KR100950971B1 - Cognitive neuro computational method of automatic lexical acquisition - Google Patents

Abstract

An object of the present invention is to provide a computational automatic vocabulary acquisition method that reflects the principles of cognitive neurological linguistic information processing.

According to the present invention, there is provided a method of acquiring cognitive nerve-based automatic vocabulary, comprising: a word acquiring function for storing the word in a word dictionary when the frequency of appearance or recentity of a word extracted from a corpus exceeds a corresponding first and second threshold values, respectively; And a morpheme acquisition function for storing in the morpheme dictionary when the frequencies of the head and tail strings of the words extracted from the corpus exceed the corresponding third and fourth threshold values, respectively.

Psychological Vocabulary, Psychological Vocabulary, Electronic Dictionary, Cognition, Neural, Natural Language Processing

Description

COGNITIVE NEURO COMPUTATIONAL METHOD OF AUTOMATIC LEXICAL ACQUISITION}

The present invention relates to a method for acquiring cognitive nerve based automatic vocabulary, and more particularly, to a method for calculating computational automatic vocabulary reflecting human cognitive neurological language information processing principles.

Cognitive neuroscience is a field of research that explains and clarifies the mechanisms and principles of human cognitive functions such as language, memory, learning, and attention using neurological and structural principles of the brain. In other words, cognitive neuroscience is the interdisciplinary research of cognitive science that seeks to reveal the basis and principle of human cognitive function and neuroscience that studies the structure, function and development of the nervous system. It can be said.

Recently, research and importance in the field of cognitive neuroscience are increasing day by day with increasing intellectual interest in clarifying the principle of human cognitive function and increasing necessity of brain science for developing intelligent robot. Therefore, in order to identify information processing of brain and understand and model brain and apply it to intelligent engineering system, it is necessary not only to combine cognitive science and neuroscience but also neuropsychology, computer science, electronic engineering. Multidisciplinary is needed by integrating statistics, statistics, and physics.

The necessity and importance of cognitive neural computation are as follows.

First, by analyzing and simulating cognitive neuron-based models, we can contribute to verifying and deepening existing theories obtained through behavioral experiments on humans.

Second, the manipulation of variables and parameter changes in computational models and systems enables predictive experiments that are difficult to perform in behavioral experiments using human subjects.

Third, through the observation and analysis of errors generated by damaged models by artificially damaging computational models, damage studies can be performed that cannot be performed by human subjects.

On the other hand, the necessity of developing an automatic vocabulary acquisition model reflecting the human language information processing principle is as follows.

First, understanding of vocabulary acquisition process is important for understanding vocabulary information processing process of human. Computational model can provide simulation environment for vocabulary acquisition process in cerebrum.

Second, vocabulary acquisition and vocabulary representation, which play an important role in understanding linguistic information processing, are closely related. Vocabulary acquisition research is also very important for mental lexicon representation.

Third, the computational automatic vocabulary acquisition model can verify the human vocabulary acquisition process, enable prediction experiments, and enable in-depth studies through damage studies.

Fourth, the study of automatic vocabulary acquisition can contribute to the development of the lexical information processing system of the automatic language processing system for the development of intelligent robots. In the past decades, we have been continuously researching in the field of artificial intelligence and natural language for the development of automated language information processing system. As a result, a system capable of semantic analysis with limited grammar in the proposed domain has been developed.

However, the current natural language processing technology is very poor compared to human language information processing technology, and the development of human-like intelligent language information processing system is a mysterious state. In order to improve the performance of the linguistic information processing system, both rule-based and statistical-based approaches have attempted to improve performance from a computational point of view, such as efficient language processing algorithms and the development of electronic dictionary structures. The point is presented.

Therefore, in order to develop the natural language processing system that can overcome the limitations of the existing language information processing system and simulate the human language processing ability, it is necessary to identify the human language information processing principle and to integrate it into the natural language processing system development. A natural language processing system needs to be developed.

An object of the present invention is to provide a computational automatic vocabulary acquisition method that reflects the principles of cognitive neurological linguistic information processing.

