Dependency Parser Using Neural Networks for NLTK

The Natural Language Toolkit (NLTK) is a leading suite of libraries for natural language processing, and with the rise of artificial intelligence, its popularity has only increased in the recent years. Since existing parsers in NLTK are either outdated or not implemented in Python, its primary programming language, this project proposes an implementation of a fast and accurate dependency parser to include in the toolkit. In this work, the transition-based parsing algorithm created by Danqi Chen and Christopher D. Manning is implemented and incorporated into NLTK. The algorithm uses artificial neural networks for predicting transitions and introduces several groundbreaking concepts. The parser employs an efficient new way of feature representation, particularly for part-of-speech tags and arc labels. Instead of one-hot encoding, dense representations are used for words, tags, and labels. Another innovation is the use of cube activation function in the neural network classifier instead of the commonly used tanh and sigmoid functions. Index Terms — natural language processing, dependency parsing, artificial neural networks, NLTK.