KR20210023640A - Transition-based Korean Dependency Analysis System Using Semantic Abstraction - Google Patents

Abstract

The present invention relates to a transition-based Korean dependency analysis system, which enables high-order words to be extracted only by learning a relatively small number of sentences through semantic abstraction, and derives dependencies in extended sentences. The system of the present invention includes: a morpheme analysis module for analyzing a morpheme of an inputted sentence; a language abstraction module for acquiring surface-type abstract information of morpheme from the analyzed morpheme information from the database; and a dependency analysis module for analyzing the dependency relationship between words using the abstract superficial information of the language and the quality information learned from the learning corpus.

Description

Transition-based Korean Dependency Analysis System Using Semantic Abstraction}

The present invention relates to a language dependence analysis system, and more specifically, a transfer that allows a higher level language to be extracted only by learning a relatively small number of sentences through semantic abstraction to derive a dependency relationship in an extended sentence. It is about a Korean-based dependency analysis system.

In the analysis of dependency relations between Korean, especially Korean, learning-based research is dominated among rule-based research and research based on learning corpus. The corpus for the analysis of dependence contains information on the dependency relationship between words, and the word is expressed as a surface layer of morphemes and part-of-speech tag information.

Therefore, the combination of the surface-of-speech tag and the morpheme of the learning corpus is learned. Among them, the part-of-speech tag has a high reproducibility, but there is a limit to analyzing the dependency relationship only with the part-of-speech tag. On the other hand, in the case of the superficial type, when the learned content is reproduced, it shows a high accuracy rate, but it is difficult to generalize the combination of the superficial type constituting the word because there are many things, and the reproducibility of the learning quality is lower than that of the POS tag. Therefore, it is difficult to learn all combinations of the superficial form, and even if the superficial form included in the learning corpus is learned, the probability of reproducing the superficial form in the test sentences input after learning is low.

As such, the existing learning-based research judges the presence or absence of a dependency relationship based on the combination of the surface type of morpheme and the POS tag, so there is a limit to accurately determining the dependency relationship of Korean sentences with various sentences due to changes in word order and omission of essential components. That is, because Korean has free word order and can omit the constituent elements of sentences.

The methods used to analyze the rules of dependence in Korean have been proposed in various ways. For example, in rule-based dependency analysis, it is possible to accurately analyze relatively short sentences, but if the length of the sentence is longer and changes and omissions in word order are compounded, exception rules for defining dependency relationships are required. These rules have the advantage of being able to determine dependency relationships with defined rules even without a learning corpus, and also having a fast processing speed. However, it is difficult to define rules that can accurately analyze all sentences, and as exception rules are added, it is difficult to manage the defined rules because it is necessary to directly determine the correlation with the existing rules.

Another example is transition-based dependency analysis. The transfer-based dependency analysis model was proposed by'Joakim Nivre' and is a relatively simple algorithm. For the study of Korean dependency analysis using the transfer-based model, the arc-based corpus of CoNLL-U (the conference on computational natural language learning) was used. There is a study using the eager algorithm. In this study, we experimented using SVM (support vector machine) as a learning algorithm for features, and a corpus of 2,532 sentences was used. The experimental results showed that the size of the corpus was small due to the characteristics of the corpus used, and the error rate was high, and the dependence accuracy rate of 63.39% was low using only the limited feature information in the word unit.

Another example is deep learning-based dependency analysis. The deep run-in-based method uses a feed-forward neural network as the oracle role of the arc-eager algorithm, and studies using embeddings such as NNLM (neural network language model) and Word2Vec, and RNN (recurrent neural network) models. A study using long short-term memory (LSTM) cells, a study using an attention technique using an end-to-end neural network model, a study using a transition based recurrent unit (TBRU) combining a transition-based algorithm with a neural network cell, There are studies using a pointer network model using GRU (gated recurrent unit) cells, and studies using biaffine attention techniques. These methods show a high dependence accuracy rate of about 90% in Korean, but are required for learning and processing. The time is larger than that of the machine learning method, and it is difficult to accurately analyze the cause in error analysis, so there is a disadvantage that an empirical error analysis is required to confirm the result after retraining by indirectly modifying the model.

Therefore, the present invention abstracts the surface form of a noun by using a pre-built vocabulary semantic network (UWordMap) to increase the reproducibility of the surface form of morphemes and the accuracy of the dependence relationship, and analyzes the transformation-based Korean dependence using semantic abstraction that is used as a feature. It is to provide a system.

In addition, this transition-based Korean dependency analysis system classified the learning corpus for superficial abstraction, and abstracted nouns according to the word hierarchy of the vocabulary semantic network based on the separated information. In consideration of the learning speed, a transfer-based analysis model was applied.

