be generated by the input statement based on the context of the input statement and searching for the presence of the semantic error rule in the first rule dictionary The specific meaning error correction rule corresponding to the candidate meaning error rule can be extracted.

The step of extracting the semantic error correcting rule may include extracting at least one generalized rule related to the object to be corrected in a second rule dictionary, extracting at least one generalized rule related to the object from among the extracted generalized rules Generating a semantic error rule of the input query by applying the input statement to the context structure of the validated generalized rule; and in the first rule dictionary, And extracting the specific meaning error correction rule corresponding to the first meaning error rule of the input query. The second rule dictionary may store a generalized rule in which the plurality of semantic error correction rules are generalized to information indicating a context structure.

Generating the second rule dictionary, wherein generating the second rule dictionary comprises converting at least one semantic error correction rule associated with the calibration subject into the generalized rule, and wherein each generalized rule The encoded value calculated according to the context detection direction, the number of contexts, and the type of the context can be stored corresponding to the calibration object.

The step of generating the second rule dictionary may include calculating a coding value of each generalized rule on the basis of a binary mask having a binary value according to the context detection direction, the number of contexts, and the type of context, And extracting the at least one generalized rule extracts at least one encoded value related to the object to be corrected in accordance with a priority in the second rule dictionary and decodes the extracted encoded value to obtain a generalized rule can do.

Wherein the step of selecting a valid rule for the context structure of the input statement includes the steps of: determining whether a first rule having the highest priority in the extracted generalized rule is valid in the context structure of the input query; Or if the first semantic error rule of the input query generated by the first rule does not exist in the first rule dictionary, the second rule of the next rank of the first rule out of the extracted generalized rules Can be determined to be appropriate for the context structure of the input query.

The method may further include determining whether an arbitrary semantic error correcting rule can be applied to the input sentence based on a result of analyzing the syntax of the input sentence before extracting the specific semantic correcting rule.

According to another embodiment of the present invention, there is provided an error correction system operating by at least one processor, the error correction system comprising: an error data pair collecting device for extracting, from a large amount of data, data pairs having the same left and right context, And a semantic error correcting rule including a correcting object in which a semantic error occurred, a context, a direction in which a context is searched from the object to be corrected, and a correction to be corrected when the context is detected, And a semantic error correcting device for correcting the error of the input query based on the semantic error correcting rules described in the semantic correcting rule description device.

The semantic error correcting rule description device may describe each of the calibration object, the calibration content, and the context as at least one of a word phrase, morpheme, and part of speech information. The part of speech information may be one of a part- have.

Wherein the semantic error correcting rule description device comprises: a first direction for instructing to search for a context existing before the object to be corrected; a second direction for instructing to detect a context existing after the object to be corrected; And a third direction for instructing detection of a context existing before and after the object to be calibrated.

The semantic error correcting apparatus classifies the input sentence into a phrase, morpheme, and part-of-speech information, extracts a subject of correction of the input sentence that is likely to cause a semantic error based on the classified information, Extracting a specific meaning error correction rule corresponding to a context included in the input subject and a correction target of the input query from the meaning correction correction rules and correcting the correction subject of the input query with the correction meaning of the specific meaning error correction rule can do.

Wherein the semantic error correcting device generates at least one candidate semantic error rule that can be generated by the input statement based on the context of the input statement and selects one of the semantic error correction rules described in the semantic error correction rule description device, The specific meaning error correction rule corresponding to the error rule can be extracted.

Wherein the semantic error correction system includes a first rule dictionary storing semantic error correction rules described in the semantic error correction rule description device and a generalized rule of semantic error correction rules stored in the first rule dictionary as information representing a context structure And a second rule dictionary that stores the rules that have been generated. Wherein the semantic error correcting device extracts at least one generalized rule relating to an object of correction of the input query in a second rule dictionary and extracts at least one generalized rule related to the input query from among the extracted generalized rules, The candidate semantic error rule can be generated by selecting a generalized rule suitable for the context structure and then applying the input statement to the context structure of the validated generalized rule.

The second rule dictionary stores a coded value corresponding to a context detection direction and a context structure included in each generalized rule, and stores the coded value corresponding to the generalized rule, and each generalized rule stores a context detection direction and a context A binary value in the binary mask may be given to indicate the priority determined according to the structure.

According to the embodiment of the present invention, it is possible to collect a large number of semantic error candidates and semantic error data pairs using a large amount of data, thereby generating semantic error rules with high accuracy. According to the embodiment of the present invention, because the rules are described using the parts-of-speech information as a context, various errors can be corrected with a minimum rule, thereby saving time and cost. According to the embodiment of the present invention, the recall rate of the error correction based on the semantic error correcting rule can be increased, and in particular, the speech recall rate can be increased by variously applying the part of speech information to the context and the length of the sentence structure .

1 is a schematic block diagram of an error correction system according to an embodiment of the present invention.
2 is an exemplary diagram illustrating a rule search path according to an embodiment of the present invention.
3 is a flowchart of a semantic error correction method of a semantic error correcting apparatus according to an embodiment of the present invention.
4 is a flowchart of an error correction method of the error correction system according to an embodiment of the present invention.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those skilled in the art can easily carry out the present invention. The present invention may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein. In order to clearly illustrate the present invention, parts not related to the description are omitted, and similar parts are denoted by like reference characters throughout the specification.

Throughout the specification, when an element is referred to as "comprising ", it means that it can include other elements as well, without excluding other elements unless specifically stated otherwise. Also, the terms " part, "" module," and " module ", etc. in the specification mean a unit for processing at least one function or operation and may be implemented by hardware or software or a combination of hardware and software have.

FIG. 1 is a schematic configuration diagram of an error correction system according to an embodiment of the present invention, and FIG. 2 is an exemplary diagram illustrating a rule search path according to an embodiment of the present invention.

Referring to FIG. 1, an error correction system (hereinafter simply referred to as an 'error correction system') 10 operated by at least one processor is a device for correcting Korean spelling errors using a rule-based semantic error correction technique , We collect error candidates and correction pairs in the large log (find out the rules), extend the semantic error correction rule by using the correction pair (rule extension), use the part of speech information in the context information, (Rule generalization), error correction rules are encoded (encoded), and matching rules can be extracted quickly.