According to the present invention, there is provided a method of acquiring cognitive nerve-based automatic vocabulary, comprising: a word acquiring function for storing the word in a word dictionary when the frequency of appearance or recentity of a word extracted from a corpus exceeds a corresponding first and second threshold values, respectively; And a morpheme acquisition function for storing in the morpheme dictionary when the frequencies of the head and tail strings of the words extracted from the corpus exceed the corresponding third and fourth threshold values, respectively.

Preferably, the word acquisition function may include extracting a character string from the corpus in word units; Registering the appearance frequency information and the latestness information of the extracted character string in a database; Comparing the appearance frequency and the appearance frequency threshold of the extracted string, and comparing the recentness frequency and the recentness threshold of the extracted string; If the frequency of appearance of the extracted string is higher than the appearance frequency threshold, or if the frequency of retrieval of the extracted string is higher than the recentness threshold value, registering in the word dictionary.

Preferably, the morpheme acquiring function may include: dividing the word into a head string and a tail string at a syllable boundary of a word input from the corpus; Extracting occurrence frequency and simultaneous appearance string information of the head string and the tail string; Comparing a threshold value corresponding to a frequency of the header or footer string; Calculating an occurrence entropy up to the last syllable of the header string and an occurrence entropy up to the first syllable of the tail string when the frequency of the header string or the tail string exceeds a corresponding threshold value; And if it is determined that the last syllable entropy of the head string increases from the previous syllable entropy, and the first syllable entropy of the tail string increases immediately after the syllable entropy, the head string is designated as the head morpheme, and the tail string is designated as the tail morpheme. Storing in a morpheme dictionary.

According to the present invention, when there is only a corpus of a language to be learned, there is an effect of automatically obtaining a combination form and morpheme of the language.

On the other hand, according to the present invention, as a result of experiment using a Korean corpus of 10 million word size, 2,097 words with 38.63% of total frequency were learned, and 3,488 morphemes were extracted with an accuracy of 99.87%.

According to the present invention, it is possible to provide a computational automatic vocabulary acquisition method that reflects the principles of cognitive neurological linguistic information processing.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. Prior to this, terms or words used in the specification and claims should not be construed as having a conventional or dictionary meaning, and the inventors should properly explain the concept of terms in order to best explain their own invention. Based on the principle that can be defined, it should be interpreted as meaning and concept corresponding to the technical idea of the present invention. Therefore, the embodiments described in the specification and the drawings shown in the drawings are only one of the most preferred embodiments of the present invention and do not represent all of the technical idea of the present invention, various modifications that can be replaced at the time of the present application It should be understood that there may be equivalents and variations.

1 is a block diagram of an automatic vocabulary acquisition model according to the cognitive nerve-based automatic vocabulary acquisition method of the present invention.

The automatic lexical acquisition model according to the cognitive nerve-based automatic lexical acquisition method of the present invention is an automatic lexical acquisition model of a hybrid model that receives a corpus and constructs a lexical dictionary of a combined model and a morpheme dictionary of a decomposition model. The corpus that is input for the vocabulary learning of the model is to model the language life that human beings hear and see as they live the world.

The principle of word acquisition in word acquisition model is two kinds of acquisition considering frequency information and acquisition considering relativity information. The high frequency words in the language life show a frequency effect that is faster than the low frequency words. One of the reasons for the frequency effect is that the whole word is recognized as one unit, not the whole word is recognized by each word and the words that combine each one are recognized. The reason why the whole word should be stored in the dictionary is valid for the efficiency of word recognition. It is more efficient to understand the language by searching the whole word quickly in the dictionary for the higher frequency word. Storing word by storing high frequency word in word dictionary reflects this principle. The principle of relativity is that in the case of intensively repeated words at a specific period or time, the words are stored in the word dictionary in a combined form. It is easy to be.

Two parameters are used in the word acquisition model: the combined form frequency threshold ( ft ) and the recentness threshold ( rt ). Each threshold may be set differently according to the characteristics of the language to be obtained.

To this end, the automatic vocabulary acquisition model according to the cognitive nerve-based automatic vocabulary acquisition method of the present invention includes a corpus 100, a word acquisition model 200, a morpheme acquisition model 300, a word dictionary 400, and a morpheme dictionary 500. It includes.

When the corpus 100 is input, a word having a blank space is input to the word acquisition model 200 and the morpheme acquisition model 300.

