KR101126186B1 - Apparatus and Method for disambiguation of morphologically ambiguous Korean verbs, and Recording medium thereof - Google Patents

Abstract

Disclosed are a verb analyzing apparatus, a method, and a recording medium thereof, in which morphological neutrality exists. According to an embodiment of the present invention, a verb analysis apparatus having a form neutrality may include a first verb utilization and a context characteristic related to the first verb utilization, in which a basic form is identified in an unprocessed cork without a label. Basic data collection unit for collecting the learning data including a; And a training performer configured to learn a classifier between a basic type and a context based on the learning data. And a verb analysis unit extracting a contextual feature around the second verb utilization in which the basic form is doubled in the corpus, and grasping the basic form based on the learned classifier. A difficult case determination unit for grasping an example that is difficult to determine with a classifier derived from existing learning data; Communication unit for collecting a web count for the basic judgment on the difficult case; Include additional training units to add and relearn difficult cases. According to embodiments of the present invention, it is possible to improve the accuracy and efficiency of text analysis without the need for artificial classification of corpus, and to automate the entire process of text analysis including database construction.

Description

Apparatus and Method for disambiguation of morphologically ambiguous Korean verbs, and Recording medium

TECHNICAL FIELD The present invention relates to morphological analysis and tagging of text, and more particularly, to an apparatus, a method, and a recording medium having a verbal analysis.

Machine learning methods based on probabilistic models using corpus have recently been successfully applied to solve natural language processing problems. For example, the statistical parts-of-speech tagging system is about 97% accurate regardless of the language. However, the error rate of 3% can still be overlooked in real-world applications that require high-performance, accurate language processing. Most of these errors cannot be solved by limited context and word information that are considered in existing tagging systems. Even if rule-based post-processing is performed, it is difficult to achieve performance improvement because the rules themselves are generalized from error-prone cases.

Among these morphological errors, errors due to verb-verb morphological ambiguity are recognized as the most difficult and important problems in morphological analysis and tagging in Korean. Correspondence errors are due to morphological variations that occur in the conjugation of verbs. Conjugation refers to verb changes and the use of a word in multiple forms to indicate grammatical functions in a sentence. Means. In the case of Korean, morphological variation occurs frequently in the process of combining the stem and the mother.

In general, natural language processing systems such as search, translation, and mining use the morphological analysis and tagging steps to find the base form of words and parts of speech from text or text as the most basic process. Morphological significance means that the solution of the prototypical form is significant in finding the prototype of the word from the morphological variations in the text. For example, when the word "selling" appears, the stem can be of two forms: "sell" and "sell", and the correct stem stems from the surrounding context of the sentence or the subject of the document ( depending on the topic).

The problem of finding the corresponding prototype of verbs that induce morphological neutrality is hereinafter referred to as verb-verb morphological disambiguation, VVMD, and verb-verb morphological neutrality. (verb-verb morphological ambiguity) is abbreviated as VVMA.

In addition, most verbs that induce morphological significance when used often belong to a multilingual word in terms of sense, and thus they appear frequently in texts. Big.

The first technical problem to be achieved by the present invention is general unlabeled general learning without the process of classifying or class (labeling) information by hand for learning corpus required for the machine learning technique applied for verb prototype recognition. It is to provide a verb analysis device that can improve morphological analysis performance by automatically analyzing verbs with morphological significance using unlabeled corpus and web count.

The second technical problem to be achieved by the present invention is to provide a machine learning method for effective verb analysis applied to the verb analysis device having the above form morphism.

The third technical problem to be achieved by the present invention is to provide a computer-readable recording medium for executing the above-described form of verbal analysis method on a computer.

In order to achieve the first technical problem, the Korean verb analyzing apparatus according to an embodiment of the present invention finds a prototype by using surrounding context information. Peripheral context data used for verb prototyping is extracted from examples of conjugations and surrounding contexts in which the verbs can be derived in one case without actual significance. According to an embodiment of the present invention, a verb analysis apparatus having a morphological neutrality maps a neutral problem to a problem of the same area without neutrality, and when a verb conjugation type without neutrality is used, the label is not labeled. From unlabeled raw corpus, we can know in advance about the class of the verb, that is, the prototype, so we can use it like a learning cork with class information already attached.