To this end, the error correction system 10 includes an error data pair collection device 100, a semantic error correction rule description device 200, and a semantic error correction device 300. The error correction system 10 may further include a semantic error correction rule dictionary (simply referred to as a " rule dictionary ") 400 for storing the rules described in the semantic error correction rule description device 200. [ The error correction system 10 includes a generalized meaning error rule (simply referred to as a generalized meaning error correction rule) obtained by converting a semantic error correction rule into information representing a context structure, and a generalized semantic error rule (Simply referred to as a " generalized rule dictionary ") 500.

The error correction system 10 includes a basic error correction device 10 for performing at least one of a preprocessing check of input inquiry, morpheme analysis, spacing error check, spelling error check, grammar error check, and probability based check, (600). The semantic error correcting apparatus 300 and the basic error correcting apparatus 600 may be interlocked to generate a final error corrected output statement. The interworking of the semantic error correcting apparatus 300 and the basic error correcting apparatus 600 may be variously set in accordance with the error correcting order, 300 and the basic error correcting apparatus 600 may be implemented as a single apparatus. In the description, the operation of the semantic error correcting apparatus 300 will be mainly described.

The error data pair collection device 100 collects error data pairs in a large amount of unstructured documents. Error correction rules are generated from the error data pairs. The error data pair collecting apparatus 100 refines data using an n-gram pattern (e.g., 5-gram) in a large amount of data. The error data pair collecting apparatus 100 collects error data pairs having the same left and right context but the editing distance 1 (phoneme difference 1) in the refined data. For example, error data pairs as shown in Table 1 can be collected and rules are generated from error data pairs. A meaningful error correction rule is derived from the collected data pairs. At this time, the rules may be generated automatically by the machine or may be generated by spelling-qualified language experts.

Error data pair Error correction rules One My mother's dumpling comb is in her hands
My mother's debt to the dumplings is so handsome When there is a context of 'dumpling' in front of the word 'combing', it creates a rule to correct it as 'owed' 2 And every morning a shoulder bag
Mail a shoulder bag and early morning If there is a context called 'bag' in front of the word 'hanging', create a rule to correct it with 'hanging' 3 At first, this approach amounts gradually assuming its cerise
Initially, the band approached the amount of scarlet is gradually If there is a context of 'reddish' in front of the word 'stand up', create a rule that corrects it with 'reddish'

The error data pair collecting apparatus 100 can classify and analyze a large amount of data with a Korean parser (parser) and extract error data pairs from a pair of header-dependent relations. The error data pair collecting apparatus 100 generates error correction rules based on the error data pair. Table 2 shows the syntax of the Korean parser. In Table 2, '←' indicates dependency between words and is expressed in the form of 'dependency ← dominant'.

sign category Example 1 Example 2 NP The Clan (noun, pronoun, investigation) School, him, three VP Yongin Ward Verb (verb, adjective, auxiliary verb) Eat, warm, I want VNP Positive Designation Dune Yes, It's a son, a person, a thing. SBJ subject Narcissistic sphere NP_SBJ
Noun Noun School is good ←. NP_SBJ
← It is hard to study. VP_SBJ OBJ direct object Objective Clan NP_OBJ
Destination Noun Destiny Noun VP_OBJ I love you ←. NP_OBJ
Pursue beauty ←. VP_OBJ MOD Censor Tube-type cymbal NP_MOD
Tubular spoken word VP_MOD Her ← sister NP_MOD
Separated ← Lovers VP_MOD AJT Vernacular Affirmative spearhead NP_AJT I lived in Seoul ←. NP_AJT

The error data pair collecting apparatus 100 extracts a pair of data parsed from a large amount of data in a cognitive phrase (dependency) -fundamental (dominant) relation. The error data pair collecting apparatus 100 may be configured such that when the parsing mark is NP (shorthand), NP_SBJ (nominal shorthand), NP_OBJ (destination shorthand) related to Chennai, Data pairs, or dependencies, are the same and the dominator collects an error data pair with an edit distance of 1.

The error data pair collecting apparatus 100 extracts a pair of data parsed from a large amount of data in a dependency relation (dependency) - dependency relation (dominant) relationship. The error data pair collecting apparatus 100 collects error data pairs whose dominant is the same and whose dependency is the edit distance 1 when the parsing mark is VP_MOD (cube phrase) modifying the cognition.

The error data pair collecting apparatus 100 extracts a pair of data parsed from a large amount of data in a cognitive sphere (dependent) - cognitive sphere (dominant) relationship. If the parse markup is NP (sibling), NP_SBJ (nominal sphere), NP_OBJ (objective sphere), NP_MOD (nominal sphere) related to the cognition, , Or the dependency element is the same and the dominant element has an edit distance of 1.

Data pairs parsed in relation to the cognition (dependency) - spoken (dominant) relation, data pairs parsed in relation to the verb phrase (dependency) - cognitive phrase (dependent) The error data pairs parsed in relation to the old (dominant) are shown in Table 3, for example.

Error data pair Explanation Cheng Wu (Dependency) - Yonghui (Domination) Relationship Rotting
It's fast. NP, the genus (dependent), the rotten / rotten (dominant)
Data pairs with the same dependency and whose dominant position is edit distance 1 The day is ripe.
Be familiar with NP_SBJ, Day / Native (dependency),
Idi (the ruling cattle),
A data pair with the same dominion and a dependency of edit distance 1 Remove one's anger
Fry NP_OBJ, anchor (dependency),
It is called "
Data pairs with the same dependency and whose dominant position is edit distance 1 Yonghwang (Dependency) - Chengnung (domination) relationship A broken lover
Lover VP_MOD, broken / broken (dependent), lover (dominant)
A data pair with the same dominion and a dependency of edit distance 1 With struck eyes
With drooping eyes VP_MOD, struck / defeated (dependent), eye and (dominant)
A data pair with the same dominion and a dependency of edit distance 1 Pickled cabbage
Chinese cabbage VP_MOD, pickle / girin (dependence cattle), cabbage (dominant cattle)
A data pair with the same dominion and a dependency of edit distance 1 Ching Ward (Dependency) - Ching Ward (Ching Ward) Relationship Preserve ecosystem
Conserving ecosystems NP_OBJ, ecosystem (dependent), conservation / conservation (dominant), dependent Dense hat
Straw hat NP, dense / straw (dependency), cap (dominant), dominance is the same, Restraint of architecture
Materials of building NP_MOD, structure (dependency), constraint / material (control), dependency are the same,

The error data pair collecting apparatus 100 collects n-gram data for a calibration pair already collected, in a large amount of data, in order to increase the error data pair. The error data pair collecting apparatus 100 then extends the error correction rules using n-gram data for the correction pair.