The word acquisition model 200 includes a word frequency information extraction module 210, a word frequency information DB 220, a word threshold value 230, a word frequency threshold verification module 240, and a word recency information extraction module 250. The word recency information DB 260, the recency threshold value 270, and the word frequency recency verification module 280 store the word frequency or the word recency in a word dictionary.

The morpheme acquisition model 300 includes a morpheme candidate extraction module 310, a morpheme candidate appearance information DB 320, an entropy threshold 330, and a morpheme candidate verification module 340. Save it.

2 is a word acquisition flowchart according to a word acquisition model of the present invention.

In the word acquisition model 200 according to the present invention, the word frequency information extraction module 210 extracts a word input from the corpus 100 and stores the word word input in the word frequency information DB 220. The word frequency threshold verification module 240 ) Is regarded as a high frequency word for a word whose frequency of the word stored in the word frequency information DB 220 is higher than the appearance frequency threshold value ( ft, 230) and stored in the word dictionary 400.

In addition, the word acquisition model 200 verifies whether the extracted word repeatedly appears in words of a specific size and stores the word in the word dictionary. The word recentness information extraction module 250 is input from the corpus 100. the extracts Eojeol stores the Eojeol recency information DB (260), Eojeol frequency recent validation module 280 is a recency frequency of Eojeol stored in Eojeol recency information DB (260) recency threshold (rt, The word higher than the word 270 is regarded as a word having a high recurrence frequency and stored in the word dictionary 400.

That is, the character string is extracted from the corpus by word units (S210), and the word occurrence frequency information and the latestness information of the extracted string are registered in the DB (S220). If the appearance frequency of the extracted string is higher than the appearance frequency threshold value (ft, 230) or the recentness frequency of the extracted string is higher than the recentity threshold value (rt, 270) (S230), it is registered in the word dictionary (S240). However, if the frequency of appearance of the extracted string is not higher than the appearance frequency threshold (ft, 230) and the frequency of recency of the extracted string is not higher than the recentness threshold (rt, 270) (S230), the process returns to step S210. .

According to the present invention, a morpheme is divided into two strings and defines a morpheme or morpheme string composed of a string at the front as a head morpheme, and a morpheme or morpheme sequence composed of a sentence at the rear is defined as a tail morpheme. Detects strings frequently appearing in corpus words, verifies that they can be stemmed, and stores them in Mdic, a morpheme dictionary. The method of verifying that a specific string can be morpheme is verified by using a sucessor entropy and a preceding syllable entropy of the specific string.

An example of the stemming of a specific string using entropy will be described.

3 is a morpheme acquisition flowchart according to the morpheme acquisition model of the present invention.

In the morpheme acquisition model 300 of the present invention, the head string / tail string is divided at all possible syllable boundaries of the input word (S310), and the appearance frequency and the simultaneous appearance string information of each head string and the tail string are extracted (S320). ).

By comparing the frequency and the threshold of the extracted header or footer string (S330), if the frequency and the threshold of the extracted header or footer string are exceeded, the occurrence entropy (forward entropy) and tail string of the last syllable of the heading string are exceeded. The calculated entropy (reverse entropy) value until the first syllable is calculated (S340).

If it is determined that the last syllable entropy of the head string is increased from the previous syllable entropy, and the first syllable entropy of the tail string is increased from the syllable entropy immediately (S350), the head string is the head morpheme and the tail string is the tail morpheme dictionary ( 500) (S360). That is, when the calculated change in the forward and reverse entropy values show the morpheme likelihood, the corresponding head string and the tail string are extracted into the head morpheme and the tail morpheme, respectively, and stored in the morpheme dictionary 500.

Table 1 is an example showing the words starting with the string "Dad" and the frequency of occurrence of each word in the corpus.

Word frequency Even my father 8 Father only 4 To my father 4 father 16 Even my father 8 Sum 40

In the example of the word of Table 1, the frequency of the occurrences of the strings “ar”, “father”, and “father” are all 40 times higher than other substrings. Therefore, there is a possibility that all these strings are head stemming. However, morphemes combine words with other morphemes to create a word, so if a string is a head morpheme, various tail morphemes can be connected after it. For example, the string that can appear after the string "father" may be "ji", but there can be several tail stems after the string that can be the head stem of "father".