The present invention for achieving the above object, the morpheme analysis module for analyzing the morpheme of the input sentence; A body language abstraction module for acquiring surface-layered abstracted information of body language from the analyzed morpheme information from a database; And a dependency relationship analysis module that analyzes the dependency relationship between words using abstracted superficial information of the body language and feature information learned from the learning corpus.

The morpheme analysis module provides a morpheme analysis result of the input sentence and a homozygous word classification result, and the morpheme analysis result is output by attaching a surface type, isomorphic synonym number, polyphony number, and part-of-speech tag in morpheme units, and the From the analysis results, the superficial and homozygous numbers and parts of speech tag information are collected in terms of words and morphemes of the input sentence.

The body language abstraction module uses a vocabulary semantic network to abstract the surface type of a noun and uses it as a feature, and the vocabulary semantic network provides a hierarchical structure of nouns in a tree structure, and provides superficial information of a noun through upper language information. Abstract.

In the superficial abstraction, the learning corpus is identified as homozygous, and nouns are abstracted according to the word hierarchy of the vocabulary semantic network based on the information identified by the homozygous language.

In the superficial abstraction of the noun, the basic features of the word unit features including the superficial features and the abstract features, morpheme units, and syllable unit expansion features are all learned.

The body language abstraction module identifies the body language morphemes by checking the parts of speech tag of the morpheme unit of the input sentence, and checks whether there is corresponding word information in the vocabulary semantic network for the body language morphemes, and the words not in the vocabulary meaning network are abstracted. For words that do not have information and have information, the word hierarchy of the vocabulary semantic network is calculated, all word hierarchies are calculated for all polymorphisms of the homozygous word based on the homozygous word, and the higher layer by the set abstraction level for all the calculated polymorphisms. It is abstracted into words, and a word with the highest frequency among abstracted words is determined as an abstracted word of a corresponding homozygous word.

The Cheon abstraction is a first method of abstracting into a layer 1 word under a root node when the abstraction level exceeds the highest level of the vocabulary semantic network when the level of abstraction exceeds the highest level, and without abstraction. It is carried out by one of the second methods of using the original as it is.

The dependency relationship analysis module inputs the order of input sentences into the input buffer in reverse order according to the transition-based model considering the characteristics of Korean, obtains the value of each feature through a feature function, and determines a transition. Dependence of all words in the input buffer is determined according to the transition decision, and if there are no words in the input buffer, the dependency analysis is terminated, and the position of the word of the dominant word of each word of the input sentence is recorded, and the dependency relationship result is output. Characterized in that.

In the evaluation of the basic feature performance according to the dependency analysis, the generalized POS tag feature provides the highest reproducibility, and the surface type and homogeneous objection number, and the POS tag sum feature (BPS) provide the highest accuracy rate. .

In the evaluation of the extended feature performance according to the dependency relationship analysis, the reproducibility and accuracy rate of the extended feature are higher than that of the basic feature, and the accuracy rate is low.

In the evaluation of the extended feature performance according to the dependency relationship analysis, the case in which both morpheme features and syllable features are used provides the highest dependency relationship accuracy rate.

The dependency relationship analysis method analyzes the transition-based dependency relationship using a modified backward arc-eager algorithm using the Korean dominant postfix principle, and the modified backward arc-eager algorithm is the right- The arc, shift transition is removed, and the left-arc transition does not remove the word of the stack and moves to the buffer.

The dependence relationship accuracy and the feature accuracy rate according to the dependency relationship analysis are calculated by the following equation.

,

,

According to the transition-based Korean dependence analysis system using semantic abstraction of the present invention as described above, by proposing a dependency relationship analysis method used as a feature by abstracting the surface type of nouns using a pre-built vocabulary semantic network, The reproducibility of the superficial layer and the accuracy of the dependence relationship are improved. In addition, for superficial abstraction, the learning corpus was identified as homozygous words, and nouns were abstracted according to the word hierarchy of the vocabulary semantic network based on the information identified by the homomorphic synonyms. In addition, a transfer-based analysis model was used in consideration of the analysis and learning speed, and was modified to suit the characteristics of Korean.

As a result, when evaluating using only superficial features through the superficial abstraction of nouns, the dependence accuracy rate increased by up to 7.55%. As a result of learning basic features with word unit features including abstract features, 90.43% accuracy rate and 1,002 sentences/sec. It showed processing speed, 1,294 sentences/sec learning speed. When learning both morpheme units and syllable unit expansion features, 90.75% dependence analysis accuracy, 631 sentences/sec, processing speed, and 562 sentences/sec learning speed were shown. In addition, compared to the rule-based study, it is possible to confirm the result of improved accuracy, learning speed, and processing speed.

