id	sid	tid	token	lemma	pos
ajst-905	1	1	academic	academic	ADJ
ajst-905	1	2	journal	journal	NOUN
ajst-905	1	3	of	of	ADP
ajst-905	1	4	science	science	NOUN
ajst-905	1	5	and	and	CCONJ
ajst-905	1	6	technology	technology	NOUN
ajst-905	1	7	issn	issn	NOUN
ajst-905	1	8	:	:	PUNCT
ajst-905	1	9	2771	2771	NUM
ajst-905	1	10	-	-	SYM
ajst-905	1	11	3032	3032	NUM
ajst-905	1	12	|	|	NOUN
ajst-905	1	13	vol	vol	NOUN
ajst-905	1	14	.	.	PROPN
ajst-905	2	1	2	2	NUM
ajst-905	2	2	,	,	PUNCT
ajst-905	2	3	no	no	INTJ
ajst-905	2	4	.	.	NOUN
ajst-905	2	5	1	1	NUM
ajst-905	2	6	,	,	PUNCT
ajst-905	2	7	2022	2022	NUM
ajst-905	2	8	114	114	NUM
ajst-905	2	9	feature‐fusion	feature‐fusion	NOUN
ajst-905	2	10	parallel	parallel	ADJ
ajst-905	2	11	decoding	decoding	NOUN
ajst-905	2	12	transformer	transformer	NOUN
ajst-905	2	13	for	for	ADP
ajst-905	2	14	image	image	NOUN
ajst-905	2	15	captioning	caption	VERB
ajst-905	2	16	chenhao	chenhao	ADJ
ajst-905	2	17	zhu	zhu	PROPN
ajst-905	3	1	*	*	PROPN
ajst-905	3	2	,	,	PUNCT
ajst-905	3	3	xia	xia	PROPN
ajst-905	3	4	ye	ye	PROPN
ajst-905	3	5	,	,	PUNCT
ajst-905	3	6	qiduo	qiduo	ADV
ajst-905	3	7	lu	lu	PROPN
ajst-905	3	8	xi’an	xi’an	PROPN
ajst-905	3	9	research	research	PROPN
ajst-905	3	10	institute	institute	PROPN
ajst-905	3	11	of	of	ADP
ajst-905	3	12	high	high	ADJ
ajst-905	3	13	-	-	PUNCT
ajst-905	3	14	tech	tech	NOUN
ajst-905	3	15	,	,	PUNCT
ajst-905	3	16	xi’an	xi’an	PROPN
ajst-905	3	17	,	,	PUNCT
ajst-905	3	18	710025	710025	NUM
ajst-905	3	19	,	,	PUNCT
ajst-905	3	20	china	china	PROPN
ajst-905	4	1	*	*	SYM
ajst-905	4	2	972778371@qq.com	972778371@qq.com	NUM
ajst-905	4	3	abstract	abstract	ADJ
ajst-905	4	4	:	:	PUNCT
ajst-905	4	5	image	image	NOUN
ajst-905	4	6	caption	caption	NOUN
ajst-905	4	7	is	be	AUX
ajst-905	4	8	an	an	DET
ajst-905	4	9	important	important	ADJ
ajst-905	4	10	research	research	NOUN
ajst-905	4	11	direction	direction	NOUN
ajst-905	4	12	at	at	ADP
ajst-905	4	13	the	the	DET
ajst-905	4	14	intersection	intersection	NOUN
ajst-905	4	15	of	of	ADP
ajst-905	4	16	computer	computer	NOUN
ajst-905	4	17	vision	vision	NOUN
ajst-905	4	18	and	and	CCONJ
ajst-905	4	19	natural	natural	ADJ
ajst-905	4	20	language	language	NOUN
ajst-905	4	21	processing	processing	NOUN
ajst-905	4	22	.	.	PUNCT
ajst-905	5	1	it	it	PRON
ajst-905	5	2	is	be	AUX
ajst-905	5	3	based	base	VERB
ajst-905	5	4	on	on	ADP
ajst-905	5	5	object	object	NOUN
ajst-905	5	6	detection	detection	NOUN
ajst-905	5	7	,	,	PUNCT
ajst-905	5	8	enabling	enable	VERB
ajst-905	5	9	machines	machine	NOUN
ajst-905	5	10	to	to	PART
ajst-905	5	11	describe	describe	VERB
ajst-905	5	12	image	image	NOUN
ajst-905	5	13	content	content	NOUN
ajst-905	5	14	in	in	ADP
ajst-905	5	15	human	human	ADJ
ajst-905	5	16	language	language	NOUN
ajst-905	5	17	,	,	PUNCT
ajst-905	5	18	generating	generate	VERB
ajst-905	5	19	sentences	sentence	NOUN
ajst-905	5	20	with	with	ADP
ajst-905	5	21	correct	correct	ADJ
ajst-905	5	22	grammar	grammar	NOUN
ajst-905	5	23	.	.	PUNCT
ajst-905	6	1	most	most	ADJ
ajst-905	6	2	of	of	ADP
ajst-905	6	3	the	the	DET
ajst-905	6	4	existing	exist	VERB
ajst-905	6	5	methods	method	NOUN
ajst-905	6	6	employ	employ	VERB
ajst-905	6	7	a	a	DET
ajst-905	6	8	transformer	transformer	NOUN
ajst-905	6	9	-	-	PUNCT
ajst-905	6	10	based	base	VERB
ajst-905	6	11	structure	structure	NOUN
ajst-905	6	12	which	which	PRON
ajst-905	6	13	achieve	achieve	VERB
ajst-905	6	14	the	the	DET
ajst-905	6	15	cutting	cutting	NOUN
ajst-905	6	16	-	-	PUNCT
ajst-905	6	17	edge	edge	NOUN
ajst-905	6	18	performance	performance	NOUN
ajst-905	6	19	.	.	PUNCT
ajst-905	7	1	however	however	ADV
