Repository files navigation

Implement some keyphrase generation algorithm

CopyRNN

Deep Keyphrase Generation (Meng et al., 2017)

CopyCNN

CopyTransformer

1. vocab_file: word line by line (don't with index!!!!)

   thispaperproposes

2. training, valid and test file

json line format, every line is a dict:

{'tokens': ['this', 'paper', 'proposes', 'using', 'virtual', 'reality', 'to', 'enhance', 'the', 'perception', 'of', 'actions', 'by', 'distant', 'users', 'on', 'a', 'shared', 'application', '.', 'here', ',', 'distance', 'may', 'refer', 'either', 'to', 'space', '(', 'e.g.', 'in', 'a', 'remote', 'synchronous', 'collaboration', ')', 'or', 'time', '(', 'e.g.', 'during', 'playback', 'of', 'recorded', 'actions', ')', '.', 'our', 'approach', 'consists', 'in', 'immersing', 'the', 'application', 'in', 'a', 'virtual', 'inhabited', '3d', 'space', 'and', 'mimicking', 'user', 'actions', 'by', 'animating', 'avatars', '.', 'we', 'illustrate', 'this', 'approach', 'with', 'two', 'applications', ',', 'the', 'one', 'for', 'remote', 'collaboration', 'on', 'a', 'shared', 'application', 'and', 'the', 'other', 'to', 'playback', 'recorded', 'sequences', 'of', 'user', 'actions', '.', 'we', 'suggest', 'this', 'could', 'be', 'a', 'low', 'cost', 'enhancement', 'for', 'telepresence', '.'] ,'keyphrases': [['telepresence'], ['animation'], ['avatars'], ['application', 'sharing'], ['collaborative', 'virtual', 'environments']]}

download thekp20k

mkdir datamkdir data/rawmkdir data/raw/kp20k_new# !! please unzip kp20k data put the files into above folder manuallypython -m nltk.downloader punktbash scripts/prepare_kp20k.shbash scripts/train_copyrnn_kp20k.sh# start tensorboard# enter the experiment result dir, suffix is time that experiment startscd data/kp20k/copyrnn_kp20k_basic-20191212-080000# start tensorboard servicestenosrboard --bind_all --logdir logs --port 6006

1. compared with the originalseq2seq-keyphrase-pytorch

   1. fix the implementation error:

      1. copy mechanism
      2. train and inference are not correspond (training doesn't have input feeding and inference has input feeding)

   2. easy data preparing
   3. tensorboard support
   4. faster beam search (6x faster used cpu and more than 10x faster used gpu)