1 is a block diagram of a transition-based Korean dependency analysis system using semantic abstraction according to an embodiment of the present invention
Figure 2 is a basic structure of a basic transition-based dependency analysis module (dependency parser) provided to explain the present invention
3 to 11 are graphs showing various indicators showing the experiment and performance evaluation of the present invention,
3 is a graph showing the accuracy and utilization rate of each step of the Cheon abstraction
4 is a graph showing feature reproducibility for each abstraction step of the'one-step abstraction method when the range is exceeded'
5 is a graph showing the accuracy rate of dependence for each abstraction step of the'one-step abstraction method when the range is exceeded'
6 is a graph showing feature accuracy for each abstraction step of the'one-step abstraction method when the range is exceeded'
Fig. 7 is a graph showing the accuracy of dependence by step of abstraction in the'method of using a circle when the range is exceeded'
Figure 8 is a graph showing the feature accuracy of each abstraction step in the'method of using a circle when the range is exceeded'
9 is a graph showing the feature accuracy of each abstraction step of'how to use a circle when the range is exceeded'
10 is a comparison graph of the accuracy rate of dependence for each learning feature
11 is a comparison graph of learning and processing speed for each learning feature

Objects and effects of the present invention, and technical configurations for achieving them will become apparent with reference to the embodiments described later in detail together with the accompanying drawings. In describing the present invention, if it is determined that a detailed description of a known function or configuration may unnecessarily obscure the subject matter of the present invention, a detailed description thereof will be omitted.

In addition, terms to be described later are terms defined in consideration of functions in the present invention, which may vary according to the intention or custom of users or operators.

However, the present invention is not limited to the embodiments disclosed below and may be implemented in various different forms. However, the present embodiments are provided to complete the disclosure of the present invention and to fully inform the scope of the invention to those of ordinary skill in the art, and the present invention is defined by the scope of the claims. It just becomes. Therefore, the definition should be made based on the contents throughout the present specification.

Hereinafter, the present invention will be described in more detail based on the embodiment shown in the drawings.

The vocabulary information used in the present invention uses'UWordMap' of the University of Ulsan, which is built with a multilingual level vocabulary semantic network based on a standard Korean dictionary, and the'UWordMap' refers to vocabularies such as nouns, verbs, adverbs, etc. Speaks the vocabulary. In addition,'UWordMap' contains the hierarchical structure of nouns in a tree structure, so information on the upper or lower words of the noun can be obtained. Since the high-level word information has a larger category meaning than the target word, the superficial information of a noun can be abstracted through the high-order word information. By selecting the number of nodes based on the highest node using the tree structure, the superficial abstraction layer can be determined, and the number of layers to be abstracted based on the current node can be determined.

For example, the surface type of'apple__05' (apple) included in'UWordMap' can be abstracted into the first layer word'object' or the 5th layer word'fruit' according to the abstraction level as shown in Table 1 below. .

Word hierarchy Word superficial 0 UWIN One stuff 2 object 3 matter 4 Fruit 5 fruit 6 Apple

Table 2 below shows an example of how the surface layer of sentences according to the abstraction stage is actually abstracted. The abstraction level represents the number of lower level hierarchies based on UWIN (highest word) in the word layer of UWordMap, and if the level of the word surface type of the original text is a word at a lower level than the abstraction level, it is expressed as an abstraction level word.

Original text Usage fee Is basic Call charges To 3 minute Per 2 back ∼3 Hundred won Is added. Level 1 abstraction Stuff is added to the two units or three units of 3 units per unit of the properties object. 2 level abstraction Goods will be added to the three languages Language Unit 2 to 3 pound per language unit based on the article. 3 level abstraction Wealth is added to the end of two to three per malwon 3 on the end of the present money. Level 4 abstraction Money is added to the two words or three words per source word in 3 fundamental money. Level 5 abstraction For the fee , 2 parts of speech ~ 3 parts of speech per 3 parts of speech are added to the basic rate.

When abstracting the surface type of morpheme, the same surface type can have different abstraction trees. These words are semantically different even if they have the same surface type because of the difference in homozygous or polymorphisms in the corresponding vocabulary.

Therefore, in the present embodiment, a corpus with homozygous classification was used, and if there is no homozygous tag in the test sentence, a homozygous tag attached using a morpheme analyzer (UTagger) was used. As shown in Table 3 below, when an apple and an apology of the same surface type are classified and abstracted, different words are abstracted according to the abstraction layer.