ajst-905	7	2	,	,	PUNCT
ajst-905	7	3	most	most	ADJ
ajst-905	7	4	methods	method	NOUN
ajst-905	7	5	focus	focus	VERB
ajst-905	7	6	on	on	ADP
ajst-905	7	7	improving	improve	VERB
ajst-905	7	8	visual	visual	ADJ
ajst-905	7	9	feature	feature	NOUN
ajst-905	7	10	information	information	NOUN
ajst-905	7	11	extraction	extraction	NOUN
ajst-905	7	12	,	,	PUNCT
ajst-905	7	13	optimizing	optimize	VERB
ajst-905	7	14	and	and	CCONJ
ajst-905	7	15	improving	improve	VERB
ajst-905	7	16	between	between	ADP
ajst-905	7	17	grid	grid	NOUN
ajst-905	7	18	features	feature	NOUN
ajst-905	7	19	and	and	CCONJ
ajst-905	7	20	region	region	NOUN
ajst-905	7	21	features	feature	NOUN
ajst-905	7	22	,	,	PUNCT
ajst-905	7	23	and	and	CCONJ
ajst-905	7	24	improving	improve	VERB
ajst-905	7	25	the	the	DET
ajst-905	7	26	performance	performance	NOUN
ajst-905	7	27	of	of	ADP
ajst-905	7	28	the	the	DET
ajst-905	7	29	final	final	ADJ
ajst-905	7	30	model	model	NOUN
ajst-905	7	31	.	.	PUNCT
ajst-905	8	1	in	in	ADP
ajst-905	8	2	this	this	DET
ajst-905	8	3	paper	paper	NOUN
ajst-905	8	4	,	,	PUNCT
ajst-905	8	5	we	we	PRON
ajst-905	8	6	tried	try	VERB
ajst-905	8	7	to	to	PART
ajst-905	8	8	improve	improve	VERB
ajst-905	8	9	the	the	DET
ajst-905	8	10	final	final	ADJ
ajst-905	8	11	effect	effect	NOUN
ajst-905	8	12	of	of	ADP
ajst-905	8	13	the	the	DET
ajst-905	8	14	model	model	NOUN
ajst-905	8	15	from	from	ADP
ajst-905	8	16	the	the	DET
ajst-905	8	17	perspective	perspective	NOUN
ajst-905	8	18	of	of	ADP
ajst-905	8	19	the	the	DET
ajst-905	8	20	model	model	NOUN
ajst-905	8	21	structure	structure	NOUN
ajst-905	8	22	and	and	CCONJ
ajst-905	8	23	the	the	DET
ajst-905	8	24	visual	visual	ADJ
ajst-905	8	25	features	feature	VERB
ajst-905	8	26	extraction	extraction	NOUN
ajst-905	8	27	.	.	PUNCT
ajst-905	9	1	we	we	PRON
ajst-905	9	2	proposed	propose	VERB
ajst-905	9	3	featurefusion	featurefusion	NOUN
ajst-905	9	4	parallel	parallel	ADJ
ajst-905	9	5	decoding	decode	VERB
ajst-905	9	6	transformer	transformer	NOUN
ajst-905	9	7	(	(	PUNCT
ajst-905	9	8	fpdt	fpdt	NOUN
ajst-905	9	9	)	)	PUNCT
ajst-905	9	10	which	which	PRON
ajst-905	9	11	adopts	adopt	VERB
ajst-905	9	12	parallel	parallel	ADJ
ajst-905	9	13	decoding	decode	VERB
ajst-905	9	14	mode	mode	NOUN
ajst-905	9	15	and	and	CCONJ
ajst-905	9	16	uses	use	VERB
ajst-905	9	17	both	both	DET
ajst-905	9	18	grid	grid	NOUN
ajst-905	9	19	features	feature	NOUN
ajst-905	9	20	and	and	CCONJ
ajst-905	9	21	region	region	NOUN
ajst-905	9	22	features	feature	NOUN
ajst-905	9	23	.	.	PUNCT
ajst-905	10	1	we	we	PRON
ajst-905	10	2	conducted	conduct	VERB
ajst-905	10	3	a	a	DET
ajst-905	10	4	large	large	ADJ
ajst-905	10	5	number	number	NOUN
ajst-905	10	6	of	of	ADP
ajst-905	10	7	experimental	experimental	ADJ
ajst-905	10	8	studies	study	NOUN
ajst-905	10	9	on	on	ADP
ajst-905	10	10	the	the	DET
ajst-905	10	11	mscoco	mscoco	ADJ
ajst-905	10	12	dataset	dataset	NOUN
ajst-905	10	13	.	.	PUNCT
ajst-905	11	1	and	and	CCONJ
ajst-905	11	2	fpdt	fpdt	NOUN
ajst-905	11	3	's	's	PART
ajst-905	11	4	performance	performance	NOUN
ajst-905	11	5	on	on	ADP
ajst-905	11	6	mscoco	mscoco	ADJ
ajst-905	11	7	datasets	dataset	NOUN
ajst-905	11	8	is	be	AUX
ajst-905	11	9	also	also	ADV
ajst-905	11	10	at	at	ADP
ajst-905	11	11	the	the	DET
ajst-905	11	12	cutting	cut	VERB
ajst-905	11	13	edge	edge	NOUN
ajst-905	11	14	.	.	PUNCT
ajst-905	12	1	keywords	keyword	NOUN
ajst-905	12	2	:	:	PUNCT
ajst-905	12	3	enter	enter	VERB
ajst-905	12	4	key	key	ADJ
ajst-905	12	5	words	word	NOUN
ajst-905	12	6	or	or	CCONJ
ajst-905	12	7	phrases	phrase	NOUN
ajst-905	12	8	in	in	ADP
