Convolutional Recurrent Neural Networks for Relation Extraction

Deep Learning Approach for Relation Extraction Challenge(SemEval-2010 Task #8: Multi-Way Classification of Semantic Relations Between Pairs of Nominals) using Convolutional Recurrent Neural Networks.

Experimental results

Parameters	Test Data Accuracy	F1 score
CRNN-Max	73%	74.28
CRNN-Att	65.95%	70.14

Usage

Train

• train data is located in "SemEval2010_task8_all_data/SemEval2010_task8_training/TRAIN_FILE.TXT"

• Display help message:

   $ python train.py --help

   optional arguments:
  -h, --help            show this help message and exit
  --train_dir TRAIN_DIR
                       Path of train data
  --dev_sample_percentage DEV_SAMPLE_PERCENTAGE
                       Percentage of the training data to use for validation
  --max_sentence_length MAX_SENTENCE_LENGTH
                       Max sentence length in train(98)/test(70) data
                       (Default: 100)
  --word2vec WORD2VEC   Word2vec file with pre-trained embeddings
  --text_embedding_dim TEXT_EMBEDDING_DIM
                       Dimensionality of character embedding (Default: 300)
  --layers LAYERS       Size of rnn output, no (Default: 100
  --dropout_keep_prob DROPOUT_KEEP_PROB
                       Dropout keep probability (Default: 0.5)
  --pooling_type POOLING_TYPE
                       pooling method, max or att (Default: max)
  --l2_reg_lambda L2_REG_LAMBDA
                       L2 regularization lambda (Default: 3.0)
  --f1 F1               f1 filter size (Default : 2)
  --f2 F2               f2 filter size (Default : 5)
  --n_channels N_CHANNELS
                       the number of channels-output vector size, nc(Default
                       : 100
  --batch_size BATCH_SIZE
                       Batch Size (Default: 64)
  --num_epochs NUM_EPOCHS
                       Number of training epochs (Default: 100)
  --display_every DISPLAY_EVERY
                       Number of iterations to display training info.
  --evaluate_every EVALUATE_EVERY
                       Evaluate model on dev set after this many steps
  --checkpoint_every CHECKPOINT_EVERY
                       Save model after this many steps
  --num_checkpoints NUM_CHECKPOINTS
                       Number of checkpoints to store
  --learning_rate LEARNING_RATE
                       Which learning rate to start with. (Default: 1e-3)

• Train Example:

   $ python train.py --train_dir "TRAIN_FILE.TXT" 

Evalutation

• test data is located in "SemEval2010_task8_all_data/SemEval2010_task8_testing_keys/TEST_FILE_FULL.TXT".

• You must give "checkpoint_dir" argument, path of checkpoint(trained neural model) file, like below example.

• Evaluation Example:

   $ python eval.py --checkpoint_dir "runs/1523902663/checkpoints"

• Official Evaluation of SemEval 2010 Task #8

  1. After evaluation like the example, you can get the "prediction.txt" and "answer.txt" in "result" directory.
  2. Install perl.
  3. Move to SemEval2010_task8_all_data/SemEval2010_task8_scorer-v1.2.

     $ cd SemEval2010_task8_all_data/SemEval2010_task8_scorer-v1.2

  4. Check your prediction file format.

     $ perl semeval2010_task8_format_checker.pl ../../result/prediction.txt

  5. Scoring your prediction.

     $ perl semeval2010_task8_scorer-v1.2.pl ../../result/prediction.txt ../../result/answer.txt

  6. The scorer shows the 3 evaluation reuslts for prediction. The official evaluation result, (9+1)-WAY EVALUATION TAKING DIRECTIONALITY INTO ACCOUNT -- OFFICIAL, is the last one. See the README for more details.

SemEval-2010 Task #8

• Given: a pair of nominals
• Goal: recognize the semantic relation between these nominals.
• Example:
  • "There were apples, pears and oranges in the bowl."
    → CONTENT-CONTAINER(pears, bowl)
  • “The cup contained tea from dried ginseng.”
    → ENTITY-ORIGIN(tea, ginseng)

The Inventory of Semantic Relations

1. Cause-Effect(CE): An event or object leads to an effect(those cancers were caused by radiation exposures)
2. Instrument-Agency(IA): An agent uses an instrument(phone operator)
3. Product-Producer(PP): A producer causes a product to exist (a factory manufactures suits)
4. Content-Container(CC): An object is physically stored in a delineated area of space (a bottle full of honey was weighed) Hendrickx, Kim, Kozareva, Nakov, O S´ eaghdha, Pad ´ o,´ Pennacchiotti, Romano, Szpakowicz Task Overview Data Creation Competition Results and Discussion The Inventory of Semantic Relations (III)
5. Entity-Origin(EO): An entity is coming or is derived from an origin, e.g., position or material (letters from foreign countries)
6. Entity-Destination(ED): An entity is moving towards a destination (the boy went to bed)
7. Component-Whole(CW): An object is a component of a larger whole (my apartment has a large kitchen)
8. Member-Collection(MC): A member forms a nonfunctional part of a collection (there are many trees in the forest)
9. Message-Topic(CT): An act of communication, written or spoken, is about a topic (the lecture was about semantics)
10. OTHER: If none of the above nine relations appears to be suitable.

Distribution for Dataset

• SemEval-2010 Task #8 Dataset [Download]

Relation	Train Data	Test Data	Total Data
Cause-Effect	1,003 (12.54%)	328 (12.07%)	1331 (12.42%)
Instrument-Agency	504 (6.30%)	156 (5.74%)	660 (6.16%)
Product-Producer	717 (8.96%)	231 (8.50%)	948 (8.85%)
Content-Container	540 (6.75%)	192 (7.07%)	732 (6.83%)
Entity-Origin	716 (8.95%)	258 (9.50%)	974 (9.09%)
Entity-Destination	845 (10.56%)	292 (10.75%)	1137 (10.61%)
Component-Whole	941 (11.76%)	312 (11.48%)	1253 (11.69%)
Member-Collection	690 (8.63%)	233 (8.58%)	923 (8.61%)
Message-Topic	634 (7.92%)	261 (9.61%)	895 (8.35%)
Other	1,410 (17.63%)	454 (16.71%)	1864 (17.39%)
Total	8,000 (100.00%)	2,717 (100.00%)	10,717 (100.00%)

Reference

• Learning local and global contexts using a convolutional recurrent network model for relation classification in biomedical text (CoNLL 2017), D Raj et al. [paper] [github]