TODO 6/6/15
- Train baseline and recurrent on 100k set to compare with recursive
- Train a separate model to predict relation/no relation and integrate with baseline and recurrent for kbp pipeline
- Generate plots from the data in rnn_xxx_dev_perf.log
5/26/15:
- Melvin: concat 2 entity vectors, predict softmax of relation at the top
- Melvin: avg of trained word vectors, predict relation at the top
- Ankur: recurrent NN, using dep path
- Mikhail: recursive NN (hw3 implementation); use NLTK to get constituency parse for sentences (but only keep the subtree that contains the mentions)
5/17/15:
- Compute word2vec for KBP corpus
- Keras - implement basic MLP with triples
- Keras - implement average of word vectors
- Keras - implement average of words in dependency path
- Figure out about the training and test data mismatch
5/10/15:
- mikhail: batch_size = 20, lr = 0.001, reg = 0.001
- ankur: batch_size = 20, lr = 0.01, reg = 0.001
- melvin: batch_size = 20, lr = 0.01, reg = 0.01
- train average model on 1 million sentences and test using validation set
- utilities to save and load a model
5/3/15:
- DL word vectors (glove)
- get code to extract dep path between mentions
- generate micro datasets in addition to the 10k (maybe 100k, 1mil)
- write a recursive NN
Desc of Training data CSV:
gloss | text | | extended | dependencies_conll | text | | extended | words | text[] | | extended | lemmas | text[] | | extended | pos_tags | text[] | | extended | ner_tags | text[] | | extended | subject_id | bigint | | plain | subject_entity | text | | extended | subject_link_score | real | | plain | subject_ner | text | | extended | object_id | bigint | | plain | object_entity | text | | extended | object_link_score | real | | plain | object_ner | text | | extended | subject_begin | smallint | | plain | subject_end | smallint | | plain | object_begin | smallint | | plain | object_end | smallint | | plain | known_relations | text[] | | extended | incompatible_relations | text[] | | extended | annotated_relation | text | | extended |
Desc of Test data CSV:
Column | Type | Modifiers | Storage | Description
--------------------+----------+-----------+----------+------------- gloss | text | | extended | dependencies_conll | text | | extended | words | text[] | | extended | lemmas | text[] | | extended | pos_tags | text[] | | extended | ner_tags | text[] | | extended | subject_id | bigint | | plain | subject_entity | text | | extended | subject_link_score | real | | plain | subject_ner | text | | extended | object_id | bigint | | plain | object_entity | text | | extended | object_link_score | real | | plain | object_ner | text | | extended | subject_begin | smallint | | plain | subject_end | smallint | | plain | object_begin | smallint | | plain | object_end | smallint | | plain |