ajst-905	12	9	alphabetical	alphabetical	ADJ
ajst-905	12	10	order	order	NOUN
ajst-905	12	11	,	,	PUNCT
ajst-905	12	12	separated	separate	VERB
ajst-905	12	13	by	by	ADP
ajst-905	12	14	commas	comma	NOUN
ajst-905	12	15	.	.	PUNCT
ajst-905	13	1	1	1	X
ajst-905	13	2	.	.	X
ajst-905	13	3	introduction	introduction	NOUN
ajst-905	13	4	image	image	NOUN
ajst-905	13	5	caption	caption	NOUN
ajst-905	13	6	has	have	VERB
ajst-905	13	7	many	many	ADJ
ajst-905	13	8	applications	application	NOUN
ajst-905	13	9	in	in	ADP
ajst-905	13	10	the	the	DET
ajst-905	13	11	real	real	ADJ
ajst-905	13	12	world	world	NOUN
ajst-905	13	13	,	,	PUNCT
ajst-905	13	14	such	such	ADJ
ajst-905	13	15	as	as	ADP
ajst-905	13	16	:	:	PUNCT
ajst-905	13	17	augmented	augment	VERB
ajst-905	13	18	reality	reality	NOUN
ajst-905	13	19	technology	technology	NOUN
ajst-905	13	20	,	,	PUNCT
ajst-905	13	21	helping	help	VERB
ajst-905	13	22	visually	visually	ADV
ajst-905	13	23	impaired	impair	VERB
ajst-905	13	24	people	people	NOUN
ajst-905	13	25	,	,	PUNCT
ajst-905	13	26	human	human	ADJ
ajst-905	13	27	-	-	PUNCT
ajst-905	13	28	machine	machine	NOUN
ajst-905	13	29	interaction	interaction	NOUN
ajst-905	13	30	technology	technology	NOUN
ajst-905	13	31	for	for	ADP
ajst-905	13	32	smart	smart	ADJ
ajst-905	13	33	glasses	glass	NOUN
ajst-905	13	34	,	,	PUNCT
ajst-905	13	35	etc	etc	X
ajst-905	13	36	.	.	X
ajst-905	14	1	it	it	PRON
ajst-905	14	2	is	be	AUX
ajst-905	14	3	an	an	DET
ajst-905	14	4	important	important	ADJ
ajst-905	14	5	technical	technical	ADJ
ajst-905	14	6	issue	issue	NOUN
ajst-905	14	7	in	in	ADP
ajst-905	14	8	the	the	DET
ajst-905	14	9	process	process	NOUN
ajst-905	14	10	of	of	ADP
ajst-905	14	11	realizing	realize	VERB
ajst-905	14	12	a	a	DET
ajst-905	14	13	high	high	ADJ
ajst-905	14	14	degree	degree	NOUN
ajst-905	14	15	of	of	ADP
ajst-905	14	16	intelligence	intelligence	NOUN
ajst-905	14	17	in	in	ADP
ajst-905	14	18	future	future	ADJ
ajst-905	14	19	machines	machine	NOUN
ajst-905	14	20	.	.	PUNCT
ajst-905	15	1	in	in	ADP
ajst-905	15	2	recent	recent	ADJ
ajst-905	15	3	years	year	NOUN
ajst-905	15	4	,	,	PUNCT
ajst-905	15	5	most	most	ADJ
ajst-905	15	6	image	image	NOUN
ajst-905	15	7	caption	caption	NOUN
ajst-905	15	8	tasks	task	NOUN
ajst-905	15	9	[	[	X
ajst-905	15	10	2	2	NUM
ajst-905	15	11	-	-	SYM
ajst-905	15	12	6	6	NUM
ajst-905	15	13	]	]	PUNCT
ajst-905	15	14	are	be	AUX
ajst-905	15	15	based	base	VERB
ajst-905	15	16	on	on	ADP
ajst-905	15	17	encoder	encoder	NOUN
ajst-905	15	18	-	-	PUNCT
ajst-905	15	19	decoder	decoder	NOUN
ajst-905	15	20	model	model	NOUN
ajst-905	15	21	which	which	PRON
ajst-905	15	22	has	have	AUX
ajst-905	15	23	shown	show	VERB
ajst-905	15	24	good	good	ADJ
ajst-905	15	25	performance	performance	NOUN
ajst-905	15	26	on	on	ADP
ajst-905	15	27	this	this	DET
ajst-905	15	28	task	task	NOUN
ajst-905	15	29	.	.	PUNCT
ajst-905	16	1	among	among	ADP
ajst-905	16	2	them	they	PRON
ajst-905	16	3	,	,	PUNCT
ajst-905	16	4	the	the	DET
ajst-905	16	5	cnn	cnn	PROPN
ajst-905	16	6	and	and	CCONJ
ajst-905	16	7	rnn	rnn	PROPN
ajst-905	16	8	are	be	AUX
ajst-905	16	9	used	use	VERB
ajst-905	16	10	as	as	ADP
ajst-905	16	11	encoders	encoder	NOUN
ajst-905	16	12	and	and	CCONJ
ajst-905	16	13	decoders	decoder	NOUN
ajst-905	16	14	to	to	PART
ajst-905	16	15	extract	extract	VERB
ajst-905	16	16	image	image	NOUN
ajst-905	16	17	features	feature	NOUN
ajst-905	16	18	and	and	CCONJ
ajst-905	16	19	generate	generate	VERB
ajst-905	16	20	corresponding	correspond	VERB
ajst-905	16	21	image	image	NOUN
ajst-905	16	22	descriptions	description	NOUN
ajst-905	16	23	,	,	PUNCT
