$ python3 main.py --gpu_idx 0 1 \  # number of GPUs used for training and their indices
    --mode train \  # training model or infer
    --resume_training false \  # if true, will resume previous trained parameters
    --neg_sample 10 \  # number of negative samples
    --word_dim 300 \  # input pre-trained / randomly initialized word embedding dimension
    --num_units 300 \  # number of units for rnn cell and hidden layer of feed-forward network
    --k 100 \  # number of units for output layer
    --use_ntn false \  # if use neural tensor network to fuse left and right contexts, otherwise just simply concatenate them
    --tune_emb false \  # whether the input word embedding are tunable while training
    --lr 0.0001 \  # learning rate
    --decay_step 10000 \  # learning rate decay step
    --decay_rate 0.9994 \  # decay rate
    --batch_size 1000 \  # batch size
    --epochs 30 \  # total training epochs
    --ckpt ckpt/ \  # checkpoint path to save model
    --max_to_keep 3 \  # maximal checkpoints can be saved
    --model_name conceptual_primitives \  # model name
    --save_step 10000 \  # save models per steps
    --print_step 1000 \  # show sample test result per steps
    --ukwac_path <raw ukwac dataset path> \  # raw ukwac dataset path
    --glove_path <pre-trained glove embedding path> \  # pre-trained glove word embedding path
    --save_path <processed data save path> \  # path for saving processed dataset
    --word_threshold 90 \  # word threshold, minimal occurrence of words to be kept
    --word_lowercase true  # whether lowercase the text

$ python3 main.py --gpu_idx 0 1 --mode infer --use_ntn false
restored model from conceptual_primitives-1000000, done...
Top 10 canidates:
['nibbling', 'drinking', 'munching', 'snacking', 'feeding', 'gorging', 'tasting', 'swallowing', 'chewing', 'feasting']