Table 4 shows that when the collected error data pairs are 'comb' and 'borrow' behind the context of 'dumplings', and when there is a context of' buns' in front of the word 'comb', ' . The correction pair here is <comb, borrow>. The error data pair collecting apparatus 100 can further extract the context information (for example, 'head') by collecting the data pairs in which the left and right contexts of this calibration pair 'comb' and 'borrow' are the same. The error data pair collecting apparatus 100 adds an error correction rule including new context information (for example, a rule of correcting 'comb' when there is a context of 'head' before the word 'borrow'). In this way, the error data pair collecting apparatus 100 further increases the error correction rule by further extracting the context information of the error data pair / correction pair already collected. This is possible because the error data pair collecting apparatus 100 can collect error data pairs from a large amount of data.

Error data pairs and rules Added error data pairs Added error correction rules One Comb - borrowing,
Rule: When there is a context of 'buns' in front of the word 'comb', it is corrected by 'borrowing' Hair comb
Head shape When there is a context of 'head' in front of the phrase 'borrowed,' create a rule to correct it with 'comb' 2 Mago - Mago,
Rule: When there is a context of 'bag' in front of the word 'hanging', it is corrected by 'hanging' I was carrying the field.
I was tying the field. When there is a context of 'field' in front of the word 'Meggo', it creates a rule to correct it with 'tagged'

The semantic error correcting rule description device 200 describes semantic error correcting rules of error data pairs collected by the error data pair collecting device 100 so that the semantic error correcting device 300 can refer to them. The semantic error correction rule description device 200 stores semantic error correction rules in the rule dictionary 400. Meaning error correction rules are expressed as [key] [context search direction] [context] [correction contents (value)] [whether to extend rules], and any of these rule items may be omitted . If the input statement including the key includes a context through the semantic error correcting rule, the semantic error correcting apparatus 300 determines that the semantic error is generated and changes the object to be corrected to a value To correct for semantic errors.

Meaning Error Correction Rule The description device 200 can describe a correction key of a word or morpheme unit. Meaning that the subject to be corrected is described as 'word'. The error correction rule can be applied only to the same word as the correction subject included in the rule. Meaning that the subject of correction is described as 'morpheme'. Error correction rules can be applied to all types of proofs included in the rule. Referring to Table 5, the semantic error correction rules are expressed as [key] [context search direction] [context] [correction value] [whether to extend the rule]. [Context] refers to the context that should be found in order to calibrate the calibration object with the calibration contents. The '*' in the tag description field of the [key] or [context] means that all parts of the morpheme can come from. 'C' and 'P' attached to the rule indicate whether to extend the rule. 'C' indicates that the semantic error correcting rule is applied to the case where the rule is explicitly specified before and after the 'C' Meaning It indicates that error correction rule is extended and applied.

If we refer to Table 5, the rule of correcting 'disappeared' to 'disjointed' when the phrase 'lover' appears after the word 'disappeared'is' Can be described as [disjointed *] [C] '. This means error correction rules, the expression "Hare true lovers" is corrected to "lovers that Hare" is not the same and can be corrected as "two ex lovers', 'Hare Jean," but is "breaking up lovers' Can not. The rule that corrects' Bidda 'to' Bidda 'when it comes to the context of' Mandu 'in front of the stem' Bidda 'is described as' [Bidda morpheme *] [previous] [Morpheme = Mandu *] [Bidda *] [C] . The meaning of this error correction rule is to use 'Bidda', 'Bid', 'Comb', 'Comb', and 'Bid' .

Technical Method Meaning Error Correction Rule (for explanation to be corrected) General technology [Keyword word / morpheme] [Context search direction (before / after / after)] [context] [correction value] [whether to extend the rule (C / P)] Eulogy [Declining] * [after] [morpheme = lover *] [separated *] [C] morpheme [Criminal morpheme *] [previous] [morpheme = dumplings *] [cuddle *] [C]

Meaning Error Correction Rule The description device 200 can describe a phrase, a morpheme, a representative tag, and a detail tag in a context in [context description]. Since the context can be described in a context without specifying a word or morpheme only, it is possible to deal with many error sentences with few rules.

Part-of-speech information is expressed by dividing the part of speech in Korean morpheme into clan, verb, adjective, adverb, exclamation, investigation, ending, affix, root, and code. The part of speech information includes a representative tag indicating the part of speech of the word and a detailed tag indicating more information such as left and right access information even in the same part of speech. The detailed tag is a tag used in the morpheme analysis stage. There are various types of detailed tags for the same part of speech. For example, 'Yi Soon, Seoul and New York' have the same tags as 'proper noun' but they have detailed tags such as 'domestic person, domestic place, overseas place'.

The rules describing the context as part of speech information can be expressed as shown in Table 6. The context search direction can be set to before, after, and backward directions of the object to be calibrated. A plurality of (n) contexts can be defined, and each context is represented by one of a word, a morpheme, a representative tag, and a detail tag. Each context is described by [description method of the context x (word / morpheme / representative tag / detail tag)] = [context x], and a plurality of these are connected to define a whole context related to the object to be corrected. On the other hand, if the direction of the object to be corrected is 'forward or backward direction', it can be described as '|'

Technical Method Meaning Error Correction Rules (Context Details) General technology [Correction target] [Context search direction (Before / After / After)] {Context description part} [Correction contents] [Extension rule (C / P)] Navigated to before / after calibration target
{Context Engineering Department}  [Description method of context 1 (word / morpheme / representative tag / detail tag)] = [context 1], ... , [Description method of context n (word / morpheme / representative tag / detail tag)] = [context n] {Contextual technology department} Previous [description method of context 1 (word / morpheme / representative tag / detail tag)] = [context 1], ... , [Description method of context n (word / morpheme / representative tag / detail tag)] = [context n] | After that [description method of context 1 (word / morpheme / representative tag / detail tag)] = [context 1], ... , [Description method of context n (word / morpheme / representative tag / detail tag)] = [context n]

The error correction rules of 'BIGDA' - 'BUDDHA' expressed in the meaning error correction rule of Table 6 can be variously described as shown in Table 7. In the example shown in Table 7, since the search direction is 'previous', if there is a context designated by the context description unit before the object to be corrected, it is a rule to correct the object to be corrected to the contents of the correction.