In order to achieve the second technical problem, a verb analysis method having morphological significance according to an embodiment of the present invention includes: defining a verb prototype set V; Defining a verb conjugation set E _{A in} which morphological significance is expressed during conjugation and a conjugation set E _{u in} which no primitive appears in the use of the prototypes; E _u A basic data collection step of constructing training data D _train by extracting relevant sentences from examples of the set; Implementing, according to a machine learning method, a classifier for assigning a verb prototype class using a surrounding context that appears with each verb prototype from the collected data as a learning feature; And constructing an experimental data set D _test from sentences including the E _A word.

According to another embodiment of the present invention, a method of analyzing verbs with morphological significance is an optional method of extracting a hard example set D _hard which is difficult to classify to experimental data using a classifier derived from the training data. A selective sampling step; Obtaining the count to the web information retrieval for _hard D includes the step of re-learning the classifier include the D D _hard to _train more.

According to embodiments of the present invention, by providing a verb analysis method that can have the effect of teacher learning (supervised learning) without the artificial classification of the corpus can improve the accuracy and efficiency of text analysis. In addition, when applying verb prototype classifiers derived from data learned to date to new data, the training set is added to the existing training data set only for difficult examples that differ from previously learned data or whose classifiers are unclear. You can automate the entire process of verb analysis by presenting incremental learning methods that increase the number of points of interest and web data together.

1A is a diagram of a morphological neutral verb analyzing apparatus according to an embodiment of the present invention.
1B is a diagram of a morphological neutral verb analyzing apparatus according to another embodiment of the present invention.
2 is a flowchart of a method of analyzing morphological neutral verbs according to an embodiment of the present invention.

Hereinafter, with reference to the drawings will be described a preferred embodiment of the present invention. However, embodiments of the present invention illustrated below may be modified in various other forms, and the scope of the present invention is not limited to the embodiments described below.

Verbs can vary in tenses, phases, and other grammatical categories, especially in verb endings. Table 1 shows verbs with various endings.

word Stemming went
Went away
Goes
going
Going
Going (Verb) + (former ending premature ending) + c (ending ending)
Ga (verb) + sour (past term tense fresh ending ending) + da (terminating ending)
A (verb) + b (present tense ending ending) + multi (ending ending)
A (verb) + is (a tubular malleable mother)
(Verb) + flag (noun-type malleable mother)
A (verb) + high (equivalent connection ending)

Table 2 shows examples of verb conjugations where the analysis shows morphological significance, examples of its possible morphological analysis, and the non-neutral conjugations of the prototype.

Although the number of verb prototypes that show morphological significance in use is not large (about 12), the utilization forms take up a considerable part, and the verbs are used in many different meanings, and there are many important words that appear frequently in Korean text.

VVMA is deeply involved in resolving the verbality of verbs, even though it is formally. In general, in a given context or discourse, words represent only one meaning (Yarowsky 1995), so words in the surrounding context can provide clues to resolving the significance of the word.

In addition, not all verb conjugations have morphological significance. For example, in the case of "selling," there is neutrality, but in the case of "selling" or "digging," there is no neutrality in the form of "sell" and "dig." Therefore, an embodiment of the present invention extracts the surrounding context of examples that can be derived from the verb utilization form without neutrality and uses it to solve the utilization type problem with the neutrality.

In one embodiment of the present invention, as learning data, after learning the surrounding context and verb prototype of unneutral conjugations from an unlabeled corpus, looking at the surrounding context of conjugation forms in which neutrality occurs, and similar to the previously learned context. Assign a prototype class. That is, a conjugation type without neutrality can be used as learning data for words whose base form is unclear. In this process, contextual information appearing with each basic type can be collected.

1A is a block diagram of a verb analyzing apparatus 100 having conformational importance according to an embodiment of the present invention.

