Skip to content

Latest commit

 

History

History
55 lines (40 loc) · 3.01 KB

SUBMISSION_README.MD

File metadata and controls

55 lines (40 loc) · 3.01 KB
Raw

BASIC MODEL:

Steps to run the basic model with word, pos and shortest path: [DEFAULT]

  1. Use shortest_path = find_shortest_path(doc, start, end) in data.py [Check around line 71, Comment out line 74]
  2. RUN: python train_basic.py --embed-file data/glove.6B.100d.txt --embed-dim 100 --batch-size 10 --epochs 5

For the experiments: a. Only Word Embeddings:

  1. Use shortest_path = [] in data.py [Check around line 74, Comment out line 71]
  2. Use rnn_outputs = self.gru_layer(word_embed, mask=mask) in model.py [Check around line 51, Comment out line 48]
  3. RUN: python train_basic.py --embed-file data/glove.6B.100d.txt --embed-dim 100 --batch-size 10 --epochs 5

b. Only Word + PoS Embeddings:

  1. Use shortest_path = [] in data.py [Check around line 74, Comment out line 71]
  2. Use rnn_outputs = self.gru_layer(rnn_input, mask=mask) in model.py [Check around line 48, Comment out line 51]
  3. RUN: python train_basic.py --embed-file data/glove.6B.100d.txt --embed-dim 100 --batch-size 10 --epochs 5

c. Only Word + Shortest Path Dependency Structure:

  1. Use shortest_path = find_shortest_path(doc, start, end) in data.py [Check around line 71, Comment out line 74]
  2. Use rnn_outputs = self.gru_layer(word_embed, mask=mask) in model.py [Check around line 51, Comment out line 48]
  3. RUN: python train_basic.py --embed-file data/glove.6B.100d.txt --embed-dim 100 --batch-size 10 --epochs 5

ADVANCED MODEL:

For the advanced model, changes were made to data.py in order to gather positional features which are inputs to the advanced model.

Changes were also made to the call method of the basic model as the implementation for the basic and advanced model were tied at the train_lib.py file which requires inputs to the model be a certain way.

Changes added had an impact on generating training and val instances due to the additional steps involved. It can be further improved

Steps to run the advanced model:

  1. Use shortest_path = [] in data.py [Check around line 74, Comment out line 71]
  2. RUN: python train_advanced.py --embed-file data/glove.6B.300d.txt --embed-dim 300 --batch-size 20 --epochs 6

Forward Prop of our Advanced Model:

  1. The input to our advanced model is a concatenations of word embeddings with that positional embeddings. Positional embeddings are essentially based on how close or far other words are to our entities in question.

  2. The advanced model has one layer of convolutional filters of different sizes being applied to the input. We convolve with filters of different sizes [2, 3, 4 and 5] with use ReLU as our activation function. To each of the results of the filters we apply Max Pooling using a filter of size according the size of the filter we convolved with.

  3. Next we concat the outputs of each of our CNN filters that is (Conv, ReLU, MaxPool), flatten it out, add then add dropout to address for over-fitting. This generates a sentence representation for us to use for classification into our 19 classes for relationship.

Advanced Model Architecture