id	sid	tid	token	lemma	pos
kw52j67652r	1	1	natural	natural	ADJ
kw52j67652r	1	2	language	language	NOUN
kw52j67652r	1	3	processing	processing	NOUN
kw52j67652r	1	4	(	(	PUNCT
kw52j67652r	1	5	nlp	nlp	PROPN
kw52j67652r	1	6	)	)	PUNCT
kw52j67652r	1	7	models	model	NOUN
kw52j67652r	1	8	the	the	DET
kw52j67652r	1	9	different	different	ADJ
kw52j67652r	1	10	techniques	technique	NOUN
kw52j67652r	1	11	computers	computer	NOUN
kw52j67652r	1	12	use	use	VERB
kw52j67652r	1	13	to	to	PART
kw52j67652r	1	14	understand	understand	VERB
kw52j67652r	1	15	and	and	CCONJ
kw52j67652r	1	16	interpret	interpret	VERB
kw52j67652r	1	17	human	human	ADJ
kw52j67652r	1	18	languages	language	NOUN
kw52j67652r	1	19	.	.	PUNCT
kw52j67652r	2	1	nlp	nlp	PROPN
kw52j67652r	2	2	covers	cover	VERB
kw52j67652r	2	3	a	a	DET
kw52j67652r	2	4	wide	wide	ADJ
kw52j67652r	2	5	range	range	NOUN
kw52j67652r	2	6	of	of	ADP
kw52j67652r	2	7	sub	sub	NOUN
kw52j67652r	2	8	-	-	NOUN
kw52j67652r	2	9	topics	topic	NOUN
kw52j67652r	2	10	such	such	ADJ
kw52j67652r	2	11	as	as	ADP
kw52j67652r	2	12	syntax	syntax	NOUN
kw52j67652r	2	13	(	(	PUNCT
kw52j67652r	2	14	analyzing	analyze	VERB
kw52j67652r	2	15	if	if	SCONJ
kw52j67652r	2	16	words	word	NOUN
kw52j67652r	2	17	in	in	ADP
kw52j67652r	2	18	an	an	DET
kw52j67652r	2	19	utterance	utterance	NOUN
kw52j67652r	2	20	are	be	AUX
kw52j67652r	2	21	well	well	ADV
kw52j67652r	2	22	arranged	arranged	ADJ
kw52j67652r	2	23	)	)	PUNCT
kw52j67652r	2	24	,	,	PUNCT
kw52j67652r	2	25	semantics	semantic	NOUN
kw52j67652r	2	26	(	(	PUNCT
kw52j67652r	2	27	understanding	understand	VERB
kw52j67652r	2	28	the	the	DET
kw52j67652r	2	29	meaning	meaning	NOUN
kw52j67652r	2	30	of	of	ADP
kw52j67652r	2	31	combined	combine	VERB
kw52j67652r	2	32	words	word	NOUN
kw52j67652r	2	33	)	)	PUNCT
kw52j67652r	2	34	,	,	PUNCT
kw52j67652r	2	35	and	and	CCONJ
kw52j67652r	2	36	discourse	discourse	NOUN
kw52j67652r	2	37	.	.	PUNCT
kw52j67652r	3	1	most	most	ADJ
kw52j67652r	3	2	state	state	NOUN
kw52j67652r	3	3	-	-	PUNCT
kw52j67652r	3	4	of	of	ADP
kw52j67652r	3	5	-	-	PUNCT
kw52j67652r	3	6	the	the	DET
kw52j67652r	3	7	-	-	PUNCT
kw52j67652r	3	8	art	art	NOUN
kw52j67652r	3	9	nlp	nlp	NOUN
kw52j67652r	3	10	systems	system	NOUN
kw52j67652r	3	11	feed	feed	VERB
kw52j67652r	3	12	large	large	ADJ
kw52j67652r	3	13	amounts	amount	NOUN
kw52j67652r	3	14	of	of	ADP
kw52j67652r	3	15	natural	natural	ADJ
kw52j67652r	3	16	language	language	NOUN
kw52j67652r	3	17	text	text	NOUN
kw52j67652r	3	18	into	into	ADP
kw52j67652r	3	19	different	different	ADJ
kw52j67652r	3	20	models	model	NOUN
kw52j67652r	3	21	for	for	ADP
kw52j67652r	3	22	training	training	NOUN
kw52j67652r	3	23	and	and	CCONJ
kw52j67652r	3	24	testing	testing	NOUN
kw52j67652r	3	25	.	.	PUNCT
kw52j67652r	4	1	one	one	NUM
kw52j67652r	4	2	problem	problem	NOUN
kw52j67652r	4	3	with	with	ADP
