An LSTM Method for Natural Pronunciation Expression of Foreign Words in Sentences

문장에 포함된 외국어의 자연스러운 발음 표현을 위한 LSTM 방법

Abstract

Korea language has postpositions such as eul, reul, yi, ga, wa, and gwa, which are attached to nouns and add meaning to the sentence. When foreign notations or abbreviations are included in sentences, the appropriate postposition for the pronunciation of the foreign words may not be used. Sometimes, for natural expression of the sentence, two postpositions are used with one in parentheses as in "eul(reul)" so that both postpositions can be acceptable. This study finds examples of using unnatural postpositions when foreign words are included in Korean sentences and proposes a method for using natural postpositions by learning the final consonant pronunciation of nouns. The proposed method uses a recurrent neural network model to naturally express postpositions connected to foreign words. Furthermore, the proposed method is proven by learning and testing with the proposed method. It will be useful for composing perfect sentences for machine translation by using natural postpositions for English abbreviations or new foreign words included in Korean sentences in the future.

한국어는 "을/를/이/가/와/과"와 같은 조사가 체언에 붙어 문장의 의미를 더해준다. 문장 중에 외국어 표기를 그대로 사용하는 경우나 외국어의 약자가 포함되어 있는 경우, 외국어의 발음에 따른 적절한 조사가 연결되지 않는 경우가 있다. 때로는 문장의 자연스러운 표현을 위하여 "을(를)"과 같이 괄호 형식으로 표현하여 조사를 두 개 다 수용 가능한 형태로 사용되어지기도 한다. 본 연구에서는 문장 내에 외국어가 포함되어 있는 경우, 조사가 부자연스럽게 연결되는 예를 찾고 체언의 종성 발음을 학습하여 자연스러운 조사 연결을 위한 방법을 알아보고자 한다. 제안하는 방법은 순환신경망 모델을 이용하여 외국어에 연결된 조사를 자연스럽게 표현하는 것이다. 제안된 모델로 학습 및 테스트하여 방법의 필요성을 입증함으로써, 향후 기계 번역에서 영문 약자나 새로운 외국어 삽입 시 자연스러운 조사 연결로 완전한 문장을 연결하는데 사용될 수 있을 것으로 기대한다.

Fig. 1. Data Composition by One-hot Encoding with Last Five Characters of the Word

Fig. 2. The Accuracy with Training and Testing Data Set. (A) Drop-out technique is not applied on any three stacked layers, (B) Drop-out technique is applied on the only bottom layer among three stack layer, (C) Drop-out technique is applied all three stack layers

Fig. 3. The Accuracy with Training and Testing Data Set. (A) Drop-out technique is applied all three stack layers, (B) Drop-out technique is applied all four-stacked-layers, (C) Drop-out technique is applied all five stack layers

Fig. 4. The Framework of Suggested Model

Fig. 5. The Screen-shot of Tensor Board forthe Suggested Model

Fig. 6. The Min, Max, and Average Accuracy for Each Epoch. The Red Point Stands for Average of Accuracy and Blue Range is Min and Max of Accuracy

Fig. 7. The ROC Curve

Fig. 8. Average Confusion Matrix Values of 10-fold Cross Validation for Each Epoch

Table 1. The Examples of Automatic Translation by Google, Naver and Kakao Applications

Table 2. The Number of Vowel and Consonant Word in Dataset

Table 3. The Number of Parts of Speech in Dataset

Table 4. The Examples of Wrong Transliteration in Korean

Table 5. Data Classification Depending on the Korean Pronunciation. “1” at postposition class stands for “eul - 을” and “0” means “reul-를”

Table 6. The Number and Distribution of Data for Each Class

Table 7. The Number of Word for Each Length in Dataset

Table 8. Confusion Matrix