pepes97 / SRL

Semantic Role Labeling

Semantic Role Labeling is a fundamental NLPtask, which has the goal of finding semantic roles for each predicate in a sentence. The goal of the SRL is to extract predicate-argument structure of a sentence, identifying ”who did what to whom”, ”when”, ”where” etc.For example, consider this sentence:The cat eats a fish. Eats is the verb,The cat is the subject and a fish is the object complement. We are not interested in the meaning of ”cat” or ”fish”, but we want to identify and classify them, i.e, associate each argument with its corresponding role. To solve this task, LSTM-based models in different configurations were used in this paper, including pre-trained word embeddings, contextualized word embedding from BERT and Graph Convolutional Network. Furthermore, the subtask of disambiguation of predicates is also taken into consideration, because every often the datasets that are provided have information on the predicates present in the sentences, but not the clarification of the meaning of them.

Table of contents

• Development Setup
• Requirements
  • Download the glove data
  • Run

Development Setup

Requirements

conda create --name srl python=3.7
conda activate srl
pip install -r requirements.txt

Download Dataset

$ cd dataset
$ python download_dataset.py

Run

python main.py type-bert TYPE-BERT --batch-size BATCH-SIZE --embedding-dim-word EMBEDDING-DIM-WORD --embedding-dim-pretrained -- EMBEDDING-DIM-PRETRAINED --embedding-dim-pos EMBEDDING-DIM-POS --embedding-dim-pred EMBEDDING-DIM-PRED embedding-dim-dep-rel EMBEDDING-DIM-DEP-REL --embedding-dim-lemma EMBEDDING-DIM-LEMMA --hidden-dim HIDDEN-DIM --epochs EPOCHS --batch-size BATCH-SIZE --lr LEARNING-RATE --dropout DROPOUT --bidirectional BIDIRECTIONAL --num-layers NUM-LAYERS --only-test ONLY-TEST --pred-disamb PRED-DISAMB --pred-dim PRED-DIM

where

• TYPE-BERT type of BERT: base or large, default is base
• EMBEDDING-DIM-WORD is the dimension of the word embedding, default is 768
• EMBEDDING-DIM-PRETRAINED is the dimension of the word pretrained embedding, default is 50
• EMBEDDING-DIM-POS is the dimension of pos embedding, default is 32
• EMBEDDING-DIM-PRED is the dimension of predicate embedding, default is 50
• EMBEDDING-DIM-DEP-REL is the dimension of dependency relations embedding, default is 50
• EMBEDDING-DIM-LEMMA is the dimension of lemma embedding, default is 50
• HIDDEN-DIM is the dimension of the hidden layer, default is 256
• EPOCHS is the number of epochs, default is 15
• BATCH-SIZE is the batch size, default is 64
• LEARNING-RATE is the learning rate, default is 0.001
• DROPOUT is the dropout rate, default is 0.5
• BIDIRECTIONAL is the bidirectional flag, default is True
• NUM-LAYERS is the number of layers, default is 2
• ONLY-TEST is the only test flag, default is False
• PRED-DISAMB is the flag for addition of predicate dimbiguation, default is False
• PRED-DIM is the dimension of the predicates, default is 457

for example, for training:

python main.py

for testing

python main.py --only-test True