kw52j67652r	4	4	natural	natural	ADJ
kw52j67652r	4	5	language	language	NOUN
kw52j67652r	4	6	corpora	corpora	NOUN
kw52j67652r	4	7	is	be	AUX
kw52j67652r	4	8	the	the	DET
kw52j67652r	4	9	unbalanced	unbalanced	ADJ
kw52j67652r	4	10	frequency	frequency	NOUN
kw52j67652r	4	11	of	of	ADP
kw52j67652r	4	12	rare	rare	ADJ
kw52j67652r	4	13	terms	term	NOUN
kw52j67652r	4	14	against	against	ADP
kw52j67652r	4	15	commonly	commonly	ADV
kw52j67652r	4	16	used	use	VERB
kw52j67652r	4	17	words	word	NOUN
kw52j67652r	4	18	.	.	PUNCT
kw52j67652r	5	1	the	the	DET
kw52j67652r	5	2	word	word	NOUN
kw52j67652r	5	3	-	-	PUNCT
kw52j67652r	5	4	level	level	NOUN
kw52j67652r	5	5	frequency	frequency	NOUN
kw52j67652r	5	6	in	in	ADP
kw52j67652r	5	7	natural	natural	ADJ
kw52j67652r	5	8	language	language	NOUN
kw52j67652r	5	9	creates	create	VERB
kw52j67652r	5	10	irregular	irregular	ADJ
kw52j67652r	5	11	sparsity	sparsity	NOUN
kw52j67652r	5	12	patterns	pattern	NOUN
kw52j67652r	5	13	,	,	PUNCT
kw52j67652r	5	14	and	and	CCONJ
kw52j67652r	5	15	these	these	DET
kw52j67652r	5	16	patterns	pattern	NOUN
kw52j67652r	5	17	generate	generate	VERB
kw52j67652r	5	18	sparse	sparse	ADJ
kw52j67652r	5	19	data	datum	NOUN
kw52j67652r	5	20	structures	structure	NOUN
kw52j67652r	5	21	that	that	PRON
kw52j67652r	5	22	do	do	AUX
kw52j67652r	5	23	not	not	PART
kw52j67652r	5	24	perform	perform	VERB
kw52j67652r	5	25	well	well	ADV
kw52j67652r	5	26	on	on	ADP
kw52j67652r	5	27	parallel	parallel	ADJ
kw52j67652r	5	28	architectures	architecture	NOUN
kw52j67652r	5	29	.	.	PUNCT
kw52j67652r	6	1	asynchronous	asynchronous	ADJ
kw52j67652r	6	2	methods	method	NOUN
kw52j67652r	6	3	work	work	VERB
kw52j67652r	6	4	best	well	ADV
kw52j67652r	6	5	on	on	ADP
kw52j67652r	6	6	specific	specific	ADJ
kw52j67652r	6	7	sparse	sparse	ADJ
kw52j67652r	6	8	distributions	distribution	NOUN
kw52j67652r	6	9	.	.	PUNCT
kw52j67652r	7	1	ideally	ideally	ADV
kw52j67652r	7	2	,	,	PUNCT
kw52j67652r	7	3	the	the	DET
kw52j67652r	7	4	entire	entire	ADJ
kw52j67652r	7	5	computation	computation	NOUN
kw52j67652r	7	6	time	time	NOUN
kw52j67652r	7	7	should	should	AUX
kw52j67652r	7	8	be	be	AUX
kw52j67652r	7	9	spent	spend	VERB
kw52j67652r	7	10	on	on	ADP
kw52j67652r	7	11	dense	dense	ADJ
kw52j67652r	7	12	values	value	NOUN
kw52j67652r	7	13	only	only	ADV
kw52j67652r	7	14	,	,	PUNCT
kw52j67652r	7	15	and	and	CCONJ
kw52j67652r	7	16	computation	computation	NOUN
kw52j67652r	7	17	time	time	NOUN
kw52j67652r	7	18	on	on	ADP
kw52j67652r	7	19	sparse	sparse	ADJ
kw52j67652r	7	20	elements	element	NOUN
kw52j67652r	7	21	should	should	AUX
kw52j67652r	7	22	be	be	AUX
kw52j67652r	7	23	minimized.graphics	minimized.graphic	NOUN
kw52j67652r	7	24	processing	processing	NOUN
kw52j67652r	7	25	units	unit	NOUN
kw52j67652r	7	26	(	(	PUNCT
kw52j67652r	7	27	gpu	gpu	X
