DepND

Rule-based negation detection using dependency parse trees.

How to use:

• Make sure python 2.7 and GDep beta2 is installed on your machine. And the program share the the same directory as GDep parser.

• Run the command in terminal: python DepND.py YOUR_TESTING_DATA_FILEPATH OUTPUT_FILEPATH.

e.g. python DepND.py ./data/testing/bioscope_abstracts_cleaned.txt ./data/result.txt

Basic Workflow

• Separate text input into sentences. Filter all sentences not containing negation triggers. For sentences containing negation triggers, parse each of them using a dependency parser.

This task is done by DepNeg class, using, by default, GDep beta2 created by Prof. Kenji Sagae.
You certainly can use your favorite parser instead of GDep, just notice that you need modify Sent class if the parser you use doesn't generate output in CoNLL 2007 format.

• Use parse tree (CoNLL 2007 format) and rules to determine the scope of negation.

This task is done by DepND class

Rules for determining negation scope

I. Default rules:

Major rules:

• gMST

maximal spanning tree from its immediate governor;

• sMST

maximal spanning tree from itself;

• withinPUNC

All MST rules should not cross punctuation marks during spanning, no matter it spans towards left or right (relatively to the position of trigger word). But some arcs can (such as SUB, OBJ and PRD).

Specification

Trigger_POS	Trigger examples	Rule
RB, DT	no, not, hardly, never	gMST
JJ	absent, negative, unable, unlikely	gMST
CC	neither, nor	gMST
VB*, IN	fail, lack, lacking, excluding, without, except	sMST
NN	none, lack, absence, failure (to/of)	sMST

II. However, default rules above are not panaceas:

Exceptions that cannot be coped with default rules:

trg	trg_pos	trg_dep	gvn	gvn_pos	gvn_dep	rule	comment or alter-rule
not	RB	DEP	but, and	CC	NMOD	ggMST	DEP_Elevate; SUB&Right
not	RB	PMOD	in	IN	PMOD	gMST	PMOD arc can cross punctuation?
not	RB	VMOD	does, did, is	VBZ	ROOT	gMST	SUB&Right
hardly	RB	AMOD	any	DT	NMOD	ggMST	AMOD_Elevate
never	RB	NMOD	effect	NN	OBJ	gMST	SUB&Right
rather (than)	RB	PMOD	for	IN		gMST	SUB&Right
negative	JJ	NMOD	regulation, factors	NN, NNS	ROOT, PMOD	gMST	only span through of/IN NMOD arc; "factors" has no children
absence	NN	PMOD	in	IN	NMOD	gMST	PMOD_Elevate
none	NN	SUB	had	VBD	ROOT	gMST	SUB&Right
without	IN	PMOD	with	IN	PMOD	gMST	only span through PMOD towards right
lack	VBP	SBAR				sMST	forbid VMOD branching to MD or VB*
lacking	VBG	VC	is	VBZ	SUB	gMST	VC_Elevate
denied	VBN	VC	are	VBP	ROOT	gMST	VC_Elevate; SUB&Right
excluded	VBN	VC	be	VB	VC	ggMST	VC_Elevate

• P.S. Whether some exceptions (e.g. "negative", "without", "lack" in above table) arose depends on the performance of the dependency parser.
• ggMST (double "g" indicates governor of governor) can be replaced by $_Elevate.
• There're also minor rules to deal with subjunctive moods.

So, some minor rules are added:

• $_Elevate

elevate root node through all possible $ arcs then do a default MST (either sMST or gMST);

• SUB&Right

only span towards right or span left through SUB arc, span nothing if there's no SUB arc or right part. (notice that this rule only apply to root node)

III. Performance of designed rules are sensitive to:

• Performance of PP attachment in the parser;

e.g. The prepositional phrase "on sth" in "no effect on sth" is often wrongly attached to other NN rather than "effect".

• Inconsistency of decision in whether including the left part of parsed tree governed by a SUB arc.

e.g. In BioScope corpus, human annotator inconsistently include or exclude the subjects of sentences in the negation scope. A safer solution than the SUB&Right rule may be discarding SUB arcs and just spanning through towards right, sacrificing false negative for false positive.

Next Steps?

• Coping with double/triple negation.
• Coreference and pronoun resolution.
• Output formatting.

License

This projected is licensed under the terms of the Apache License, Version 2.0.