Error correction rules and error correction results Rule 1: [Binda morphemes *] [previous] [ morpheme = buns * ] [bold *] [C]
Context 1: morpheme = buns *
Error correction result: The dumplings were combed. → I made buns. Rule 2: [bitter morpheme *] [previous] [ word = buns *] [borrow *] [C]
Context 1: Echoes = Buns
Error correction result: I diced the dumplings. → I made buns. Rule 3: [Binda morpheme *] [previous] [ morpheme = bunt *, detail tag = assistant] [bold *] [C]
Context 1: morpheme = buns *
Context 2: Detail tag = assistant
Error correction result: dumplings + assistant (degrees / moon / s) combed. → dumplings / dumplings / dumplings. Rule 4: [abridged morpheme *] [previous] [ morpheme = bunt *, representative tag = adverb ]
Context 1: morpheme = buns *
Context 2: Representative tag = adverb
Error Correction Result: Dumpling Adverb (lot / well / together / fast / bunch).
→ a lot of buns / well / together / fast / I made a lot. Rule 5: [abridged morpheme *] [previous] [ verse = dumplings *, detail tag = adjective, morpheme =
Context 1: Eatings = Buns *
Context 2: Detail tag = adjective
Context 3: Morpheme = crab
Error correction result: dumplings with adjective stem + mother-crab (deliciously / beautifully / nicely / carefully) comb
. → Delicious dumplings / beautifully / nicely / carefully.

The error correction rules of 'jigae' - 'jigae' expressed in the error correction rule of Table 6 can be variously described as shown in Table 8. In Table 8, the example rule is a rule that corrects the object to be corrected to the contents of correction if there is a context specified by the context description unit after the object to be corrected since the search direction is 'after'.

Error correction rules and error correction results Rule 1: [stemming morpheme *] [after] [ morpheme = boiling *] [stewing *] [C]
Context 1: morpheme = boil *
Error correction result: In the evening, it was boiled. → I cooked the stew in the evening. Rule 2: [stemming morpheme *] [after] [ Representative tag = assistant, morpheme = boil *] [stew *] [C]
Context 1: Main Tag = Assistant
Context 2: morpheme = boil *
Error correction result: Steaming + auxiliary (boiling / boiling / boiling) boiling. → The stew / stew / stew boiled. Rule 3: [stemming morpheme *] [after] [ Representative tag = adjective, morpheme = stem, morpheme = boil *] [stew *] [C]
Context 1: Representative tag = adjective
Context 2: Morpheme = crab
Context 3: morpheme = boil *
Error correction result: Stubborn adjective stem + mother-crab (delicious / hot / spicy / flavor) To boil → Stew delicious / hot / spicy / flavor to boil

Table 9 shows the context search rules expressed by the error correction rules in Table 6 and the error correction results thereof. In the example shown in Table 9, since the search direction is 'before and after', if there is a context designated by the context description unit before and after the object to be corrected, it is a rule to correct the object to be corrected to the contents of correction.

Error correction rules and error correction results Rule 1: Raised Eojeol *] [after] [= Eojeol water * | Morpheme = drink *] [long *] [C]
Context 1: Word = water *
Context 2: Morphology = Drink *
Error correction Result: I drank water from a well. → I drank water from a well. Rule 2: [low Eojeol *] after [the Eojeol = sick * | Morpheme = do *] [better *] [C]
Context # 1:
Context 2: morpheme = do *
Error correction results: The doctor lowered the sick. The doctor healed the sick. Rule 3: [tied up] * [before and after] [ representative tag = subject survey | Morpheme = hotel *] [stay [*]] [C]
Context 1: Representative Tag = Subject Search
Context 2: Morpheme = Hotel *
Error Correction Result: We will tie them to the hotel → we will arrange the hotel / Rule 4: [Etiquette *] [Before and After] [ Morphea = Morpheme = I am *] [pretty *] [C]
Context 1: Morphology =
Context 2: Morpheme = Ida *
Error Correction Result: Coal is / is heated. → The coal is quite hot.

The rules described with reference to Table 7 to Table 9 up to now correct the calibration target to the calibration content if the context specified in the error correction rule exists. When the error correction rule is applied in this way, it can not cope with a context that has not been found yet. There are also cases where the position of the context is out of the allowable range of semantic error correction rules. In order to solve such a problem, the semantic error correction rule description device 200 attaches a value (for example, 'C' or 'P') indicating whether the rule is extended to the error correction rule so as to strictly apply the error correction rule ('C'), and whether to apply the semantic error correction rules adaptively ('P') through parsing. For example, as shown in Table 10, a rule may be described by including a value indicating a rule extension (for example, 'P').

Rule 1: [verb morpheme verb] [previous] [verse = in the eye *] [verb verb] [P] Rule 2: [occlusion morpheme *] [since] [morpheme = dark *] [red *] [P]

When the error correction rule detected by the error correction rule of the input query includes a value ('P') indicating the rule extension, the semantic error correction device 300 analyzes the syntax of the input query with a syntax analyzer, It is possible to judge whether or not to apply it.

The scope of semantic error correction through parsing may be different depending on whether it is a unit-of-word context or a morpheme context. For example, if the rule searched by the error correcting rule of the input query is a rule containing the unit-of-word context (the word in the eye) [(the verb is the morpheme verb) ]), Semantic error correction can be promised to correct errors by applying the corresponding rule to all dependency types ([NP-AJT] ← in the eye).

A rule that is searched by the error correction rule of the input query is a rule that includes a morpheme unit context and the rule is that the key is dependent and the context is a header (for example, Apply the corresponding rule to all dependency types (occlusion is [NP-SBJ] ← dark) at the time of semantic error correction, if the occlusion morpheme *] [after] [morpheme = dark *] [ It can be promised to correct errors.

On the other hand, if the rule retrieved by the error correction rule of the input query is a rule including the morpheme unit context and the key is the header and the context is the dependent rule, Only when it is a dependency type, error correction can be applied by applying the rule. The specified dependency type may include at least one of NP, NP_SBJ, NP_OBJ, NP_MOD, VP, VP_SBJ, VP_OBJ, VP_MOD, VNP_SBJ, VNP_OBJ and VNP_MOD, for example. For example, as shown in Table 11, the rule ([bark morphological verb] [previous] [morpheme = home *] [compose verb] [P]) is a 'house large bark "to" nidify significantly "as rules for calibrating inde, you accept these rules to 'the dog barks loudly at home "can cause errors by modifying the" puppy smiles loudly at home. Therefore, in this case, the rule is applied only when the specified dependency type is specified to instruct error correction.