kw52j67652r	7	28	)	)	PUNCT
kw52j67652r	7	29	are	be	AUX
kw52j67652r	7	30	widely	widely	ADV
kw52j67652r	7	31	used	use	VERB
kw52j67652r	7	32	to	to	PART
kw52j67652r	7	33	process	process	VERB
kw52j67652r	7	34	a	a	DET
kw52j67652r	7	35	large	large	ADJ
kw52j67652r	7	36	quantity	quantity	NOUN
kw52j67652r	7	37	of	of	ADP
kw52j67652r	7	38	operations	operation	NOUN
kw52j67652r	7	39	in	in	ADP
kw52j67652r	7	40	parallel	parallel	NOUN
kw52j67652r	7	41	.	.	PUNCT
kw52j67652r	8	1	a	a	DET
kw52j67652r	8	2	problem	problem	NOUN
kw52j67652r	8	3	with	with	ADP
kw52j67652r	8	4	the	the	DET
kw52j67652r	8	5	use	use	NOUN
kw52j67652r	8	6	of	of	ADP
kw52j67652r	8	7	these	these	DET
kw52j67652r	8	8	accelerators	accelerator	NOUN
kw52j67652r	8	9	is	be	AUX
kw52j67652r	8	10	that	that	SCONJ
kw52j67652r	8	11	not	not	PART
kw52j67652r	8	12	all	all	DET
kw52j67652r	8	13	computation	computation	NOUN
kw52j67652r	8	14	problems	problem	NOUN
kw52j67652r	8	15	can	can	AUX
kw52j67652r	8	16	be	be	AUX
kw52j67652r	8	17	parallelized	parallelize	VERB
kw52j67652r	8	18	,	,	PUNCT
kw52j67652r	8	19	and	and	CCONJ
kw52j67652r	8	20	some	some	DET
kw52j67652r	8	21	parallel	parallel	ADJ
kw52j67652r	8	22	adaptations	adaptation	NOUN
kw52j67652r	8	23	run	run	VERB
kw52j67652r	8	24	slower	slow	ADJ
kw52j67652r	8	25	than	than	ADP
kw52j67652r	8	26	a	a	DET
kw52j67652r	8	27	serial	serial	ADJ
kw52j67652r	8	28	cpu	cpu	NOUN
kw52j67652r	8	29	counterpart	counterpart	NOUN
kw52j67652r	8	30	.	.	PUNCT
kw52j67652r	9	1	using	use	VERB
kw52j67652r	9	2	gpus	gpus	PROPN
kw52j67652r	9	3	to	to	PART
kw52j67652r	9	4	process	process	VERB
kw52j67652r	9	5	sparse	sparse	ADJ
kw52j67652r	9	6	structures	structure	NOUN
kw52j67652r	9	7	of	of	ADP
kw52j67652r	9	8	different	different	ADJ
kw52j67652r	9	9	sizes	size	NOUN
kw52j67652r	9	10	poses	pose	VERB
kw52j67652r	9	11	additional	additional	ADJ
kw52j67652r	9	12	problems	problem	NOUN
kw52j67652r	9	13	.	.	PUNCT
kw52j67652r	10	1	a	a	DET
kw52j67652r	10	2	large	large	ADJ
kw52j67652r	10	3	part	part	NOUN
kw52j67652r	10	4	of	of	ADP
kw52j67652r	10	5	the	the	DET
kw52j67652r	10	6	computation	computation	NOUN
kw52j67652r	10	7	time	time	NOUN
kw52j67652r	10	8	will	will	AUX
kw52j67652r	10	9	be	be	AUX
kw52j67652r	10	10	spent	spend	VERB
kw52j67652r	10	11	on	on	ADP
kw52j67652r	10	12	sparse	sparse	ADJ
kw52j67652r	10	13	regions	region	NOUN
kw52j67652r	10	14	if	if	SCONJ
kw52j67652r	10	15	the	the	DET
kw52j67652r	10	16	parallel	parallel	ADJ
kw52j67652r	10	17	implementations	implementation	NOUN
kw52j67652r	10	18	do	do	AUX
kw52j67652r	10	19	not	not	PART
kw52j67652r	10	20	take	take	VERB
kw52j67652r	10	21	advantage	advantage	NOUN
kw52j67652r	10	22	of	of	ADP
kw52j67652r	10	23	the	the	DET
kw52j67652r	10	24	partially	partially	ADV
kw52j67652r	10	25	dense	dense	ADJ
kw52j67652r	10	26	properties	property	NOUN