Therefore, the present invention calculates using the occurrence entropy of syllables appearing behind the head morphemes to investigate whether a particular string that appears at high frequency is a head morpheme. The entropy of syllables that can appear after a specific string is defined as in Equation 1.

Here, pi denotes the probability value of the syllable C _i after the string S.

The higher the Entropy (S) value, the more likely the string S is to be a morpheme. In the example of Table 1, the high-frequency strings "Father" to "A", "Father", "Father", and "Father" are graphs of entropy as shown in FIG.

4 is an entropy graph of a forward string according to an embodiment of the present invention, in which the entropy value decreases from Entropy (ar) to Entropy (arbor), and the entropy value increases in the string “father” which may become a head morpheme. can see. In addition, it can be seen that the entropy value decreases in the string in the process of forming the stem, and the entropy value increases at the position of the string in which the stem is made.

The present invention finds the character string of the position where the entropy value rises and determines the head morpheme candidate and investigates the possibility that the candidate is the head morpheme. The reason why entropy does not determine the rising string is head morphology because there can be various morphemes that share a specific string. For example, in Table 2, the substring “A” appears repeatedly, and the syllables after “A” are also variously shown as “Y”, “A”, “Ki”, and “Spit”.

Word frequency Ice cream 4 Sweetheart 8 Baby 8 In the morning 8 Ice cream 4

Therefore, the entropy value of "a" is calculated to a high value and an error may be registered as a head morpheme. If "Ah" is a head morpheme, "Iscream", "Gay", "Ki", "Spit", and "Ice Cream" should be tail morphemes. Ah ”cannot be a head morph.

According to the present invention, in order to investigate the possibility of the head morpheme, a method of investigating whether the remaining part except the head morpheme candidate in the word may be the tail morpheme is used. Possibility of tail morpheme uses entropy value in the same way as selecting head morpheme candidate, and calculates entropy value for preceding syllables using frequency information of string created from the end of word in reverse direction. For example, in order for the string "Father" to be a head stem in the string "Father," there must be a possibility that "Father" is a tail stem. Therefore, the entropy of the syllable that appears after the reverse order of the "book book" is high.

FIG. 5 is an entropy graph of a reverse string according to an embodiment of the present invention. Similar to FIG. 4, an entropy value is increased at a position of a tail morpheme.

The morpheme acquisition algorithm using the entropy of the forward string and the entropy of the reverse string as described above is expressed as pseudo code as follows.

// input: corpus of words (E)

// word E is a string of n + 1 syllables (C), E = C ₀ C ₁ C ₂ ..., C _n

// Si = C ₀ substring consisting of C ₁ , ..., C _i

// output: Morpheme-Dic (Mdic)

/ * Forward entropy increase check * /

1. Set tmp_dic to NULL // tmp_dic is a table that stores cumulative frequency and entropy values of temporary morphemes.

// SUCC _k (S _i ) is C _k after S _i

// tmp_dic is a table that stores {frequency of SU _i , SUCC _k (S _i ), SUCC _k (S _i )}

2. Divide the input word into syllable units (C ₀ , C ₁ , C ₂ ... C _n ); // ex) Arbor / Ger / G

3.for each i = 0 to (n-1)

4.{

5. Regenerate the string (S _i ) with the letters C ₀ + C ₁ + C ₂ + .... + C _i ; father, father, father

6. Recreate the string (S _j ) with the characters C _{(i + 1)} + C _{(i + 2)} + ... C _n ; // ex

If (is S _i registered in tmp_dic?) Then increases the frequency of S _{i in} tmp_dic by 1;

8. // ex) Father, Father

9. Register S _i with else tmp_dic;

10. If SUCC _k (S _i ) is registered in if (tmp_dic?) Then SUCC _k (S _i ) Increase frequency 1;

                                           father / father, father / father, father / father

11. Register SUCC _k (S _i ) in else tmp_dic;

12.

13.for each j = 0 to (n-1)

14. {

15. Forward_E (S _j ) =;

16. // (is), m is the number of SUCC _k (S _j ) in tmp_dic

17.}

18.}

19.for each i = 1 to (character-1)

20. {

21. if (entropy value of K _(i-1) <entropy value of K _i ) S _i Temporary storage (H _i );

22.}

/ * Check reverse entropy increase * /

23. Set Rtmp_dic to NULL // rtmp_dic is a table that stores cumulative frequency and entropy values of temporary morphemes.