ajst-905	16	24	respectively	respectively	ADV
ajst-905	16	25	.	.	PUNCT
ajst-905	17	1	after	after	SCONJ
ajst-905	17	2	the	the	DET
ajst-905	17	3	self	self	NOUN
ajst-905	17	4	-	-	PUNCT
ajst-905	17	5	attention	attention	NOUN
ajst-905	17	6	mechanism	mechanism	NOUN
ajst-905	17	7	and	and	CCONJ
ajst-905	17	8	transformer	transformer	NOUN
ajst-905	17	9	model	model	NOUN
ajst-905	17	10	was	be	AUX
ajst-905	17	11	proposed	propose	VERB
ajst-905	17	12	by	by	ADP
ajst-905	17	13	vaswani	vaswani	PROPN
ajst-905	17	14	et	et	NOUN
ajst-905	17	15	al.[7	al.[7	PROPN
ajst-905	17	16	]	]	PUNCT
ajst-905	17	17	,	,	PUNCT
ajst-905	17	18	transformer	transformer	NOUN
ajst-905	17	19	model	model	NOUN
ajst-905	17	20	and	and	CCONJ
ajst-905	17	21	attention	attention	NOUN
ajst-905	17	22	mechanism	mechanism	NOUN
ajst-905	17	23	have	have	AUX
ajst-905	17	24	been	be	AUX
ajst-905	17	25	proved	prove	VERB
ajst-905	17	26	to	to	PART
ajst-905	17	27	have	have	VERB
ajst-905	17	28	good	good	ADJ
ajst-905	17	29	mobility	mobility	NOUN
ajst-905	17	30	and	and	CCONJ
ajst-905	17	31	adaptability	adaptability	NOUN
ajst-905	17	32	to	to	ADP
ajst-905	17	33	multiple	multiple	ADJ
ajst-905	17	34	downstream	downstream	ADJ
ajst-905	17	35	tasks	task	NOUN
ajst-905	17	36	,	,	PUNCT
ajst-905	17	37	and	and	CCONJ
ajst-905	17	38	perform	perform	VERB
ajst-905	17	39	better	well	ADV
ajst-905	17	40	than	than	ADP
ajst-905	17	41	cnn	cnn	PROPN
ajst-905	17	42	and	and	CCONJ
ajst-905	17	43	rnn	rnn	VERB
ajst-905	17	44	in	in	ADP
ajst-905	17	45	image	image	NOUN
ajst-905	17	46	retrieval	retrieval	NOUN
ajst-905	17	47	,	,	PUNCT
ajst-905	17	48	natural	natural	ADJ
ajst-905	17	49	language	language	NOUN
ajst-905	17	50	understanding	understanding	NOUN
ajst-905	17	51	and	and	CCONJ
ajst-905	17	52	generation	generation	NOUN
ajst-905	17	53	tasks	task	NOUN
ajst-905	17	54	.	.	PUNCT
ajst-905	18	1	since	since	SCONJ
ajst-905	18	2	then	then	ADV
ajst-905	18	3	,	,	PUNCT
ajst-905	18	4	most	most	ADJ
ajst-905	18	5	efforts	effort	NOUN
ajst-905	18	6	to	to	PART
ajst-905	18	7	improve	improve	VERB
ajst-905	18	8	image	image	NOUN
ajst-905	18	9	captioning	captioning	NOUN
ajst-905	18	10	model	model	NOUN
ajst-905	18	11	focus	focus	NOUN
ajst-905	18	12	on	on	ADP
ajst-905	18	13	optimizing	optimize	VERB
ajst-905	18	14	visual	visual	ADJ
ajst-905	18	15	feature	feature	NOUN
ajst-905	18	16	extraction	extraction	NOUN
ajst-905	18	17	methods	method	NOUN
ajst-905	18	18	and	and	CCONJ
ajst-905	18	19	improving	improve	VERB
ajst-905	18	20	model	model	NOUN
ajst-905	18	21	structure	structure	NOUN
ajst-905	18	22	for	for	ADP
ajst-905	18	23	feature	feature	NOUN
ajst-905	18	24	processing	processing	NOUN
ajst-905	18	25	.	.	PUNCT
ajst-905	19	1	in	in	ADP
ajst-905	19	2	terms	term	NOUN
ajst-905	19	3	of	of	ADP
ajst-905	19	4	feature	feature	NOUN
ajst-905	19	5	extraction	extraction	NOUN
ajst-905	19	6	,	,	PUNCT
ajst-905	19	7	two	two	NUM
ajst-905	19	8	main	main	ADJ
ajst-905	19	9	methods	method	NOUN
ajst-905	19	10	are	be	AUX
ajst-905	19	11	region	region	NOUN
ajst-905	19	12	features	feature	NOUN
ajst-905	19	13	[	[	X
ajst-905	19	14	8	8	NUM
ajst-905	19	15	]	]	PUNCT
ajst-905	19	16	and	and	CCONJ
ajst-905	19	17	grid	grid	NOUN
ajst-905	19	18	features	feature	NOUN
ajst-905	19	19	[	[	X
ajst-905	19	20	4	4	X
ajst-905	19	21	]	]	PUNCT
ajst-905	19	22	respectively	respectively	ADV
ajst-905	19	23	.	.	PUNCT
ajst-905	20	1	at	at	ADP
ajst-905	20	2	first	first	ADV
ajst-905	20	3	,	,	PUNCT
ajst-905	20	4	region	region	NOUN
ajst-905	20	5	features	feature	VERB
ajst-905	20	6	outperform	outperform	ADJ
