Open JudePark96 opened 3 years ago
κΈ°μ‘΄μ ν μ€νΈ λΆλ₯ κ³Όμ μ νκΈ° μν λͺ λͺ μ°κ΅¬μμλ convolutional neural networks λ₯Ό μ¬μ©νμλ€. λ³Έ λ Όλ¬Έμμλ convolutional neural networks κ° μλ graph convolutional networks λ₯Ό ν μ€νΈ λΆλ₯ κ³Όμ μ μ μ©νλ κ²μ μ μνλ€.
We propose a novel graph neural network method for text classification. To the best of our knowledge, this is the first study to model a whole corpus as a heterogeneous graph and learn word and document embeddings with graph neural networks jointly.
Results on several benchmark datasets demonstrate that our method outperforms state-of-the-art text classification methods, without using pre-trained word embeddings or external knowledge. Our method also learn predictive word and document embeddings automatically.
κΈ°μ‘΄μ ν μ€νΈ λΆλ₯ κ³Όμ μ μν μ£Όμ λ₯λ¬λ λͺ¨λΈλ‘λ CNN, LSTM λ±μ΄ μλ€. μ΄λ¬ν λͺ¨λΈλ€μ local consecutive word sequences μ semantic, synthetic information μ μ caputre ν μ μμ§λ§ global word co-occurence λ₯Ό 무μν μ μλ κ°λ₯μ±μ΄ μλ€.
λ³Έ λ Όλ¬Έμμλ μ΄μ μ μ κ·Όκ³Όλ λ€λ₯΄κ² graph representation μ νμ΅νμ¬ classification task λ₯Ό μννλ€λ μ μ΄λ€. μ΄λ₯Ό μννκΈ° μνμ¬ corpus -> graph μ λ°©λ²μΌλ‘ PMI, TF-IDF λ₯Ό μκ°νμκ³ , PMI μ κ²½μ°κ° μ±λ₯μ΄ λ μ’μμ μλ €μ€λ€. λν two-layer GCN μ΄ one-layer GCN λ³΄λ€ μ±λ₯μ΄ μ’μμ§λ§ λ λ§μ layer λ μ±λ₯μ ν° μλ―Έκ° μμμ λνλ΄μλ€.
Contents
1. μ΄λ‘μ λλΌκ³ λ§νκ³ μμ΄ ?
2. μ£Όμ κΈ°μ¬μ μ λμΌ ?
3. μ΄μ μ μ κ·Όκ³Όλ λκ° λ€λ₯Έ κ² κ°μ ?
4. μ΄λ€ κ±Έ μ μν μ μμκΉ ?
5. λ€μ λ Όλ¬Έμ 무μμ μ½μ΄μΌν κΉ ?