Word hierarchy Word superficial 0 UWIN UWIN One stuff Act 2 object origin 3 matter Apple_08(apology) 4 Fruit 5 fruit 6 Apple_05

It looks at the learning qualities applied to the present invention. Learning qualities can be divided into basic learning qualities and extended learning qualities.

The basic learning features include a combination of morphemes of part-of-speech tags and superficial layers, and information such as adjacent words and relative distances of the stack and the top of the buffer are used as features, and the word unit is used. We selected the qualities to learn as a standard. Among the learning qualities, the generalized part-of-speech tag generalizes the part-of-speech combination with low reproducibility to increase the reproducibility. In addition, to determine the long term distance, the number of verbs between the target words was learned, and the position of the word to be learned is the previous or the next word based on the top word along with the top word of the stack and buffer. stack top We learned the parts of speech tag of the dominant word of the word.

The reason for limiting the range of the target word is that as the range of nodes to be referenced is widened, the more nodes information is referred, the more noise is generated in the learning information, resulting in a lower dependence accuracy rate than the case where the range is limited. The basic feature was set to arc (dependency setting), and the reason for doing this is that if all learning features are not reproduced, it is highly accurate to establish a dependency relationship with an adjacent node by the probability of the dominant distance probabilistically.

Table 4 below shows examples of basic learning qualities and types of generalized parts of speech.

Existing Qualitative Elements [ stack top word, buffer top word ]: generalized POS tag,
body,
BSP(body +sense number +POS tag),
abstracted body,
stack first morpheme BSP + buffer POS,
stack POS + count of verb between buffer top word.
[ stack previous word, stack top word ], [ stack top word, stack next word ], [ stack top word, stack governor word ], [ buffer previous word, buffer top word ]: POS
default feature (arc) Generalized POS NNB, NNP, NP, NR, JKV, XPN, XSN, XSA, XR, SO, SE, SH, SL, SN, SW, SS (deduplication after generalization to NNG) body: surface type of morpheme
sense number: homozygous number
POS: POS tags in the form
BSP: Surface type of morpheme + homozygous number + part-of-speech tag

The extended learning feature is not a word unit, but a combination of morpheme units and syllable unit features. As an extended learning feature, in order to supplement the lost information through the generalization of POS tags and the superficial abstraction from the basic features, the original features were learned. For the extended features, all possible combinations were learned and the average value was calculated with the number of morphemes or the number of words in the word unit, and the probability of the word unit was calculated.

Table 5 shows examples of extended learning qualities.

Expansion qualities factor [ stack top word, buffer top word ]: POS combination,
morpheme POS with word POS combination,
body combination,
morpheme body with word body combination,
BSP combination,
morpheme BSP with word BSP combination,
syllable combination with morpheme POS body: surface type of morpheme
sense number: homozygous number
POS: POS tags in the form
BSP: Surface type of morpheme + homozygous number + part-of-speech tag

1 is a block diagram of a system for analyzing a transition-based Korean dependency relationship using semantic abstraction according to an embodiment of the present invention. Looking at this, the analysis system 100 is a configuration including a sentence input module 110, a morpheme analysis module 120, a body language abstraction module 130, a dependency relationship analysis module 140, and a dependency relationship output module 150. .

In the present embodiment, the sentence input module 110 is a module for inputting a series of sentences for analyzing a dependency relationship, and may use various media.

In the present embodiment, the morpheme analysis module 120 analyzes the morpheme analysis result of the input sentence and the homozygous word classification result using the morpheme analyzer 122, for example,'UTagger'. Examples of such analysis results may be shown in Table 6.

Input sentence I saw the beautiful younger brother of Younghee. morpheme
Analysis Beautiful__000001/VA+ㄴ/ETM Younghee/NNP+U/JKG
Brother__010001/NNG+/JKO Bo__010101/VV+D/EP+Da/EF+./SF

Looking at this, it can be seen that the morpheme analysis module 120 attaches and outputs a surface type, a homogeneous objection number, a polyphonic number, and a part of speech tag in units of morphemes. The synonym number and part-of-speech tag information are collected.

In this embodiment, the body language abstraction module 130 acquires surface-type abstraction information of the body language by using the vocabulary semantic network 132, for example, information of the'UWordMap' of Ulsan University. Identify the part-of-speech tag of the morpheme unit of the input sentence to identify the morphemes of the body language, and check whether there is information about the word in'UWordMap' for the morphemes of the body language.