kw52j67652r	10	27	of	of	ADP
kw52j67652r	10	28	the	the	DET
kw52j67652r	10	29	input.significant	input.significant	NOUN
kw52j67652r	10	30	speedups	speedup	NOUN
kw52j67652r	10	31	are	be	AUX
kw52j67652r	10	32	achieved	achieve	VERB
kw52j67652r	10	33	when	when	SCONJ
kw52j67652r	10	34	a	a	DET
kw52j67652r	10	35	parallel	parallel	ADJ
kw52j67652r	10	36	implementation	implementation	NOUN
kw52j67652r	10	37	is	be	AUX
kw52j67652r	10	38	tailored	tailor	VERB
kw52j67652r	10	39	to	to	ADP
kw52j67652r	10	40	the	the	DET
kw52j67652r	10	41	sparsity	sparsity	NOUN
kw52j67652r	10	42	pattern	pattern	NOUN
kw52j67652r	10	43	of	of	ADP
kw52j67652r	10	44	the	the	DET
kw52j67652r	10	45	problem	problem	NOUN
kw52j67652r	10	46	being	be	AUX
kw52j67652r	10	47	solved	solve	VERB
kw52j67652r	10	48	and	and	CCONJ
kw52j67652r	10	49	the	the	DET
kw52j67652r	10	50	targeted	target	VERB
kw52j67652r	10	51	architecture	architecture	NOUN
kw52j67652r	10	52	.	.	PUNCT
kw52j67652r	11	1	our	our	PRON
kw52j67652r	11	2	work	work	NOUN
kw52j67652r	11	3	adapts	adapt	VERB
kw52j67652r	11	4	methods	method	NOUN
kw52j67652r	11	5	used	use	VERB
kw52j67652r	11	6	in	in	ADP
kw52j67652r	11	7	nlp	nlp	PROPN
kw52j67652r	11	8	to	to	PART
kw52j67652r	11	9	run	run	VERB
kw52j67652r	11	10	efficiently	efficiently	ADV
kw52j67652r	11	11	on	on	ADP
kw52j67652r	11	12	a	a	DET
kw52j67652r	11	13	parallel	parallel	ADJ
kw52j67652r	11	14	architecture	architecture	NOUN
kw52j67652r	11	15	using	use	VERB
kw52j67652r	11	16	high	high	ADJ
kw52j67652r	11	17	performance	performance	NOUN
kw52j67652r	11	18	computing	computing	NOUN
kw52j67652r	11	19	concepts	concept	NOUN
kw52j67652r	11	20	.	.	PUNCT
kw52j67652r	12	1	all	all	DET
kw52j67652r	12	2	contributions	contribution	NOUN
kw52j67652r	12	3	focus	focus	VERB
kw52j67652r	12	4	mainly	mainly	ADV
kw52j67652r	12	5	on	on	ADP
kw52j67652r	12	6	the	the	DET
kw52j67652r	12	7	gpu	gpu	NOUN
kw52j67652r	12	8	device	device	NOUN
kw52j67652r	12	9	designed	design	VERB
kw52j67652r	12	10	to	to	PART
kw52j67652r	12	11	carry	carry	VERB
kw52j67652r	12	12	out	out	ADP
kw52j67652r	12	13	a	a	DET
kw52j67652r	12	14	large	large	ADJ
kw52j67652r	12	15	amount	amount	NOUN
kw52j67652r	12	16	of	of	ADP
kw52j67652r	12	17	computations	computation	NOUN
kw52j67652r	12	18	faster	fast	ADV
kw52j67652r	12	19	than	than	ADP
kw52j67652r	12	20	several	several	ADJ
kw52j67652r	12	21	off	off	ADP
kw52j67652r	12	22	-	-	PUNCT
kw52j67652r	12	23	the	the	DET
kw52j67652r	12	24	-	-	PUNCT
kw52j67652r	12	25	shelf	shelf	NOUN
kw52j67652r	12	26	cpu	cpu	NOUN
kw52j67652r	12	27	architectures	architecture	NOUN
kw52j67652r	12	28	.	.	PUNCT
kw52j67652r	13	1	this	this	DET
kw52j67652r	13	2	dissertation	dissertation	NOUN
kw52j67652r	13	3	covers	cover	VERB
kw52j67652r	13	4	different	different	ADJ
