citation-analysis/README.md

Citation Intent Classification

Project repo for Computational Linguistics Team Lab at the University of Stuttgart.

Introduction

This repository contains code and datasets for classifying citation intents in research papers.

We implemented 3 classifiers and evaluated on test dataset:

• Perceptron Classifier - Baseline model (Implemented from scratch)
• Feedforward Neural Network Classifier (using PyTorch)
• BiLSTM + Attention with ELMo Embeddings (using AllenNLP library)

This README documentation focuses on running the code base, training the models and predictions. For more information about our project work, model results and detailed error analysis, check this report.
For more information on the Citation Intent Classification in Scientific Publications, follow this link to the original published paper and their GitHub repo

Environment & Setup

TODO

Data

We have 3 different intents/classes in the dataset:

• background (background information)
• method (use of methods)
• result (comparing results)

Dataset Class distribution:

	background	method	result
train	4.8 K	2.3 K	1.1 K
dev	0.5 K	0.3 K	0.1 K
test	1 K	0.6 K	0.2 K

Methods (Classification)

1) Perceptron Classifier (Baseline Classifier)

We implemented Perceptron as a baseline classifier, from scratch (including evaluation). Perceptron is an algorithm for supervised learning of classification. It's a linear and binary classifier, which means it can only decide whether or not an input feature belongs to some specific class and it's only capable of learning linearly separable patterns.

class Perceptron:
  def __init__(self, label: str, weights: dict, theta_bias: float):
  def score(self, features: list):
  def update_weights(self, features: list, learning_rate: float, penalize: bool, reward: bool):

class MultiClassPerceptron:
  def __init__(self, epochs: int = 5000, learning_rate: float = 1, random_state: int = 42)
  def fit(self, X_train: list, labels: list)
  def predict(self, X_test: list)

Since we have 3 different classes for Classification, we create a Perceptron object for each class. Each Perceptron has score and update functions. During training, for a set of input features it takes the score from the Perceptron for each label and assigns the label with max score(for all the data instances). It compares the assigned label with the true label and decides whether or not to update the weights (with some learning rate).

Check the source code for more details on the implementation of Perceptron Classifier.

Running the Model

(citation-env) [user@server citation-analysis]$ python -m testing.model_testing

Link to the source code. All the Hyperparameters can be modified to experiment with.

Evaluation

we used f1_score metric for evaluation of our baseline classifier.

F1 score is a weighted average of Precision and Recall(or Harmonic Mean between Precision and Recall). The formula for F1 Score is:
F1 = 2 * (precision * recall) / (precision + recall)

eval.metrics.f1_score(y_true, y_pred, labels, average)  

Parameters: y_true : 1-d array or list of gold class values
y_pred : 1-d array or list of estimated values returned by a classifier
labels : list of labels/classes
average: string - [None, 'micro', 'macro'] If None, the scores for each class are returned.

Link to the metrics source code.

Results

2) Feedforward Neural Network (using PyTorch)

A feed-forward neural network classifier with a single hidden layer containing 9 units. While a feed-forward neural network is clearly not the ideal architecture for sequential text data, it was of interest to add a sort of second baseline and examine the added gains (if any) relative to a single perceptron. The input to the feedforward network remained the same; only the final model was suitable for more complex inputs such as word embeddings.

Check this feed-forward model source code for more details.

3) BiLSTM + Attention with ELMo (AllenNLP Model)

The Bi-directional Long Short Term Memory (BiLSTM) model built using the AllenNLP library. For word representations, we used 100-dimensional GloVe vectors trained on a corpus of 6B tokens from Wikipedia. For contextual representations, we used ELMo Embeddings which have been trained on a dataset of 5.5B tokens. This model uses the entire input text, as opposed to selected features in the text, as in the first two models. It has a single-layer BiLSTM with a hidden dimension size of 50 for each direction.

We used AllenNLP's Config Files to build our model, just need to implement a model and a dataset reader (with a Config file).

Our BiLSTM AllenNLP model contains 4 major components:

1. Dataset Reader - CitationDatasetReader
   • It reads the data from the file, tokenizes the input text and creates AllenNLP Instances
   • Each Instance contains a dictionary of tokens and label
2. Model - BiLstmClassifier
   • The model's forward() method is called for every data instance by passing tokens and label
   • The signature of forward() needs to match with field names of the Instance created by the DatasetReader
   • The forward() method finally returns an output dictionary with the predicted label, loss, softmax probabilities and so on...
3. Config File - basic_model.json
   • The AllenNLP Configuration file takes the constructor parameters for various objects (Model, DatasetReader, Predictor, ...)
   • We can also define a number of Hyperparameters from the Config file.
     • Depth and Width of the Network
     • Number of Epochs
     • Optimizer & Learning Rate
     • Batch Size
     • Dropout
     • Embeddings
4. Predictor - IntentClassificationPredictor
   • AllenNLP uses Predictor, a wrapper around trained model, for making predictions.
   • The Predictor uses a pre-trained/saved model and dataset reader to predict new Instances

Running the Model

TODO

Evaluation

TODO

Results

References