ajst-905	20	7	grid	grid	NOUN
ajst-905	20	8	features	feature	NOUN
ajst-905	20	9	.	.	PUNCT
ajst-905	21	1	because	because	SCONJ
ajst-905	21	2	region	region	NOUN
ajst-905	21	3	features	feature	NOUN
ajst-905	21	4	can	can	AUX
ajst-905	21	5	represent	represent	VERB
ajst-905	21	6	the	the	DET
ajst-905	21	7	object	object	NOUN
ajst-905	21	8	level	level	NOUN
ajst-905	21	9	information	information	NOUN
ajst-905	21	10	of	of	ADP
ajst-905	21	11	the	the	DET
ajst-905	21	12	most	most	ADV
ajst-905	21	13	significant	significant	ADJ
ajst-905	21	14	region	region	NOUN
ajst-905	21	15	in	in	ADP
ajst-905	21	16	the	the	DET
ajst-905	21	17	image	image	NOUN
ajst-905	21	18	,	,	PUNCT
ajst-905	21	19	it	it	PRON
ajst-905	21	20	is	be	AUX
ajst-905	21	21	represented	represent	VERB
ajst-905	21	22	by	by	ADP
ajst-905	21	23	feature	feature	NOUN
ajst-905	21	24	vectors	vector	NOUN
ajst-905	21	25	,	,	PUNCT
ajst-905	21	26	which	which	PRON
ajst-905	21	27	greatly	greatly	ADV
ajst-905	21	28	reduces	reduce	VERB
ajst-905	21	29	the	the	DET
ajst-905	21	30	difficulty	difficulty	NOUN
ajst-905	21	31	of	of	ADP
ajst-905	21	32	visual	visual	ADJ
ajst-905	21	33	semantic	semantic	ADJ
ajst-905	21	34	embedding	embedding	NOUN
ajst-905	21	35	and	and	CCONJ
ajst-905	21	36	thus	thus	ADV
ajst-905	21	37	improves	improve	VERB
ajst-905	21	38	the	the	DET
ajst-905	21	39	performance	performance	NOUN
ajst-905	21	40	of	of	ADP
ajst-905	21	41	the	the	DET
ajst-905	21	42	model	model	NOUN
ajst-905	21	43	.	.	PUNCT
ajst-905	22	1	however	however	ADV
ajst-905	22	2	,	,	PUNCT
ajst-905	22	3	it	it	PRON
ajst-905	22	4	does	do	AUX
ajst-905	22	5	still	still	ADV
ajst-905	22	6	have	have	VERB
ajst-905	22	7	many	many	ADJ
ajst-905	22	8	shortcomings	shortcoming	NOUN
ajst-905	22	9	,	,	PUNCT
ajst-905	22	10	such	such	ADJ
ajst-905	22	11	as	as	ADP
ajst-905	22	12	the	the	DET
ajst-905	22	13	loss	loss	NOUN
ajst-905	22	14	of	of	ADP
ajst-905	22	15	fine	fine	ADJ
ajst-905	22	16	granularity	granularity	NOUN
ajst-905	22	17	of	of	ADP
ajst-905	22	18	detection	detection	NOUN
ajst-905	22	19	,	,	PUNCT
ajst-905	22	20	low	low	ADJ
ajst-905	22	21	coverage	coverage	NOUN
ajst-905	22	22	of	of	ADP
ajst-905	22	23	image	image	NOUN
ajst-905	22	24	content	content	NOUN
ajst-905	22	25	and	and	CCONJ
ajst-905	22	26	timeconsuming	timeconsuming	ADJ
ajst-905	22	27	.	.	PUNCT
ajst-905	23	1	for	for	ADP
ajst-905	23	2	these	these	DET
ajst-905	23	3	problems	problem	NOUN
ajst-905	23	4	,	,	PUNCT
ajst-905	23	5	grid	grid	NOUN
ajst-905	23	6	features	feature	NOUN
ajst-905	23	7	can	can	AUX
ajst-905	23	8	be	be	AUX
ajst-905	23	9	a	a	DET
ajst-905	23	10	good	good	ADJ
ajst-905	23	11	solution	solution	NOUN
ajst-905	23	12	.	.	PUNCT
ajst-905	24	1	recently	recently	ADV
ajst-905	24	2	,	,	PUNCT
ajst-905	24	3	jiang	jiang	PROPN
ajst-905	24	4	et	et	PROPN
ajst-905	24	5	al	al	PROPN
ajst-905	24	6	.	.	PUNCT
ajst-905	25	1	[	[	X
ajst-905	25	2	9	9	NUM
ajst-905	25	3	]	]	PUNCT
ajst-905	25	4	demonstrated	demonstrate	VERB
ajst-905	25	5	that	that	SCONJ
ajst-905	25	6	grid	grid	NOUN
ajst-905	25	7	features	feature	VERB
ajst-905	25	8	not	not	PART
ajst-905	25	9	only	only	ADV
ajst-905	25	10	have	have	VERB
ajst-905	25	11	the	the	DET
ajst-905	25	12	same	same	ADJ
ajst-905	25	13	accuracy	accuracy	NOUN
ajst-905	25	14	as	as	ADP
ajst-905	25	15	region	region	NOUN
ajst-905	25	16	features	feature	NOUN
ajst-905	25	17	,	,	PUNCT
ajst-905	25	18	but	but	CCONJ
ajst-905	25	19	also	also	ADV
ajst-905	25	20	have	have	VERB
ajst-905	25	21	an	an	DET
ajst-905	25	22	order	order	NOUN
ajst-905	25	23	of	of	ADP
ajst-905	25	24	magnitude	magnitude	NOUN