According to the result of this check, the word hierarchy of'UWordMap' is calculated for words with information without abstraction for words that do not belong to'UWordMap'. On the other hand, if there is a corresponding word, abstraction is performed on the basis of the homozygous word, so all word hierarchies are calculated for all polymorphisms of the homozygous word. This is because one homozygous synonym can have two or more polymorphisms. In addition, it abstracts all polymorphisms into higher-level words as much as the set abstraction level, and determines the word with the highest frequency among the abstracted words as the abstracted word of the corresponding homozygous word.

The dependency relationship analysis module 140 determines the dependency relationship using the abstracted superficial information of the body language in the body language abstraction module 130 using the dependency relationship learning information db 142 and feature information learned from the learning corpus. do. In addition, in the analysis process of the dependency relation analysis module 140, the order of input sentences is placed in the input buffer in reverse order according to the transition-based model in which the characteristics of Korean are considered, and the value of each feature is obtained through the feature function. (transition) is determined. In addition, depending on the transition decision, all words in the buffer are dependent on dependence. If there are no words in the buffer, the dependency analysis is terminated, and the position of the dominant word of each word in the input sentence is recorded, and the dependency relationship result is recorded. Print it out.

2 is a basic structure of a basic transition-based dependency parser 140 provided to explain the present invention. As shown, a stack 144 and an input buffer 145 are provided, and a parser 146 is included to perform a transition from an Oracle function. In addition, the Oracle function for determining the transition of each step is determined as the sum of the probabilities in which the weights for each learning feature are reflected through Equation 1 below.

, here,

,

The feature function f(x) was changed from 0 to 1 in the range of 0.5 to -0.5 by reducing the ratio of arc action to the ratio of all actions of features learned in the learning corpus by half of the total action. Therefore, when the ratio of arc count is positive, the probability of arc action is higher than that of reduce action, and when it is negative, it is reversed. In order to have a lower effect on the sum of the probabilities of features, the probability value of low frequency according to the frequency of action is multiplied by the logarithm of the total action count to calculate the transition probability. When taking the logarithmic function, the minimum value α of 1 or more is summed to prevent sign reversal.

And the decision of the action is to normalize the probability value of the feature multiplied by the weight of each feature from the result of the Oracle function with a hyperbolic tangent function, and change the final value into the range of -1 to 1. When the result value is positive, the dependency relationship is determined through the left arc action, and when the result value is negative, the dependency relationship is determined by the reduce action.

The weight of each feature is determined by changing the sentence order of the learning corpus in a random order and learning iteratively. As for the weight change value of each feature, the correct answer value CA (0.5 or -0.5) and the error value E of the oracle result value are calculated as shown in Equation 2 below. It was determined by multiplying the value and the learning ratio α to determine the change value of each weight for the error.

Meanwhile, in this embodiment, the transfer-based dependency relationship was analyzed using a modified backward arc-eager algorithm that considers the dominant post-secondary principle of Korean. Therefore, as shown in Table 7, the right-arc and shift transition are removed from the basic arc-eager algorithm, and the left-arc transition is moved to the buffer without removing the word of the stack.

을

와 같이 역순으로 변경해야 할 것이고, 표 8과 같이 나타낼 수 있다.And in order to use the backward arc-eager algorithm, the input sentence

of

It will have to be changed in the reverse order as shown in Table 8.

Input buffer Input buffer [0] [One] [2]
⇒ [0] [One] [2] I rice ate. ate. rice I

Next, it looks at the performance evaluation of the present invention. The performance evaluation analyzed the reproducibility and accuracy of learning features according to the superficial abstraction level, and analyzed basic feature performance and extension quality performance.

For the performance evaluation, in this example, the corpus used in the experiment and learning used the corpus converted from the 21st century Sejong Plan syntax analysis corpus into the dependent grammar, and was used by attaching an additional homozygous tag. In addition, the dependency relations label was not used and only the dependency relations were analyzed during the syntax analysis. Therefore, in the accuracy rate evaluation, only UAS (unlabeled attachment score) evaluation is performed. In addition to the accuracy rate of the dependence analysis, the learning speed and test speed according to the learning qualities were also evaluated, and the dependence accuracy and speed were also evaluated.

In addition, the corpus excluded sentences with a mismatch in the number of words due to a spacing error, sentences with cross dominant relationships, and sentences with self-circulating dominant words in the middle of the sentences. The refined sentences consist of 57,265 sentences, and a total of 561,645 words. For the performance evaluation, 10% of the entire corpus was tested using a 10-fold cross-validation method, and the process of learning the rest was repeated. The dependence relationship accuracy rate was calculated by dividing the number of dependent relationship words correctly analyzed in the entire corpus by the total number of words as shown in Equation 3 below.

