bluesky1018/relation-classification

Relation Classification - SEMEVAL 2010 task 8 dataset

Relation-Classification

Relation Classification - SEMEVAL 2010 task 8 dataset (Master's Thesis)

• Relation classification is a task of assigning predefined relation labels to the entity pairs that occur in texts.
  • Example:
  • Sentence: [People]_e1 have been moving back into [downtown]_e2
  • Relation: Entity-Destination(e1,e2) where e1 = people, e2 = downtown

Cite Us As

@MastersThesis{Sahitya:2018,
 author = { {Sahitya Patel} and Harish Karnick}, 
 title = {Multi-Way Classification of Relations Between Pairs of Entities}, 
 school = {Indian Institute of Technology Kanpur (IITK)}, 
 address = {India},
 year = 2018, 
 month = 6
}

Presentation

Relation-Classification-github.pdf

Dataset

Paper: SemEval-2010 Task 8: Multi-Way Classification of Semantic Relations Between Pairs of Nominals

Zip: SemEval2010_task8_all_data.zip

Files

Preprocessing of data

• 01_create_train_test_attn
• 02_train_val_split
• 03_data_preprocess

Model training

• 04_CBGRU_MEA_Model

Environment

• Python 3.5.4 | Anaconda custom (64-bit)
• Keras 2.1.5
• Tensorflow 1.4.0
• CUDA compilation tools, release 8.0, V8.0.44 (nvcc --version)
• CuDNN 6.0.21
• Perl

Running the model without preprocessing

1. Get preprocessed data. Download "data_all.npy" from this-link (94.6 MB) and put it in the folder "./data/".
2. Run "04_CBGRU_MEA_Model"

Running the model with preprocessing

01_create_train_test_attn

Description: Pre-processing of dataset files

Reads:

• "./corpus/SemEval2010_task8_training/TRAIN_FILE.TXT"
• "./corpus/SemEval2010_task8_testing_keys/TEST_FILE_FULL.TXT"

Creates:

• "./files/train_attn.txt"
• "./files/test_attn.txt"

To Do:

1. Set the following path in "01_create_train_test_attn"

os.environ['CLASSPATH'] = "H:/Relation-Classification/stanford/stanford-postagger-2017-06-09"

2. Run "01_create_train_test_attn"

02_train_val_split

Description: Spliting of the training data into training and validation data

Reads:

• "./files/train_attn.txt"
• "./files/test_attn.txt"

Creates:

• "./files/train_attn_sp.txt"
• "./files/val_attn_sp.txt"
• "./files/test_attn_sp.txt"

To Do:

1. Run "02_train_val_split"

03_data_preprocess

Description: Generating a single input file for the model

Creates:

• "./data/data_all.npy"

Steps:

1. Place "GoogleNews-vectors-negative300.bin" in "./word_embeddings" folder. (Download-Link, Website-word2vec)
2. Run "./word_embeddings/GoogleNews-vectors-negative300_bin_to_txt.py" to create "./word_embeddings/GoogleNews-vectors-negative300.txt"
3. Run "03_data_preprocess"

04_CBGRU_MEA_Model

Description: Model training. Best model is saved in "./model" folder.

Steps:

1. Run "04_CBGRU_MEA_Model"

Creates:

• "./model/model.keras" - Model

Model CBGRU-MEA