ajst-905	25	25	faster	fast	ADJ
ajst-905	25	26	inference	inference	NOUN
ajst-905	25	27	speed	speed	NOUN
ajst-905	25	28	.	.	PUNCT
ajst-905	26	1	on	on	ADP
ajst-905	26	2	the	the	DET
ajst-905	26	3	basis	basis	NOUN
ajst-905	26	4	of	of	ADP
ajst-905	26	5	previous	previous	ADJ
ajst-905	26	6	work	work	NOUN
ajst-905	26	7	,	,	PUNCT
ajst-905	26	8	we	we	PRON
ajst-905	26	9	try	try	VERB
ajst-905	26	10	to	to	PART
ajst-905	26	11	fuse	fuse	VERB
ajst-905	26	12	grid	grid	NOUN
ajst-905	26	13	features	feature	NOUN
ajst-905	26	14	and	and	CCONJ
ajst-905	26	15	region	region	NOUN
ajst-905	26	16	features	feature	VERB
ajst-905	26	17	to	to	PART
ajst-905	26	18	improve	improve	VERB
ajst-905	26	19	the	the	DET
ajst-905	26	20	final	final	ADJ
ajst-905	26	21	performance	performance	NOUN
ajst-905	26	22	of	of	ADP
ajst-905	26	23	our	our	PRON
ajst-905	26	24	model	model	NOUN
ajst-905	26	25	.	.	PUNCT
ajst-905	27	1	the	the	DET
ajst-905	27	2	main	main	ADJ
ajst-905	27	3	contributions	contribution	NOUN
ajst-905	27	4	of	of	ADP
ajst-905	27	5	our	our	PRON
ajst-905	27	6	work	work	NOUN
ajst-905	27	7	are	be	AUX
ajst-905	27	8	summarized	summarize	VERB
ajst-905	27	9	as	as	SCONJ
ajst-905	27	10	follows	follow	VERB
ajst-905	27	11	:	:	PUNCT
ajst-905	28	1	1	1	X
ajst-905	28	2	.	.	X
ajst-905	28	3	we	we	PRON
ajst-905	28	4	propose	propose	VERB
ajst-905	28	5	the	the	DET
ajst-905	28	6	feature	feature	NOUN
ajst-905	28	7	augment	augment	NOUN
ajst-905	28	8	module	module	NOUN
ajst-905	28	9	to	to	PART
ajst-905	28	10	fuse	fuse	VERB
ajst-905	28	11	the	the	DET
ajst-905	28	12	region	region	NOUN
ajst-905	28	13	features	feature	NOUN
ajst-905	28	14	and	and	CCONJ
ajst-905	28	15	region	region	NOUN
ajst-905	28	16	features	feature	NOUN
ajst-905	28	17	.	.	PUNCT
ajst-905	29	1	the	the	DET
ajst-905	29	2	correlation	correlation	NOUN
ajst-905	29	3	between	between	ADP
ajst-905	29	4	features	feature	NOUN
ajst-905	29	5	can	can	AUX
ajst-905	29	6	be	be	AUX
ajst-905	29	7	judged	judge	VERB
ajst-905	29	8	by	by	ADP
ajst-905	29	9	calculating	calculate	VERB
ajst-905	29	10	the	the	DET
ajst-905	29	11	similarity	similarity	NOUN
ajst-905	29	12	between	between	ADP
ajst-905	29	13	them	they	PRON
ajst-905	29	14	.	.	PUNCT
ajst-905	30	1	2	2	X
ajst-905	30	2	.	.	X
ajst-905	30	3	for	for	ADP
ajst-905	30	4	the	the	DET
ajst-905	30	5	structure	structure	NOUN
ajst-905	30	6	of	of	ADP
ajst-905	30	7	the	the	DET
ajst-905	30	8	transformer	transformer	NOUN
ajst-905	30	9	,	,	PUNCT
ajst-905	30	10	we	we	PRON
ajst-905	30	11	propose	propose	VERB
ajst-905	30	12	the	the	DET
ajst-905	30	13	parallel	parallel	ADJ
ajst-905	30	14	input	input	NOUN
ajst-905	30	15	decoder	decoder	NOUN
ajst-905	30	16	module	module	NOUN
ajst-905	30	17	,	,	PUNCT
ajst-905	30	18	which	which	PRON
ajst-905	30	19	could	could	AUX
ajst-905	30	20	decode	decode	VERB
ajst-905	30	21	multilayer	multilayer	NOUN
ajst-905	30	22	encoding	encoding	NOUN
ajst-905	30	23	vectors	vector	NOUN
ajst-905	30	24	and	and	CCONJ
ajst-905	30	25	improve	improve	VERB
ajst-905	30	26	the	the	DET
ajst-905	30	27	performance	performance	NOUN
ajst-905	30	28	of	of	ADP
ajst-905	30	29	the	the	DET
ajst-905	30	30	model	model	NOUN
ajst-905	30	31	.	.	PUNCT
ajst-905	31	1	2	2	X
ajst-905	31	2	.	.	NUM
ajst-905	31	3	related	relate	VERB
ajst-905	31	4	work	work	NOUN
ajst-905	31	5	at	at	ADP
ajst-905	31	6	first	first	ADV
ajst-905	31	7	,	,	PUNCT
ajst-905	31	8	grid	grid	NOUN
ajst-905	31	9	features	feature	NOUN
ajst-905	31	10	of	of	ADP
ajst-905	31	11	fixed	fix	VERB
ajst-905	31	12	size	size	NOUN
ajst-905	31	13	extracted	extract	VERB