Input statements and dependencies Error correction rules and calibration results Magpies constantly bark all day.

key = make (dominant),
context = Magpie (dependency)
Magpie → make [NP_SBJ] [Preposition morpheme verb] [previous] [morpheme = magpie *] [bark verb] [P]

Correction results: Magpies constantly shout all day The puppy barks loudly at home.

key = home (dependency),
context = bark (dominion)
Home → bark [NP_AJT] [Bold] morpheme verb [previous] [morpheme = house *] [verb verb] [P]

Calibration result: [NP_AJT] is not calibrated because it is not applicable to the meaning error rule.

The semantic error correcting apparatus 300 corrects semantic errors of the input query based on the rules described in the semantic error correcting rule description device 200, that is, rules stored in the rule dictionary 400. [

First, the semantic error correcting apparatus 300 judges whether there is a correction target (key) that may cause a semantic error in the input word morpheme and morpheme. The semantic error correcting apparatus 300 extracts at least one semantic error correcting rule including a key to be calibrated and a context of an input query in the rule dictionary 400. [ The meaning error correcting apparatus 300 corrects the key to be corrected with the corrected value of the extracted meaning error correcting rule. The semantic error correcting apparatus 300 generates a semantic error-corrected output statement by generating a vocabulary in a utilization form when the corrected vocabulary / morpheme is grammatically combinable or grammatical error check.

Meanwhile, since the rule dictionary 400 is represented by a context combination of words, morpheme, and parts of speech information (representative tags, detailed tags), a rule for correcting a single calibration subject to a calibration value is very much described . Therefore, the semantic error correcting apparatus 300 must fetch at least one rule from the semantic error correcting rule dictionary 400 for error correction of the input query, and the semantic error correcting rule dictionary 400 includes at least one rule corresponding to the combination of the key and the context It is necessary to check whether the meaning error correction rule exists in the rule dictionary 400. [ However, if all rule candidates that are likely to exist for error correction of the key to be corrected are generated from the context combination of the input query and if all rule candidates are actually present in the rule dictionary 400, . Therefore, the semantic error correcting apparatus 300 can classify a path for searching for context information (phrase, morpheme, and part-of-speech information) that can be present in a sentence according to the type (keyword or morpheme) of the correction target key of the error correction rule Set as error rule. A plurality of generalized semantic error rules can be defined, and each generalized semantic error rule can be prioritized. The generalized semantic error rules can be stored in the generalized rule dictionary 500.

Referring to FIG. 2 (a), when the input statement is ' force is strong ', it can be judged that the calibration subject (key) having a possibility of occurrence of a semantic error in the input query word and morpheme is 'leaky'. The semantic error correcting apparatus 300 must extract at least one semantic error correcting rule including the subject to be calibrated and the context of the input statement in the rule dictionary 400. If the context which can exist as a rule is a word ' It can be a combination of a morpheme 'power' and its part-of-speech tag, an assistant 'doo' and its part-of-speech tag.

At this time, the semantic error correcting apparatus 300 may generate the generalized rule dictionary 500 as shown in FIG. 2B without checking whether all possible rule candidates exist in the rule dictionary 400 for error correction of the input query. A search for a context consisting of a 'morpheme' and a 'tag' indicating parts of speech by searching in a 'previous' direction of a correction subject, from a verb morpheme corresponding to the subject to be corrected (key) ).

Then, the semantic error correcting apparatus 300 generates semantic error correcting rules (a combination of a morpheme 'force' and a contextual tag of a part of speech tag 'assistant') as context information of a valid path, (For example, 'Cedar') of the rule existing in the rule dictionary 400 is obtained.

To this end, the semantic error correcting apparatus 300 generalizes semantic error correcting rules of the rule dictionary 400 to information indicating a context structure, and generates generalized semantic error rules. Generalized semantic error rules can be coded and stored. Here, the generalized semantic error rule means that the semantic error rule (for example, the rule of correcting 'misdata' to 'error') is generalized to the context structure information such as part-of-speech tag.

Multiple generalized semantic error rules per generalized key (key) can be defined, and each generalized semantic error rule can be encoded. Each generalized semantic error rule may have a priority set. The generalized semantic error rules and their encoded values can be stored in the generalized rule dictionary 500. The semantic error correcting apparatus 300 can read the rule dictionary 400 and convert the key and its rules into generalized semantic error rules in advance.

For example, referring to Table 12, if there are three rules for correcting 'Bidda' to 'Bidda', Rule # 1 is [ If an adverb] is in this context, for example, rules for correcting the "you're good comb the rice cake," rule # 2 - stemming Speech is a noun of "rice cake", the representative tag is objective investigation, Eojeol a 'clean' - the If that context, examples and rules for 'clean bitne a rice cake, proofreading, rule # 3 If you have to have [Speech' rice cake 'representatives tag Assistant regardless of] the context, for example, only rice cake comb called "the These are the rules for calibration and are generalized as shown in Table 13.

number Meaning error rule example #One [Comb morphological verb] [previous] [morpheme = noun rice cake, a representative tag = survey objective, a representative tag = adverb] [bitda verb] [C] #2 [Comb morphological verb] [previous] [morpheme = noun rice cake, a representative tag = survey objective, Eojeol = clean *] [bitda verb] [C] # 3 [Comb morphological verb] [previous] [morpheme = songpyeon *, representative tag = Assistant] [bitda verb] [C]

number Generalized semantic error rules Key
(Word / morpheme) Direction Context
(Word / morpheme / part of speech) #One Blended morpheme verbs Previous Morpheme_tag, representative tag, representative tag #2 Blended morpheme verbs Previous Morpheme _tag, representative tag, word _all # 3 Blended morpheme verbs Previous Morpheme _all, representative tag

The semantic error correcting apparatus 300 may assign the context search direction and the context information of a plurality of generalized semantic error rules to a base priority and encode (encode) the same with a value indicating the base priority. The semantic error correcting apparatus 300 can represent important priorities of each generalized semantic error rule by placing important information among the context search direction and the context information in the upper bits.