The basic data collection unit 120 collects learning data including contextual features related to verb conjugations (hereinafter, 'first verb utilization'), which can be seen in the raw corpus 110. In this process, basic learning data is generated. The contextual feature to be considered is composed of the two words appearing before the word from which the neutrality is generated, the content words of each of the two words, and the survey information are used as the contextual feature vector to form the learning data 130.

학습 방법을 이용한다. The training performer 140 obtains a conditional context probability between the context information and the verb prototype to determine each basic form using a machine learning method based on the training data 130. The training performer 140 learns a classifier for determining the basic type based on the context information having the highest probability of being used with the basic type. In one embodiment of the present invention convenient re-learning

Use the learning method.

Table 3 shows an example of the context feature features in the training data used to distinguish between the basic types "listen" and "s".

Contents
( word _One ) Contents
( word ₂ ) Word _One Word ₂ Tone
( word _One ) Check
( word ₂ ) Class
( output ) Production cost A lot of Production cost A lot of (Current Investigation) field I story of mine Stories To Listen Equal story same Stories To Listen point Reason Point For a reason of in field rapidly line rapidly Shrink field

"Class" means the base type that is output when using the context characteristic feature of Table 3.

The verb analyzing unit 150 uses the classifiers learned in the training performing unit 140 as a part for judging the basic type of the verbal form of the verb included in the corpus 110 (hereinafter, 'second verb utilization'). Figure out the basic form for the new example. When an arbitrary text is input, the verb analyzing unit 150 extracts a vector required for learning such as a search, a word, a content word, etc. from the surrounding words existing in front of each verb in the input text.

Judgment of the basic forms of verbal usage in a neutral form plays an important role in the reading of text meaning. Therefore, the verb analysis apparatus 100 of FIG. 1A may be applied to a search engine, a translator, and the like.

1B is a block diagram of a verb analyzing apparatus having form neutrality according to another embodiment of the present invention.

The basic data collector 120 collects the training data 130 including the contextual feature related to the first verb utilization in the raw corpus 110, and the training performer 140 is based on the training data 130. While learning the basic type classifier according to the context, the verb analyzing unit 150 determines the basic type of the second verb utilization included in the corpus 110.

In general, the more contextual information collected by expanding the learning data, the better the accuracy of verb analysis. In addition, the learning domain may be changed or new words may be required to acquire new learning data. In this case, a method of using a web document may be considered as a method for extracting contextual information. But web pages are too big to download. Table 4 shows an example of the web counts of "black" and "color", one of the words "close" containing neutrality and one of the non-neutral conjugations that can be used to solve it. "Daterange" outputs the results limited to documents published in a specific period. Table 4 shows the results of limiting the search to documents within a month.

Word Total count Daterange counts Pick
Laying
How
accident
live
living 492,000
1,200,000
1,020,000
37,000,000
9,200,000
21,500,000 70,500
122,000
156,000
246,000
247,000
248,000

Therefore, the difficult case determination unit 160 does not extend the training data for all cases through selective sampling, but selectively extends the training data only for the cases that are difficult to classify by the current trained classifier. Accordingly, another embodiment of the present invention can increase the accuracy of the basic prediction while avoiding repeated data addition and unnecessary data size.

Hereinafter, by extracting the surrounding context for the second verb utilization, the first basic type predicted with the first high probability value and the second basic type predicted with the second highest probability value are identified by using the existing trained classifier. The difficult case determination unit 160 calculates a difference between the first highest probability value and the second highest probability value. The difficult case determination unit 160 may be configured to determine whether to further collect new learning data for the second verb utilization based on the confidence probability value. This probability value may be calculated in the form of Equation 1.

분류자를 사용한 예이다.Where p (c | x ) is the probability that case x is predicted as a category (basic type) c, and x is the context feature set described above. In order to predict the p value

Here is an example using a classifier.

In Equation 1, c _i is the closest category (basic type) to x corresponding to the context feature vector of the unanalyzed word, and c _j is the second closest category (basic type) to x. The confidence level when classifying with is the confidence value of Equation 1.

As shown in Equation 1, the greater the difference between these two probabilities, the greater the confidence of the classifier for the data. According to an embodiment of the present invention, the case of low confidence (verb) of classification is defined as hard examples, and the learning data is extended based on this.

