kor2vec

Library for Korean morpheme and word vector representation.

• Paper : http://kiise.or.kr/e_journal/2018/5/JOK/pdf/04.pdf

Requirements

For training,

• Python 3
• Tensorflow
• numpy, scipy
• Konlpy (Twitter)

For test and visualization,

• gensim
• sklearn
• matplotlib

Model

Train Vectors

In order to learn morpheme vectors, do:

$ python3 train.py <input_corpus>

<input_corpus> format : one sentence = one line

Change Hyperparameters

$ python3 train.py -h
usage: train.py [-h] [--embedding_size EMBEDDING_SIZE]
                [--window_size WINDOW_SIZE] [--min_count MIN_COUNT]
                [--num_sampled NUM_SAMPLED] [--learning_rate LEARNING_RATE]
                [--sampling_rate SAMPLING_RATE] [--epochs EPOCHS]
                [--batch_size BATCH_SIZE]
                input

positional arguments:
  input                 input text file for training: one sentence per line

optional arguments:
  -h, --help            show this help message and exit
  --embedding_size EMBEDDING_SIZE
                        embedding vector size (default=150)
  --window_size WINDOW_SIZE
                        window size (default=5)
  --min_count MIN_COUNT
                        minimal number of word occurences (default=5)
  --num_sampled NUM_SAMPLED
                        number of negatives sampled (default=50)
  --learning_rate LEARNING_RATE
                        learning rate (default=1.0)
  --sampling_rate SAMPLING_RATE
                        rate for subsampling frequent words (default=0.0001)
  --epochs EPOCHS       number of epochs (default=3)
  --batch_size BATCH_SIZE
                        batch size (default=150)

Load Trained Morpheme Vectors

$ python3
>>>> from gensim.models.keyedvectors import KeyedVectors
>>>> pos_vectors = KeyedVectors.load_word2vec_format('pos.vec', binary=False)
>>>> pos_vectors.most_similar("('대통령','Noun')")

Generate Word Vectors

A word vector is defined by sum of its morphemes' vectors.

$ python3
>>>> from konlpy.tag import Twitter
>>>> import numpy as np
>>>> twitter = Twitter()
>>>> word = "대통령이"
>>>> pos_list = twitter.pos(word, norm=True)
>>>> word_vector = np.sum([pos_vectors.word_vec(str(pos).replace(" ", "")) for pos in pos_list], axis=0)

Test Dataset

• Word Similarity Test : Translated WordSim 353 Dataset into Korean. Translation ambiguous words were excluded.
  • WordSim 353 Dataset : http://www.cs.technion.ac.il/~gabr/resources/data/wordsim353/wordsim353.html
• Word Analogy Test : Created Semantic Pair 420 questions + Syntactic Pair 840 questions.

Test Morpheme Vectors

Similarity Test

Word similarity test using kor_ws353.csv.

$ python3 test/similarity_test.py pos.vec

Analogy Test (Semantic)

Word analogy test using kor_analogy_semantic.txt.

$ python3 test/analogy_test.py pos.vec

Visualization

Visualize the learned embeddings on two dimensional space using PCA.

$ python3 test/visualization.py pos.vec --words 밥 밥을 물 물을

Donwload Pre-trained Morpheme Vectors

Morpheme vectors are trained on Naver news corpus (218M tokens) using our model. You can download pre-trained morpheme vectors here : http://mmlab.snu.ac.kr/~djlee/pos.vec

Load Vectors using Gensim Library

$ python3
>>>> from gensim.models.keyedvectors import KeyedVectors
>>>> pos_vectors = KeyedVectors.load_word2vec_format('pos.vec', binary=False)
>>>> pos_vectors.most_similar("('대통령','Noun')")
>>>> pos_vectors.most_similar(positive=["('도쿄','Noun')", "('프랑스','Noun')"], negative=["('일본','Noun')"])