Referring to Table 14, the priority of the semantic error rule has the search direction of the context, the window size, and the context unit rank, and is located in the upper bit according to each significance level. Referring to Table 15, when searching for a maximum of five windows before and after the object to be corrected, each generalized semantic error rule can be represented by a 32-bit encoded value. At this time, the context search direction can be assigned to the upper two bits and encoded. Here, the priority assigned to the previous / next / backward direction may be given in the order of before, after, before, and after. Since the importance in the forward and backward directions is high, the bit value 11 is given. P5-P1 / N1-N5 displays the types of contexts (word / morpheme / part-of-speech information, etc.) existing before or after the object to be calibrated according to window order. Values are assigned according to the importance of each context. The importance of the context can be expressed by the bit value assigned in Table 14. [ If the importance of the phrase &quot; tag &quot; in the context is highest, the bit value of the phrase &quot; tag &quot; is assigned to a value (e.g., 110) larger than the other context. The values assigned to the word / morpheme / parts information can be given differently depending on the key or the contribution in the context.

Priority [Direction] [Window size] [Context unit]
- [Direction]: Priority in order of before, after, before, and after
- [window size]: long length (5,4,3,2,1)
- [contextual unit]: word_tag, word_ * (all), morpheme_tag, morpheme_ * (all), representative tag, detail tag Bit representation 2bit: Direction
- Before and after: 11 (3)
- Previous: 10 (2)
- After: 01 (1)

30bit: Context (3 bits before, after each 5 contexts)
- Eagle _tag: 110 (6)
- E-mail _ * (all): 101 (5)
- morpheme_tag: 100 (4)
- morpheme _ * (all): 011 (3)
- Representative Tags: 010 (2)
- Detailed Tags: 001 (1)

Direction (D) P5 P4 P3 P2 P1 N5 N4 N3 N2 N1 2bit 3bit 3bit 3bit 3bit 3bit 3bit 3bit 3bit 3bit 3bit

For example, when the generalized semantic error rule of Table 13 is encoded with the binary mask of Table 15, it is as shown in Table 16, and their encoding values and priority can be calculated as shown in Table 17. Rule # 1, # 2, The context search direction of # 3 is 'previous' and the upper two bits are assigned the same '10'. Since the context window size of the rules # 1 and # 2 is 3, bits are assigned to P3, P2, and P1, and the rule window # 3 has the context window size of 2. Therefore, bits are assigned to P2 and P1, '00' (morpheme _tag), '011' (morpheme _ *), and '010' (representative tag) may be assigned according to the context of the word '

As a result of such encoding, in the case of the semantic error rule 'BIGDA-> overflow' in Table 12, the encoding value of rule # 2 is the largest, meaning that rule # 2 has the highest priority.

rule D P5 P4 P3 P2 P1 N5 N4 N3 N2 N1 2bit 3bit 3bit 3bit 3bit 3bit 3bit 3bit 3bit 3bit 3bit #One 10 000 000 100 010 010 000 000 000 000 000 #2 10 000 000 100 010 101 000 000 000 000 000 # 3 10 000 000 000 011 010 000 000 000 000 000

rule Encoded value Priority #One 2156462080 2 #2 2156560384 One # 3 2148335616 3

Next, a method for the semantic error correcting apparatus 300 to correct the semantic error of the input statement using the generalized semantic error rule will be described below.

3 is a flowchart of a semantic error correction method of a semantic error correcting apparatus according to an embodiment of the present invention.

Referring to FIG. 3, the semantic error correcting apparatus 300 classifies the input sentence into phrases, morphemes, and parts of speech information (representative tags and detailed tags), and selects a correction target key (S110). Table 18 shows the morpheme and part-of-speech information of the input sentence when the input sentence is ' beautifully combed '. The semantic error correcting apparatus 300 can judge the morpheme 'BIDDER' as a correction subject having a possibility of a semantic error. On the other hand, it can be judged as a correction subject (key) which means a word 'combed', but it explains a case where the word 'comb' is a correction subject (key).

Eulogy songpyeon
noun_phrase beautifully
pa ec I did it.
pv ec morpheme songpyeon
noun of
Object investigation pretty
adjective to
ending of a word comb
verb
-> key The
A good mother All
ending of a word Representative tags noun Object investigation adjective ending of a word verb A good mother ending of a word Detail tag noun Object investigation adjective ending of a word verb A good mother ending of a word

The semantic error correcting apparatus 300 extracts generalized semantic error correcting rules in which the partly tagged information verb of the morpheme 'Bidda' is a key in the generalized rule dictionary 500 (S120). At this time, the generalized rule dictionary 500 stores the encoded values, so that the semantic error correcting apparatus 300 firstly extracts the encoded value '2156560384' with the encoded rule, and decodes the encoded rule '2156560384' ) To obtain the generalized semantic error correcting rule # 2. Referring to Table 12, there are three types of context information for correcting the pair of correction <BIGDA morphological verb, BUDGDA verb>, and the generalized error correction rule that can exist as a combination of them is three as shown in Table 13. Thus, even though the context of the input word can be variously combined in Table 18, the semantic error correcting apparatus 300 has the structure information that is structurally valid according to the object to be corrected. ) [Bidda morpheme verbs] that are related to the generalized error correction rules.

The semantic error correcting apparatus 300 verifies whether the extracted generalized error correcting rule is a valid rule in the context structure of the input statement (S130). For example, the semantic error correction apparatus 300 determines whether a context search path corresponding to the generalized error correction rule is generated based on the results of the word and morphological analysis of Table 18. For example, referring to Tables 13 and 18, Rule # 1 can search all of the contextual paths corresponding to the morpheme_tag, the representative tag, and the representative tag in the previous direction, Generalized error correction rules. Rule # 2 is also a valid generalized error correction rule because the context is the morpheme _tag, the representative tag, and the word _all in the previous direction, and the corresponding path can be searched. Rule # 3 is a valid generalized error correction rule because the context is the morpheme _all, the representative tag in the previous direction, and the corresponding path can be searched. On the other hand, if the generalized error correction rule is output as a coded value as shown in Table 17, the semantic error correcting device 300 can confirm the context information of the generalized error correction rule after decoding it. When the semantic error correcting apparatus 300 is programmed to sequentially operate according to the priority of the extracted generalized error correction rules, if the rule having a higher priority among the extracted generalized error correction rules is not valid in the context structure of the input query , It is possible to verify whether the generalized error correction rule extracted in subordinate order is a valid rule in the context structure of the input query.

