Improving Task specific Word embedding : Improving Hierarchical Word embedding using semantic of word
Improving Task specific Word embedding
- 주제(키워드) Word embedding , Poincare , hierarchical structure , Wordnet , NLP
- 발행기관 아주대학교 일반대학원
- 지도교수 손경아
- 발행년도 2019
- 학위수여년월 2019. 2
- 학위명 석사
- 학과 및 전공 일반대학원 컴퓨터공학과
- 실제URI http://www.dcollection.net/handler/ajou/000000028597
- 본문언어 영어
- 저작권 아주대학교 논문은 저작권에 의해 보호받습니다.
초록/요약
Using neural network models, it becomes possible to express words in a vector representation of a certain dimension. Existing word-embedding models have focused on creating semantic and syntactic features of words in vector-based textual data such as books, news, and so on. GLOVE [5], Word2Vec [6] and CBOW [8] as representative word embedding models, and recently, models such as FASTTEXT [7] have been introduced to further enhance existing models. Relational data such as graphs and networks also play an important role in the field of artificial intelligence. The network or graph embedding model such as latent space embeddings [9], DeepWalk [4] and Node2Vec [3] has also been studied and widely used in many applications. Although the text data embedding technique has become a common method of expressing the role of unit text in a context, it has solved the text mining problem using a machine learning model prior to neural network or a large lexical database such as Wordnet [14] [15]. Recently, there has been an effort to embed such a database in a graph embedding technique, since it is a type of data that does not match the expression intent of word embedding but cannot waste the value of an existing lexical database. There have been attempts to represent databases storing the semantics of words in a hierarchical manner through graph embedding techniques such as DeepWalk [4] and Node2Vec [3]. Recently, Poincare embedding has shown that it is possible to express the hierarchical relationship of words with sufficient performance even with low dimensional vector. In this paper, we embed hierarchical word data of WordNet [14] based on the Poincare embedding, and add an edge reflecting the sibling of the word to the hierarchical structure of WordNet data in order to represent more similar to the semantic meaning of itself. Converging the model with ideal objective function has been found to be difficult, but it can be improved through data-structured adjustments and more complex neural network models. Through this study, it is expected that we can solve the previously unresolved natural language processing problem through the expression of the vector of the hierarchical structure and the sibling relation of the word.
more목차
Chapter 1. Introduction
Section 1
1. Background
2. Research contents and method
3. Expected Effectiveness
4. Components of paper
Chapter 2. The Word Embeddings
Section 1. Context-based Word embedding
1. Count-based Word embedding
2. Word embedding using neural network
Section 2. Bilingual Word embedding
Section 3. Hierarchical word embedding
Chapter 3. Improving Hierarchical word embedding using semantic meaning of word
1. Research Objective
2. Method
3. Evaluation
Chapter 4. Conclusion
Reference