kw52j67652r	13	5	adaptations	adaptation	NOUN
kw52j67652r	13	6	of	of	ADP
kw52j67652r	13	7	sparse	sparse	ADJ
kw52j67652r	13	8	nlp	nlp	NOUN
kw52j67652r	13	9	algorithms	algorithm	NOUN
kw52j67652r	13	10	to	to	ADP
kw52j67652r	13	11	the	the	DET
kw52j67652r	13	12	gpu	gpu	PROPN
kw52j67652r	13	13	architecture	architecture	NOUN
kw52j67652r	13	14	.	.	PUNCT
kw52j67652r	14	1	we	we	PRON
kw52j67652r	14	2	carry	carry	VERB
kw52j67652r	14	3	out	out	ADP
kw52j67652r	14	4	experiments	experiment	NOUN
kw52j67652r	14	5	using	use	VERB
kw52j67652r	14	6	different	different	ADJ
kw52j67652r	14	7	gpu	gpu	PROPN
kw52j67652r	14	8	architectures	architecture	NOUN
kw52j67652r	14	9	and	and	CCONJ
kw52j67652r	14	10	compare	compare	VERB
kw52j67652r	14	11	the	the	DET
kw52j67652r	14	12	performance	performance	NOUN
kw52j67652r	14	13	on	on	ADP
kw52j67652r	14	14	different	different	ADJ
kw52j67652r	14	15	datasets	dataset	NOUN
kw52j67652r	14	16	.	.	PUNCT
kw52j67652r	15	1	our	our	PRON
kw52j67652r	15	2	results	result	NOUN
kw52j67652r	15	3	demonstrate	demonstrate	VERB
kw52j67652r	15	4	that	that	SCONJ
kw52j67652r	15	5	gpu	gpu	NOUN
kw52j67652r	15	6	adaptations	adaptation	NOUN
kw52j67652r	15	7	can	can	AUX
kw52j67652r	15	8	significantly	significantly	ADV
kw52j67652r	15	9	reduce	reduce	VERB
kw52j67652r	15	10	the	the	DET
kw52j67652r	15	11	execution	execution	NOUN
kw52j67652r	15	12	time	time	NOUN
kw52j67652r	15	13	of	of	ADP
kw52j67652r	15	14	different	different	ADJ
kw52j67652r	15	15	sparse	sparse	ADJ
kw52j67652r	15	16	nlp	nlp	NOUN
kw52j67652r	15	17	algorithms	algorithm	NOUN
kw52j67652r	15	18	:	:	PUNCT
kw52j67652r	15	19	6000x	6000x	NUM
kw52j67652r	15	20	speedup	speedup	NOUN
kw52j67652r	15	21	on	on	ADP
kw52j67652r	15	22	the	the	DET
kw52j67652r	15	23	viterbi	viterbi	PROPN
kw52j67652r	15	24	task	task	PROPN
kw52j67652r	15	25	,	,	PUNCT
kw52j67652r	15	26	4.5x	4.5x	NUM
kw52j67652r	15	27	speedup	speedup	NOUN
kw52j67652r	15	28	on	on	ADP
kw52j67652r	15	29	the	the	DET
kw52j67652r	15	30	composition	composition	NOUN
kw52j67652r	15	31	task	task	NOUN
kw52j67652r	15	32	,	,	PUNCT
kw52j67652r	15	33	7x	7x	NUM
kw52j67652r	15	34	speedup	speedup	NOUN
kw52j67652r	15	35	on	on	ADP
kw52j67652r	15	36	a	a	DET
kw52j67652r	15	37	batched	batched	NOUN
kw52j67652r	15	38	forward	forward	ADV
kw52j67652r	15	39	-	-	PUNCT
kw52j67652r	15	40	backward	backward	ADJ
kw52j67652r	15	41	method	method	NOUN
kw52j67652r	15	42	,	,	PUNCT
kw52j67652r	15	43	and	and	CCONJ
kw52j67652r	15	44	50x	50x	NUM
kw52j67652r	15	45	improvement	improvement	NOUN
kw52j67652r	15	46	on	on	ADP
kw52j67652r	15	47	batched	batched	NOUN
kw52j67652r	15	48	operations	operation	NOUN
kw52j67652r	15	49	seen	see	VERB
kw52j67652r	15	50	in	in	ADP
kw52j67652r	15	51	deep	deep	ADJ
kw52j67652r	15	52	learning	learning	NOUN
kw52j67652r	15	53	.	.	PUNCT