The semantic error correcting apparatus 300 generates an error correcting rule of the input query according to the generalized error correcting rule verified by a proper rule (S140). For example, referring to Table 18, it can be seen that the error correcting rule of the input query generated by rule # 1 is "[Binda morpheme verb] [previous] [morpheme = investigation, representative tag = adjective, representative tag = The error correction rule of the input query generated by # 2 is' [Binda morpheme verb] [previous] [morpheme = Song noun, representative tag = The rule can be '[BIGDA] morpheme verb [previous] [morpheme = pretty *, representative tag = mother]'.

The semantic error correcting apparatus 300 searches the rule dictionary 400 for the error correction rule of the input query to find the correction value (S150).

If there is an error correction rule of the input query in the rule dictionary 400, the meaning error correction device 300 extracts the correction value of the error correction rule (S160). If there is no error correcting rule of the input query in the rule dictionary 400, the semantic error correcting apparatus 300 moves to step S130 of verifying whether it is a valid rule and generates a new error correcting rule of the input query.

The meaning error correcting apparatus 300 calibrates the calibration target (key) with the extracted calibration value (S170).

The semantic error correcting apparatus 300 generates a semantic error-corrected output statement by generating a vocabulary in a utilization form when the corrected vocabulary / morpheme is grammatically combinable or grammatical error check (S 180).

In particular, since the generalized error correction rules are sequentially arranged in order of priority, the semantic error correction device 300 generates an error correction rule of the input query corresponding to the context information of the input query in order of priority, 400), if there is an error correction rule of the input query, the error correction can be terminated by extracting the correction value (value). That is, referring to Table 17, the semantic error correcting apparatus 300 searches whether the error correction rule of the input query generated by rule # 2 having a higher priority exists in the rule dictionary 400, and if so, And the semantic error correction procedure can be terminated by extracting the correction contents ('ode'). If the error correction rule generated by the rule # 2 does not exist in the rule dictionary 400, the error correction device 300 corrects the error correction rule of the input query generated by the rule # 1 corresponding to the next priority The rule dictionary 400 is searched. Therefore, even if there are a plurality of generalized error correcting rules, the semantic error correcting apparatus 300 generates error correcting rules of the input query corresponding to the context information of the input query, If there is no need to search for the calibration contents, the error correction rule search procedure can be terminated as soon as the calibration contents are output.

On the other hand, the semantic error correcting apparatus 300 can analyze the syntax of the input statement with the syntax analyzer to determine whether to apply the semantic error correcting rule. In this case, before the generalized error correction rule is verified in the input structure of the context structure (step S130), the semantic error correcting apparatus 300 may determine whether there is a semantic error correction rule (That is, whether the semantic error correcting rule is applied). For example, if you can apply error correction only when you specify a dependency type, and you do not apply the semantic error correction rule if it does not fall under the specified dependency type, you do not need to search semantic error correction rules from the beginning It is because there is not.

4 is a flowchart of an error correction method of the error correction system according to an embodiment of the present invention.

Referring to FIG. 4, the error correction system 10 performs preprocessing and morphological analysis when an input statement is received (S210). The error correction system 10 can identify the spell relation, the dependency relation, etc. through the classification analysis.

The error correction system 10 performs a text error check (spelling error check, spelling error check, etc.) (S220).

The error correction system 10 extracts a correction target (key) having a possibility of occurrence of a semantic error among the input word morpheme and morpheme (S230). Referring to Table 18, the candidates for the key may be a word 'BITSU', a morpheme 'BITDA', and a representative tag or a detailed tag indicating the parts of speech information.

The error correction system 10 extracts at least one semantic error correction rule capable of correcting a semantic error of a key based on the rule dictionary 400 and the generalized rule dictionary 500 at step S240. Meaning Error correction rules are a rule that includes a context search direction and a context. The context is described by not only a morpheme or a phrase, but also parts of speech information. The context may be composed of a plurality of context information.

The error correction system 10 corrects the calibration target (key) to the calibration value of the semantic error correction rule (S250).

The error correction system 10 checks whether the corrected vocabulary / morpheme is grammatically combinable or grammatical error check, generates a vocabulary in the form of a morpheme if it is a morpheme, and generates a semantic error-corrected output statement (S260).

As described above, according to the embodiment of the present invention, it is possible to collect a large number of error candidates and error data pairs by using large-capacity data, so that semantic error rules with high accuracy can be found and generated and the construction time can be shortened. According to the embodiment of the present invention, because the rules are described using the parts-of-speech information as a context, various errors can be corrected with a minimum rule, thereby saving time and cost. According to the embodiment of the present invention, it is possible to increase the recall rate of the error correction based on the error correction rule, and in particular, it can be applied variously to the length and configuration of the sentence by using the part of speech information as context.

The embodiments of the present invention described above are not implemented only by the apparatus and method, but may be implemented through a program for realizing the function corresponding to the configuration of the embodiment of the present invention or a recording medium on which the program is recorded.

While the present invention has been particularly shown and described with reference to exemplary embodiments thereof, it is to be understood that the invention is not limited to the disclosed exemplary embodiments, It belongs to the scope of right.

Claims (20)