In addition, the difficult case determination unit 160 may determine whether to add the training data based on the similarity between the training data to be added and the training data collected so far. It can be calculated as Equation 2 by defining this as a diversity value. This diversity value is inversely proportional to the similarity. If this value is reflected, it can be judged as a difficult case by assigning a high value to the example of the form different from the existing data learned so far.

Here, t _i means an instance to be added to the training data. K is a function that calculates the similarity between two input cases. If S is a set of cases used in existing learning, s _j means the most similar to t _i among the existing cases in the training data. The diversity value reflects the similarity between t _i and s _j . Find out how similar the existing training data is with the data that is currently being added.

In another embodiment of the present invention, an example in which the diversity value has the greatest value, that is, an example different from the existing learning data (the least similarity) is defined as a difficult case to expand the learning data. A combination of confidence and diversity values for about 3% of the experimental data is extracted as a difficult case.

Table 5 shows examples of difficult cases extracted from the analysis of "black" and "color" and examples of their contextual features.

hard examples Context feature Correct Category Security program
Fuck her in the presence
In panties 4 Security program security program josa1 =
She in front of her josa2 in front of her
From josa1 = in pantyham 4 Install
Slate
Install

If the communication unit 170 is determined to be a difficult case through the difficult case determination unit 160 and adds the training data for the second verb utilization, the communication word network that combines the context characteristics and possible basic forms around the second verb utilization is a query language. Or type into a search engine connected to the Internet. When the communicator 170 receives the web count from the search engine, the communicator 170 transfers the web count to the additional training performer 180 to find a basic type for the current case (difficult case).

Table 6 shows queries using contextual features of difficult cases and the web count returned by the search engine Google in response to these queries.

The web count in Table 6 is calculated only for documents published in the previous three months. The query used for the web count extraction consists of a phrase that combines the contextual features of a difficult case (eg "program") with the verbal verb forms (eg "flat") used in this context. do. That is, the usage forms of the verbs such as "color", "carb", "carb", and "carb" in Table 6 are used as examples for analyzing "car" or "car". In the query phrase "program * laying" including a wildcard (*), various phrases such as "laying a program", "laying only a program", "laying a program unconditionally" and the like can be matched with the query. That is, examples including conjugations such as "car", "car", "carouse only", and "cargo" can be used to train the context of the base form.

The additional training performer 180 expands the learning data for the difficult case. The additional training performer 180 generates training data for second verb utilization using the web count received through the communication unit 170 and adds the training data to the database 130.

Meanwhile, the analysis apparatus 100 of FIG. 1B may also include a verb analyzer 150 as in FIG. 1A.

2 is a flowchart illustrating a verb analysis method in which morphological neutrality exists according to an embodiment of the present invention.

First, in the raw corpus collect the learning data that is the contextual feature associated with the utilization type of the first verb utilization (S210).

Next, the basic type of the second verb utilization included in the corpus is determined based on the learning data (S220). More specifically, this process (S220) may be a process of determining the basic type of the second verb utilization based on the context information having the highest probability of being used in connection with the second verb utilization.

The contextual feature used in the process of determining the basic type of the second verb utilization is added as the learning data for the second verb utilization (S230).

Finally, when any text is input, the basic form of the verb included in the text is determined by comparing words existing in front of each verb in the text with the training data (S240). Words used in this process (S240) are the first two words and search, and the content words, the context words associated with each verb. For example, the process (S240) may be a process of determining the basic form of each verb by extracting two nouns located closest to each verb and comparing it with the learning data.

The invention can be implemented via software. Preferably, a program for executing a verb analysis method having morphological significance according to embodiments of the present invention may be provided by recording a program for executing in a computer on a computer-readable recording medium. When implemented in software, the constituent means of the present invention are code segments that perform the necessary work. The program or code segments may be stored on a processor readable medium or transmitted by a computer data signal coupled with a carrier on a transmission medium or network.