And the accuracy rate of the feature is calculated by dividing the number of transition decisions of the feature, such as the transition for determining the dependence relationship, by the number of reproductions of the feature in order to determine the accuracy of the dependency relationship of the feature when the feature is used to determine the dependency relationship as shown in Equation 4 below. I did.

Meanwhile, the system environment in which the experiment was performed is shown in Table 9.

operating system Windows 10 CPU Intel®Core™i7-5820K @ 3.30 GHz RAM 32 GB Development language and tools C#, .Net Framework 4.5.2, Visual Studio 2017

The results of performance experiments through such a system will be described.

First, it is the experimental result of semantic abstraction performance of vocabulary. If dependency analysis is performed by learning only superficial information as a learning feature in the learning corpus, it is possible to compare the reproducibility and accuracy of the learning features according to the superficial abstraction level.

As a result of comparison, it was found that the dependence analysis accuracy rate increased proportionally according to the use rate of the feature, and the feature accuracy rate decreased due to the inverse relationship with the degree of abstraction. For example, as shown in Table 10, when abstracting the surface type of Cheon to the unit layer 1 level of UWordMap, the accuracy rate of the surface type feature decreases by 2.92% due to the lost information, but the recall rate of the feature increases by 36.75%. The silver was improved by 7.55%.

Dependency accuracy rate Surface feature recall Quality accuracy Original text 73.95% 18.33% 98.26% abstraction 1 level 81.50% 55.08% 93.34% △ +7.55% +36.75% -2.92%

3 is a diagram showing a change in the accuracy rate and the usage rate for each step of body language abstraction in the experimental result of the present invention. From this, it can be seen that as the abstraction step progresses, the feature accuracy rate slightly increases, but the dependency relationship accuracy rate and feature recall rate decrease.

Cheon's abstraction method is relatively abstracted based on the word layer of UWordMap, and when the abstraction level exceeds the highest level, it can be divided into two methods according to the processing method. If the abstraction level is higher than the highest level of UWordMap, the '1st level abstraction method when the range is exceeded' is abstracted with the word 1 layer under the root node (UWIN), and'when the range is exceeded, the prototype is used as it is without abstraction. There's a way.

When the dependence accuracy rate and the feature reproducibility according to the abstraction step processing method exceed the range, the method of abstracting with the lowermost word as shown in Figs. 4 to 6 is the reproducibility of the feature and the dependence accuracy rate according to the abstraction step. The relationship was proportional, and the quality accuracy was inversely proportional.

In addition, as shown in Figs. 7 to 9, the'method of using prototypes when the range is exceeded' increases the number of words that are not abstracted beyond the highest level as the level of abstraction increases, so that the dependence accuracy and feature reproducibility decrease after increasing. Shows the pattern.

The second is the evaluation of basic quality performance.

Calculate dependence accuracy, processing speed, learning speed, etc. using basic features, and evaluate the performance with the average value through 10-fold cross validation. As shown in Table 11, the dependency accuracy rate and processing speed were compared with the dependency relationship analyzer constructed by the rule-based method. Both models showed stable accuracy rate and processing speed, and it was shown that the proposed model based on learning improved both the accuracy rate and the processing speed of the dependency relationship compared to the rule based model.

10-fold index Transition-based proposal system Rule-based analyzer Accuracy (%) Processing speed (sec) Learning speed (sec) Accuracy (%) Processing speed (sec) One 90.62 5.507 41.110 88.23 14.653 2 90.26 7.439 39.731 88.20 12.103 3 90.37 5.589 40.606 88.43 9.321 4 90.54 6.671 41.612 88.21 11.336 5 90.57 5.314 39.235 88.66 11.488 6 90.27 5.282 38.335 88.50 11.240 7 90.32 5.363 40.556 88.53 11.139 8 90.37 5.362 38.830 88.07 9.073 9 90.50 5.267 40.315 88.13 11.234 10 90.44 5.386 38.029 88.48 11.348 Average 90.43 5.718 39.836 88.34 11.293 △ +2.09 -5.575
(-97.5%)

In addition, when comparing the reproducibility and accuracy rate for each feature, as shown in Table 12, the generalized POS tag features the highest recall, and the surface type, homologue number, and POS tag sum feature (BSP) show the highest accuracy rate.

Qualities type Recall Accuracy rate generalized POS 98.87% 91.49% body 14.97% 94.64% abstracted body 19.14% 94.10% BSP 14.71% 94.96% stack first morpheme BSP + buffer POS 20.63% 94.05% stack POS + count of verb between buffer top. 98.76% 91.69% (stack prev, stack top) POS 97.33% 91.76% (stack top, stack next) POS 82.81% 90.87% (stack top, stack gov) POS 83.48% 90.95% (buffer prev, buffer top) POS 79.67% 91.82%

Third, look at the evaluation of extension quality performance.