A method for generating a semantic error correction rule for an error correction system operating by at least one processor,
Extracting error data pairs from a large amount of data,
A semantic error correcting rule including a correcting object in which a semantic error occurred, a context, a direction in which a context is searched from the object to be corrected, and a correction content to be corrected when the context is detected, Creating, and
And storing the semantic error correcting rule in a semantic error correcting rule dictionary,
Wherein the error data pair is a data pair having the same left and right context but a certain length of phoneme difference. The method of claim 1,
Wherein each of the calibration target, the calibration content, and the context of the semantic error correction rule is described as at least one of a word, morpheme, and part-
Wherein the part-of-speech information is any one of a part-of-speech representative tag and a part-of-speech tag. The method of claim 1,
The direction to search for the context
A second direction for instructing to detect a context existing after the calibration object, a second direction for instructing detection of a context existing after the calibration object, and an instruction to detect a context existing before and after the calibration object And a third direction in which the semantic error correction rule is generated. The method of claim 1,
Wherein the semantic error correction rule further includes information indicating whether to extend the rule,
The information indicating whether or not the rule is expanded
A first information for instructing the rule to be applied to a case clearly specified in the rule, and a second information for instructing the rule to be applied to the dependency specified through the parsing. The method of claim 1,
Extracting a specific calibration object included in the generated specific meaning error correction rule and a specific calibration content by a calibration pair,
Collecting from the bulk data the same extended error data pair of left and right contexts of the calibration pair,
Extracting a context of the extended error data pair and generating a new semantic error correcting rule for the corrected pair
The method comprising the steps of: The method of claim 1,
The step of extracting the error data pair
extracting a pair of data having the same length in the left and right context but having a predetermined length of phonemic difference in the refined data as the error data pair,
A semantic error correction rule generation method for extracting the error data pair from a dependency pair obtained by classifying and analyzing a large amount of data. A semantic error correction method of an error correction system operating by at least one processor,
Classifying the input sentence into a word phrase, a morpheme, and part-of-speech information, extracting a correction target of the input query that may cause a semantic error based on the classified information,
Extracting a specific meaning error correction rule corresponding to the context included in the input subject and the calibration subject in the first rule dictionary, and
Correcting the calibration target with the calibration content of the specific meaning error correction rule,
Wherein the first rule dictionary stores a plurality of semantic error correction rules including the specific semantic error correction rule and the specific semantic error correction rule is a rule that the correction subject is displayed together with the context, Indicates the rule to correct by content,
Wherein the context is represented by at least one of a word, morpheme, and part of speech information. 8. The method of claim 7,
The step of extracting the semantic error correction rule
And generating at least one candidate semantic error rule that can be generated by the input statement based on the context of the input statement, searching for the candidate semantic error rule in the first rule dictionary, Meaning error is to extract calibration rules, meaning error correction methods. 8. The method of claim 7,
The step of extracting the semantic error correction rule
Extracting, in a second rule dictionary, at least one generalized rule relating to the calibration object,
Selecting a generalized rule that is valid in the context structure of the input query from among the extracted generalized rules based on information obtained by classifying the input statement,
Generating a semantic error rule of the input statement by applying the input statement to the context structure of the validated generalized rule; and
Extracting the specific meaning error correction rule corresponding to the first meaning error rule of the input query in the first rule dictionary,
Wherein the second rule dictionary stores a generalized rule in which the plurality of semantic error correction rules are generalized to information representing a context structure. The method of claim 9,
Further comprising generating the second rule dictionary,
The step of generating the second rule dictionary
Converting at least one semantic error correcting rule related to the object to be corrected into the generalized rule, calculating an encoding value calculated according to a context detection direction, a number of contexts, and a context included in each generalized rule, And storing the corresponding semantic error correction method. 11. The method of claim 10,
The step of generating the second rule dictionary
Determining the priority of the generalized rule by calculating an encoding value of each generalized rule based on a binary mask having a binary value according to the context detection direction, the number of contexts, and the type of context,
The step of extracting the at least one generalized rule
Extracting at least one encoded value related to the object to be corrected in accordance with a priority order in the second rule dictionary, and decoding the extracted encoded value to obtain a generalized rule. 12. The method of claim 11,
The step of selecting a valid rule for the context structure of the input query
Determining whether the first rule having the highest priority in the extracted generalized rule is valid in the context structure of the input query,
When the first rule is not valid in the context structure of the input statement or when the first semantic error rule of the input query generated by the first rule is not present in the first rule dictionary, And judges whether the second rule of the next rank of the first rule is valid in the context structure of the input statement. 8. The method of claim 7,
Before the step of extracting the specific meaning error correction rule,
Determining whether an arbitrary semantic error correcting rule can be applied to the input sentence based on a result of analyzing the syntax of the input sentence
Wherein the error correction method further comprises: An error correction system operated by at least one processor,
An error data pair collecting device for extracting a pair of data having a certain length of phoneme difference from a large amount of data in the same right and left context but in an error data pair, and
A semantic error correcting rule including a correcting object in which a semantic error occurred, a context, a direction in which a context is searched from the object to be corrected, and a correction content to be corrected when the context is detected, Generate meaning error correction rules technology device, and
A meaning error correcting device for correcting an error of the input query based on the meaning error correcting rules described in the semantic error correcting rule description device
The error correction system comprising: The method of claim 14,
The semantic error correction rule description device
Describing the calibration object, the calibration content, and the context as at least one of a word phrase, morpheme, and part-
Wherein the part-of-speech information is one of a part-of-speech representative tag and a part-of-speech tag. The method of claim 14,
The semantic error correction rule description device
The direction to search for the context
A second direction for instructing to detect a context existing after the calibration object, a second direction for instructing detection of a context existing after the calibration object, and an instruction to detect a context existing before and after the calibration object And a second direction in which the first and second directions are orthogonal to each other. The method of claim 14,
The semantic error correcting device
Classifying the input sentence into a phrase, morpheme, and part-of-speech information, extracting a correction target of the input query that may cause a semantic error based on the classified information,
Extracting a specific meaning error correction rule corresponding to a context contained in the input subject and a correction subject of the input query from the semantic error correction rules described in the semantic error correction rule description device,
And corrects the calibration object of the input query with the calibration contents of the specific meaning error correction rule. The method of claim 17,
The semantic error correcting device
Generating at least one candidate semantic error rule that can be generated by the input statement based on the context of the input statement,
Wherein the specific meaning error correction rule corresponding to the candidate meaning error rule is extracted from the semantic error correction rules described in the semantic error correction rule description device. The method of claim 17,
A first rule dictionary storing semantic error correction rules described in the semantic error correction rule description device, and
Further comprising a second rule dictionary storing a generalized rule in which semantic error correction rules stored in the first rule are generalized to information indicating a context structure,
The semantic error correcting device
Extracting at least one generalized rule relating to an object of correction of the input query from a second rule dictionary and extracting at least one generalized rule from the generalized rules extracted from the extracted generalized rules, And then generates the candidate semantic error rule by applying the input statement to the context structure of the validated generalized rule. 20. The method of claim 19,
The second rule dictionary
A binary value is added to the context detection direction and the context structure included in each generalized rule to store the encoded value corresponding to the generalized rule,
Wherein each generalized rule is assigned a binary value in a binary mask to indicate a priority determined by the context detection direction and the context structure.

Images