Computer-readable recording media include all kinds of recording devices that store data that can be read by a computer system. Examples of the computer readable recording medium include ROM, RAM, CD-ROM, DVD 占 ROM, DVD-RAM, magnetic tape, floppy disk, hard disk, optical data storage, and the like. The computer readable recording medium can also be distributed over network coupled computer devices so that the computer readable code is stored and executed in a distributed fashion.

Although the present invention has been described with reference to the embodiments illustrated in the drawings, this is merely exemplary and will be understood by those skilled in the art that various modifications and variations may be made therefrom. And, such modifications should be considered to be within the technical protection scope of the present invention. Therefore, the true technical protection scope of the present invention will be defined by the technical spirit of the appended claims.

Claims (14)

A basic data collection unit for collecting learning data including a first verb utilization without a neutrality in which a basic form is identified in an unlabelled raw corpus and a contextual feature related to the first verb utilization; And
Extracts the context features of the two preceding words in which the basic form is the closest to the second verb usage, and based on the extracted context features and the context features of the learning data, A training performer learning a classifier for matching with the basic type of the first verb utilization
Containing, morphological neutral verb analysis device. The method of claim 1,
When the arbitrary text is input, it further comprises a verb analysis unit for extracting the contextual feature existing in front of each verb in the text and using the classifier to determine the basic form of each verb Verb analysis device. The method of claim 1,
The training performer
And wherein in terms of morphological significance, the basic type of the second verb utilization is determined based on the conditional context probability of the context information associated with each basic type. The method of claim 1,
The confidence level of classification using the difference between the probability values for the first basic type predicted with the first highest probability value and the second basic type predicted with the second highest probability value when the basic form of the neutral verb is predicted by the training performer. And calculating a similarity between the experimental data and the learning data collected to date, and then determining whether to add the training data based on the confidence degree and the similarity. , Morphological neutrality verb analysis device. The method of claim 4, wherein
In order to grasp the basic type of the difficult case in which the difficult case determination unit decides to add the learning data, the search engine transmits a query phrase using a context feature of two preceding words located closest to the difficult case to the search engine. A communication unit for receiving a web count from the communication unit; And
And an additional training performer configured to learn the classifier by extracting the basic type of the difficult case using the web count and adding it as learning data. Collecting, as learning data, the contextual features associated with the primitive from the example of the first verb utilization with no neutrality to identify the primitive in the unlabeled raw corpus;
From the corpus, the basic feature of the second verb utilization is extracted by matching the basic feature of the second verb utilization based on the extracted context feature and the context feature of the learning data. Determining; And
And adding contextual features used in determining the basic form as learning data for extracting the basic form of the second verb utilization. The method according to claim 6,
If any text is input, the method may further include determining a basic form of the verb utilization using a probability between the word and the basic form appearing before the verb usage to obtain the basic form in the text. Gender verb analysis method. The method of claim 7, wherein
The step of determining the basic form of verb usage
And determining the basic form by extracting the contents of the previous contexts and extracting the contents of the verbs. The method of claim 8,
The step of determining the basic form of verb usage
And determining the basic form of the verb utilization based on two words which are located closest to the verb utilization. The method according to claim 6,
Determining the basic type of the second verb utilization
And determining the basic type of the second verb utilization using a naive Bayesian classifier using the conditional context probability between the possible basic types of verb utilization and contextual information. The method according to claim 6,
The first basic type predicted with the first highest probability value and the second basic type predicted with the second highest probability value are identified for the second verb utilization, and the difference between the first highest probability value and the second highest probability value is determined. And determining whether or not to add the learning data for the second verb utilization based on the formal verbs verb analysis method. The method according to claim 6,
Calculating a similarity between the learning data added to identify the basic type of the second verb utilization and the learning data collected so far, and determining whether to add the learning data for the second verb utilization based on the similarity; Morphological neutral verb analysis method, characterized in that it further comprises. The method of claim 12,
In order to identify the basic type of the difficult case in which the learning data is added in the step of determining whether to add the learning data, a query phrase using a context feature of two preceding words located closest to the difficult case is added to the search engine. Delivering; And
And extracting the basic type of the difficult case using the web count returned from the search engine and adding the learning type to the training data. A computer-readable recording medium having recorded thereon a program for performing the method of claim 6.

Images