As for the extended features, the performance evaluation was performed by adding the extended features to the basic features. As for the extended features, the morpheme unit combination and the syllable unit combination were compared, and the performance was evaluated using all the combination features. Both the morpheme unit combination feature and the syllable unit combination feature improved the accuracy rate, but the improvement rate was not significant. Combinations of extended features were learned for all possible combinations. The value of the feature function of the feature in the morpheme unit or the syllable unit was calculated by summing the results of all combinations and dividing it by the feature number of the buffer top word.

The following Equation 5 is an equation for calculating word quality of extended features, and Table 13 is a performance comparison table according to the use of extended features.

10-fold index Basic qualities + morpheme unit combination qualities Basic qualities + syllable unit combination qualities Accuracy (%) Processing speed
(sec) Learning speed
(sec) Accuracy (%) Processing speed
(sec) Learning speed
(sec) One 90.98 8.49 73.38 90.90 6.00 48.49 2 90.57 8.32 80.05 90.52 5.65 50.42 3 90.63 8.52 78.99 90.54 5.57 47.78 4 90.86 8.37 80.63 90.77 5.51 49.54 5 90.80 8.47 83.12 90.75 5.60 47.95 6 90.54 8.83 81.09 90.49 5.60 48.33 7 90.63 8.54 81.55 90.62 5.68 46.45 8 90.65 8.60 82.48 90.56 5.69 50.13 9 90.83 8.68 80.44 90.69 7.71 48.45 10 90.76 11.11 77.90 90.70 5.77 49.12 Average 90.73 8.79 80.46 90.65 5.88 48.67 △ +0.30 +3.08sec
(+53.8%) +40.63sec
(+101.98%) +0.23% +0.16sec
(+2.78%) +8.83sec
(+22.16%)

On the other hand, when looking at the reproducibility and accuracy rate of the extension quality by referring to the performance of each extension feature in Table 14, it can be seen that the reproducibility rate is higher than that of the basic feature, but the accuracy rate is lower.

Qualities type Recall (%) Accuracy (%) syllable combination with morpheme POS 90.30 82.13 morpheme POS combination 99.99 82.19 morpheme POS with word POS 98.58 82.87 morpheme body combination 71.42 82.52 morpheme body with word body 71.42 88.62 morpheme BSP combination 68.84 90.94 morpheme BSP with word BSP 14.82 90.93 morpheme BSP with word POS 83.56 86.05

In addition, looking at the performance of using the integrated features in Table 15, when both morpheme features and syllable features are used, the highest dependence accuracy rate is shown, but the improvement rate is not high.

10-fold index Basic qualities + morpheme combination qualities + syllable qualities Dependency accuracy rate (%) Processing speed (sec) Learning speed (sec) One 91.01 9.37 94.00 2 90.59 9.00 94.68 3 90.62 9.05 91.16 4 90.87 8.88 90.93 5 90.80 9.05 91.14 6 90.58 8.96 90.13 7 90.71 8.93 91.18 8 90.69 9.10 90.96 9 90.84 9.06 91.13 10 90.83 9.33 91.24 Average 90.75 9.07 91.66 △ +0.33% +3.35sec(+58.66%) +51.82sec(+130.08%)

Table 16 shows the performance comparison results for different features. We compared the dependence accuracy rate, learning and processing speed according to the feature selection. The basic qualities had the lowest accuracy, but took the shortest time in terms of speed. It can be seen that the highest accuracy rate is the integration feature using all features and the speed is the slowest.

Qualities of use Dependency accuracy rate (%) Processing speed (sec) Learning speed (sec) Basic qualities 90.46 5.72 39.84 Basic + syllable quality 90.65 5.88 48.67 Basic + morpheme quality 90.73 8.79 80.46 Integrated qualities 90.75 9.07 91.66

Graphs for comparing the accuracy rate, processing speed, and learning speed according to the use features of Table 16 are shown in FIGS. 10 and 11.

As described above, the present invention proposes a dependency relationship analysis system that abstracts the surface form of a noun by using a vocabulary semantic network and then uses it as a feature, and improves accuracy, learning speed, and processing speed compared to conventional dependency analysis methods. Can be seen. In addition, the present invention may be applied to various programs such as automatic translators, chatbots, and automatic Q&A.

Although it has been described with reference to the illustrated embodiments of the present invention as described above, these are only exemplary, and those of ordinary skill in the art to which the present invention pertains, without departing from the spirit and scope of the present invention, various It will be apparent that variations, modifications and other equivalent embodiments are possible. Therefore, the true technical protection scope of the present invention should be determined by the technical spirit of the appended claims.