// RSUCC _k (RS _i ) is RC _k after RS _i

// Rtmp_dic is a table that stores {frequency of RS _i , RSUCC _k (RS _i ), RSUCC _k (RS _i )}

24. The input word is divided into syllable units in the reverse order (RC ₀ , RC ₁ , RC ₂ ... RC _n ); // ex)

25.for each i = 0 to (n-1)

26. {

27. Regenerate the string S _i with the letters RC ₀ + RC ₁ + RC ₂ + .... + RC _i ; // ex)

28.RC _{(i + 1)} + RC _{(i + 2)} + ... regenerates the string (S _j ) with characters of RC _n ; // ex

29. if (? Is the RS _i is registered in Rtmp_dic) then Rtmp_dic S _i in the frequency increased by 1;

30. // ex) fall, branch, branch

31. Register S _i with else Rtmp_dic;

32. If RSUCC _k (S _i ) is registered in if (Rtmp_dic?) Then RSUCC _k (S _i ) increases by 1;

                                          // ex) ga / gibera, eggplant / bua, gaziber / ah

33. Register RSUCC _k (S _i ) with else Rtmp_dic;

34.

35.for each j = 0 to (n-1)

36. {

37. Forward_E (RS _j ) =;

38. // (), Rm is the number of RSUCC _k (S _j ) in Rtmp_dic

39.}

40.}

41.for each i = 1 to (character-1)

42. {

43. if (entropy value of RK _(i-1) <entropy value of RK _i ) RS _i temporary storage (RH _i );

44.}

As described above, although the present invention has been described by way of limited embodiments and drawings, the present invention is not limited thereto and is intended by those skilled in the art to which the present invention pertains. Various modifications and variations are possible within the scope of equivalents of the claims to be described.

1 is a block diagram of an automatic vocabulary acquisition model according to a cognitive nerve-based automatic vocabulary acquisition method of the present invention;

2 is a word acquisition flowchart according to a word acquisition model of the present invention;

3 is a morpheme acquisition flowchart according to the morpheme acquisition model of the present invention;

4 is an entropy graph of a forward string according to an embodiment of the present invention, and

5 is an entropy graph of a reverse string according to an embodiment of the present invention.

* Description of the main parts of the drawings *

100: Corpus 200: Word Acquisition Model

300: Morphological Acquisition Model 400: Word Dictionary

500: Morse Dictionary

Claims (4)

delete A word acquiring function for storing in a word dictionary if the frequency of appearance or recentity of the word extracted from the corpus exceeds the corresponding first and second threshold values, respectively; And Morphological acquisition function for storing in the morpheme dictionary if the frequency of the head and tail strings of the words extracted from the corpus exceeds the corresponding third and fourth threshold values, respectively Including, The word acquiring function may include extracting a character string from the corpus in word units; Registering the appearance frequency information and the latestness information of the extracted character string in a database; Comparing the appearance frequency and the appearance frequency threshold of the extracted string, and comparing the recentness frequency and the recentness threshold of the extracted string; If the frequency of appearance of the extracted string is higher than the appearance frequency threshold, or if the frequency of retrieval of the extracted string is higher than the recentness threshold, cognitive neuron-based automatic vocabulary, comprising the step of registering in the word dictionary Acquisition method. The method of claim 2, wherein the morpheme acquisition function, Dividing the word into a head string and a tail string at a syllable boundary of the word inputted from the corpus; Extracting occurrence frequency and simultaneous appearance string information of the header string and the tail string; Comparing a threshold value corresponding to a frequency of the header or footer string; Calculating an occurrence entropy up to the last syllable of the header string and an occurrence entropy up to the first syllable of the tail string when the frequency of the header string or the tail string exceeds a corresponding threshold value; And If it is determined that the last syllable entropy of the head string increases from the previous syllable entropy and the first syllable entropy of the tail string increases immediately after the syllable entropy, the head string is designated as the head morpheme and the tail string is designated as the tail morpheme. Step to save in advance Cognitive nerve-based automatic vocabulary acquisition method comprising a. The method of claim 3, wherein the entropy satisfies the following equation.

Where pi is the probability that the syllable C _i will appear after the string S.

Images