Homework 1 old

Dependency Parsing

Your task is to implement a dependency parser reaching to the state-of-the-art accuracy and speed. You are allowed to work in groups of at most 2. Submit your work by Oct. 7th before the class.

• Download the standard dataset.
• Implement a transition-based dependency parser using Nivre's arc-eager algorithm using our architecture.
• Improve your parser in various ways.
• Evaluate the accuracy of your parser (UAS and LAS) for both Malt and Stanford dependency formats.
• Evaluate the speed of your parser (tokens/sec. and sentences/sec.).
• Write a report (4-8) pages in the ACL format. Your report must include abstract, introduction, related work, approach, experiments, and conclusion.
• Commit all your work to your Github repoistory.
• Create a wiki page Dependency Parsing showing instructions of how to run your parser. You must provide a pre-trained model that is ready to be run.

Data format

1	Ms.	ms.	NNP	_	2	NMOD	2	nn
2	Haag	haag	NNP	_	3	SUB	3	nsubj
3	plays	play	VBZ	_	0	ROOT	0	root
4	Elianti	elianti	NNP	_	3	OBJ	3	dobj
5	.	.	.	_	3	P	3	punct

Each column represents:

• 0: ID.
• 1: word-form.
• 2: lemma (predicted).
• 3: POS tag (predicted).
• 4: extra features (blank).
• 5: head ID from Malt (gold).
• 6: dependency label from Malt (gold).
• 7: head ID from Stanford (gold).
• 8: dependency label from Stanford (gold).

Notes

• Your dependency parser must extend NLPComponent.
• No 3rd-party library including any implementation of a dependency parser can be used for this homework.
• Your final model must not be tuned for the evaluation set.
• Your report must clearly state all of your approaches and findings including:
• Machine learning algorithm.
• Parsing algorithm.
• Search strategy.
• Feature set.
• In your wiki-page, indicate which codes are used for your dependency parser.
• You must report the following for at least your baseline and final models.
• UAS and LAS for the Malt dependency format on the development and evaluation sets.
• UAS and LAS for the Stanford dependency format on the development and evaluation sets.
• Use SpeedTest for measuring the speed of your parser.