ajst-905	31	14	by	by	ADP
ajst-905	31	15	cnn	cnn	PROPN
ajst-905	31	16	are	be	AUX
ajst-905	31	17	used	use	VERB
ajst-905	31	18	as	as	ADP
ajst-905	31	19	the	the	DET
ajst-905	31	20	input	input	NOUN
ajst-905	31	21	of	of	ADP
ajst-905	31	22	encoder	encoder	NOUN
ajst-905	31	23	.	.	PUNCT
ajst-905	32	1	after	after	ADP
ajst-905	32	2	encoding	encoding	NOUN
ajst-905	32	3	,	,	PUNCT
ajst-905	32	4	the	the	DET
ajst-905	32	5	output	output	NOUN
ajst-905	32	6	of	of	ADP
ajst-905	32	7	encoder	encoder	NOUN
ajst-905	32	8	is	be	AUX
ajst-905	32	9	passed	pass	VERB
ajst-905	32	10	into	into	ADP
ajst-905	32	11	the	the	DET
ajst-905	32	12	decoder	decoder	NOUN
ajst-905	32	13	composed	compose	VERB
ajst-905	32	14	of	of	ADP
ajst-905	32	15	rnn	rnn	NOUN
ajst-905	32	16	or	or	CCONJ
ajst-905	32	17	lstm	lstm	NOUN
ajst-905	32	18	to	to	PART
ajst-905	32	19	get	get	VERB
ajst-905	32	20	the	the	DET
ajst-905	32	21	image	image	NOUN
ajst-905	32	22	description	description	NOUN
ajst-905	32	23	.	.	PUNCT
ajst-905	33	1	vinyas	vinyas	NOUN
ajst-905	33	2	et	et	PROPN
ajst-905	33	3	al	al	PROPN
ajst-905	33	4	.	.	PUNCT
ajst-905	34	1	[	[	X
ajst-905	34	2	4	4	X
ajst-905	34	3	]	]	PUNCT
ajst-905	34	4	proposed	propose	VERB
ajst-905	34	5	neural	neural	ADJ
ajst-905	34	6	image	image	NOUN
ajst-905	34	7	caption	caption	NOUN
ajst-905	34	8	model	model	NOUN
ajst-905	34	9	(	(	PUNCT
ajst-905	34	10	nic	nic	PROPN
ajst-905	34	11	)	)	PUNCT
ajst-905	34	12	.	.	PUNCT
ajst-905	35	1	nic	nic	PROPN
ajst-905	35	2	is	be	AUX
ajst-905	35	3	based	base	VERB
ajst-905	35	4	on	on	ADP
ajst-905	35	5	encoder	encoder	NOUN
ajst-905	35	6	-	-	PUNCT
ajst-905	35	7	decoder	decoder	NOUN
ajst-905	35	8	structure	structure	NOUN
ajst-905	35	9	.	.	PUNCT
ajst-905	36	1	it	it	PRON
ajst-905	36	2	uses	use	VERB
ajst-905	36	3	cnn	cnn	PROPN
ajst-905	36	4	as	as	ADP
ajst-905	36	5	encoder	encoder	NOUN
ajst-905	36	6	to	to	PART
ajst-905	36	7	extract	extract	VERB
ajst-905	36	8	visual	visual	ADJ
ajst-905	36	9	features	feature	NOUN
ajst-905	36	10	of	of	ADP
ajst-905	36	11	the	the	DET
ajst-905	36	12	image	image	NOUN
ajst-905	36	13	and	and	CCONJ
ajst-905	36	14	rnn	rnn	NOUN
ajst-905	36	15	as	as	ADP
ajst-905	36	16	decoder	decoder	NOUN
ajst-905	36	17	to	to	PART
ajst-905	36	18	generate	generate	VERB
ajst-905	36	19	descriptions	description	NOUN
ajst-905	36	20	.	.	PUNCT
ajst-905	37	1	xu	xu	PROPN
ajst-905	37	2	et	et	PROPN
ajst-905	38	1	al	al	PROPN
ajst-905	38	2	.	.	PUNCT
ajst-905	39	1	[	[	X
ajst-905	39	2	6	6	NUM
ajst-905	39	3	]	]	PUNCT
ajst-905	39	4	improved	improve	VERB
ajst-905	39	5	the	the	DET
ajst-905	39	6	work	work	NOUN
ajst-905	39	7	of	of	ADP
ajst-905	39	8	vinyas	vinyas	NOUN
ajst-905	39	9	et	et	PROPN
ajst-905	39	10	al	al	PROPN
ajst-905	39	11	.	.	PUNCT
ajst-905	40	1	[	[	X
ajst-905	40	2	4	4	NUM
ajst-905	40	3	]	]	PUNCT
ajst-905	40	4	.	.	PUNCT
ajst-905	41	1	they	they	PRON
ajst-905	41	2	used	use	VERB
ajst-905	41	3	the	the	DET
ajst-905	41	4	soft	soft	ADJ
ajst-905	41	5	attention	attention	NOUN
ajst-905	41	6	mechanism	mechanism	NOUN
ajst-905	41	7	to	to	PART
ajst-905	41	8	weight	weight	VERB
ajst-905	41	9	the	the	DET
ajst-905	41	10	image	image	NOUN
ajst-905	41	11	features	feature	VERB
ajst-905	41	12	with	with	ADP
ajst-905	41	13	the	the	DET
ajst-905	41	14	hidden	hide	VERB
ajst-905	41	15	layer	layer	NOUN
ajst-905	41	16	state	state	NOUN
ajst-905	41	17	of	of	ADP
ajst-905	41	18	lstm	lstm	NOUN
ajst-905	41	19	at	at	ADP
ajst-905	41	20	each	each	DET
ajst-905	41	21	time	time	NOUN
ajst-905	41	22	step	step	NOUN
ajst-905	41	23	,	,	PUNCT
ajst-905	41	24	and	and	CCONJ