110: sentence input module
120: morpheme analysis module
122: morpheme analyzer
130: Cheon abstraction module
132: Vocabulary Semantic Network
140: dependency analysis module
142: dependency relationship learning information DB
150: dependency output module

Claims (13)

A morpheme analysis module that analyzes the morpheme of the input sentence;
A body language abstraction module for acquiring surface-layered abstracted information of body language from the analyzed morpheme information from a database; And
Transition-based Korean dependency analysis system comprising a dependency relationship analysis module that analyzes dependence between words using abstracted superficial information of the body language and feature information learned from the learning corpus. The method of claim 1,
The morpheme analysis module,
Provides a result of morpheme analysis of the input sentence and a result of discrimination of homozygous,
The result of the morpheme analysis is output by attaching a surface layer type, a homozygous word number, a polyphonic word number, and a part-of-speech tag in units of morphemes,
Transition-based Korean dependence analysis system using semantic abstraction that collects superficial and homozygous numbers and parts of speech tag information in units of words and morphemes of input sentences from the analysis result. The method of claim 1,
The Cheon abstraction module,
Using the vocabulary semantic network, abstract the superficial form of nouns and use them as features,
The vocabulary semantic network provides a hierarchical structure of nouns in a tree structure, and a transition-based Korean dependency analysis system that abstracts superficial information of nouns through upper language information. The method of claim 3,
The superficial abstraction,
A transfer-based Korean dependency analysis system that identifies learning corpus as homozygous and abstracts nouns according to the word hierarchy of the vocabulary semantic network based on the information identified by the homozygous language. The method of claim 4,
The superficial abstraction of the noun is a transition-based Korean dependence analysis system that learns all of the basic features of the word unit features including the superficial features and the abstract features, morpheme units and syllable unit extension features. The method of claim 3,
The Cheon abstraction module,
Identify the parts of speech tag in the morpheme unit of the input sentence to identify the morphemes of body language,
Check whether there is information about the word in the vocabulary semantic network for the body morpheme,
Words that are not in the vocabulary semantic network are not abstracted, and words with information calculate the word hierarchy of the vocabulary semantic network,
All word hierarchies are calculated for all polymorphic words of the homozygous based on the homozygous synonym,
Abstracts all calculated multi-words into higher layer words as much as the set abstraction level,
Transition-based Korean dependency analysis system that determines the word with the highest frequency among abstracted words as abstracted words of the corresponding homologous word. The method of claim 6,
The Cheon abstraction module,
When the level of abstraction exceeds the highest level, when the level of abstraction is higher than the highest level of the vocabulary semantic network, the first method of abstracting with a layer 1 word under the root node, and the second method of using the original as it is without abstraction. Transition-based Korean language dependency analysis system performed by one of the. The method of claim 1,
The dependency relationship analysis module,
The order of the input sentences is input into the input buffer in reverse order according to the transition-based model considering the characteristics of Korean,
The transition is determined by obtaining the value of each feature through the feature function,
According to the transfer decision, all words in the input buffer are dependent on the relationship. If there is no word in the input buffer, the dependency relationship analysis is terminated, and the position of the dominant word of each word in the input sentence is recorded to determine the dependency relationship result. Transition-based Korean language dependency analysis system for output. The method of claim 8,
In the evaluation of basic feature performance according to the dependence analysis, the generalized POS tag feature provides the highest reproducibility, and the surface type and homogeneous objection number, and the POS tag sum feature (BPS) provide the highest accuracy rate. Transition-based Korean dependency analysis system. The method of claim 8,
In the evaluation of the extended feature performance according to the dependency analysis, the reproducibility and accuracy of the extended feature are higher than the basic feature, and the accuracy rate is low. The method of claim 8,
In the case of using both morpheme features and syllable features in the evaluation of extended feature performance according to the dependency relationship analysis, a transition-based Korean dependency relationship analysis system that provides the highest dependency relationship accuracy rate. The method of claim 1,
The analysis method of the dependency relationship analysis module,
Transition-based dependency is analyzed using the modified backward arc-eager algorithm using the Korean dominant postfix principle,
In the modified backward arc-eager algorithm, the right-arc and shift transitions are removed from the basic arc-eager algorithm, and the left-arc transition is analyzed by moving to the buffer without removing the word of the stack. system. The method of claim 12,
Transition-based Korean dependence analysis system in which the dependence relationship accuracy rate and the feature accuracy rate according to the dependency relationship analysis are calculated by the following equation.

,

Images