ajst-905	41	25	input	input	VERB
ajst-905	41	26	the	the	DET
ajst-905	41	27	weighted	weight	VERB
ajst-905	41	28	features	feature	NOUN
ajst-905	41	29	and	and	CCONJ
ajst-905	41	30	words	word	NOUN
ajst-905	41	31	into	into	ADP
ajst-905	41	32	lstm	lstm	NOUN
ajst-905	41	33	to	to	PART
ajst-905	41	34	get	get	VERB
ajst-905	41	35	the	the	DET
ajst-905	41	36	results	result	NOUN
ajst-905	41	37	.	.	PUNCT
ajst-905	42	1	because	because	SCONJ
ajst-905	42	2	the	the	DET
ajst-905	42	3	hidden	hide	VERB
ajst-905	42	4	layer	layer	NOUN
ajst-905	42	5	state	state	NOUN
ajst-905	42	6	of	of	ADP
ajst-905	42	7	lstm	lstm	NOUN
ajst-905	42	8	contains	contain	VERB
ajst-905	42	9	context	context	NOUN
ajst-905	42	10	information	information	NOUN
ajst-905	42	11	,	,	PUNCT
ajst-905	42	12	this	this	DET
ajst-905	42	13	weighting	weighting	NOUN
ajst-905	42	14	can	can	AUX
ajst-905	42	15	select	select	VERB
ajst-905	42	16	important	important	ADJ
ajst-905	42	17	information	information	NOUN
ajst-905	42	18	to	to	ADP
ajst-905	42	19	the	the	DET
ajst-905	42	20	current	current	ADJ
ajst-905	42	21	time	time	NOUN
ajst-905	42	22	step	step	NOUN
ajst-905	42	23	to	to	ADP
ajst-905	42	24	a	a	DET
ajst-905	42	25	certain	certain	ADJ
ajst-905	42	26	extent	extent	NOUN
ajst-905	42	27	.	.	PUNCT
ajst-905	43	1	mao	mao	PROPN
ajst-905	43	2	et	et	PROPN
ajst-905	44	1	al	al	PROPN
ajst-905	44	2	.	.	PUNCT
ajst-905	45	1	[	[	X
ajst-905	45	2	13	13	NUM
ajst-905	45	3	]	]	PUNCT
ajst-905	45	4	proposed	propose	VERB
ajst-905	45	5	a	a	DET
ajst-905	45	6	multimodal	multimodal	ADJ
ajst-905	45	7	recurrent	recurrent	NOUN
ajst-905	45	8	neural	neural	ADJ
ajst-905	45	9	network	network	NOUN
ajst-905	45	10	model	model	NOUN
ajst-905	45	11	(	(	PUNCT
ajst-905	45	12	m	m	NOUN
ajst-905	45	13	-	-	PUNCT
ajst-905	45	14	rnn	rnn	PROPN
ajst-905	45	15	)	)	PUNCT
ajst-905	45	16	,	,	PUNCT
ajst-905	45	17	which	which	PRON
ajst-905	45	18	fused	fuse	VERB
ajst-905	45	19	hidden	hidden	ADJ
ajst-905	45	20	states	state	NOUN
ajst-905	45	21	of	of	ADP
ajst-905	45	22	rnn	rnn	NOUN
ajst-905	45	23	layers	layer	NOUN
ajst-905	45	24	,	,	PUNCT
ajst-905	45	25	word	word	NOUN
ajst-905	45	26	sequence	sequence	NOUN
ajst-905	45	27	features	feature	VERB
ajst-905	45	28	,	,	PUNCT
ajst-905	45	29	and	and	CCONJ
ajst-905	45	30	visual	visual	ADJ
ajst-905	45	31	features	feature	NOUN
ajst-905	45	32	to	to	ADP
ajst-905	45	33	generating	generate	VERB
ajst-905	45	34	captions	caption	NOUN
ajst-905	45	35	.	.	PUNCT
ajst-905	46	1	chen	chen	PROPN
ajst-905	46	2	et	et	PROPN
ajst-905	46	3	al	al	PROPN
ajst-905	46	4	.	.	PUNCT
ajst-905	47	1	[	[	X
ajst-905	47	2	11	11	NUM
ajst-905	47	3	]	]	PUNCT
ajst-905	47	4	proposed	propose	VERB
ajst-905	47	5	a	a	DET
ajst-905	47	6	sca	sca	PROPN
ajst-905	47	7	-	-	PUNCT
ajst-905	47	8	cnn	cnn	PROPN
ajst-905	47	9	model	model	NOUN
ajst-905	47	10	,	,	PUNCT
ajst-905	47	11	which	which	PRON
ajst-905	47	12	makes	make	VERB
ajst-905	47	13	full	full	ADJ
ajst-905	47	14	use	use	NOUN
ajst-905	47	15	of	of	ADP
ajst-905	47	16	the	the	DET
ajst-905	47	17	spatial	spatial	ADJ
ajst-905	47	18	and	and	CCONJ
ajst-905	47	19	channel	channel	NOUN
ajst-905	47	20	characteristics	characteristic	NOUN
ajst-905	47	21	of	of	ADP
ajst-905	47	22	cnn	cnn	PROPN
ajst-905	47	23	to	to	PART
ajst-905	47	24	decode	decode	VERB
ajst-905	47	25	at	at	ADP
ajst-905	47	26	each	each	DET
ajst-905	47	27	time	time	NOUN
ajst-905	47	28	step	step	NOUN
ajst-905	47	29	.	.	PUNCT
ajst-905	48	1	considering	consider	VERB
ajst-905	48	2	that	that	SCONJ
ajst-905	48	3	the	the	DET
ajst-905	48	4	generated	generate	VERB
ajst-905	48	5	words	word	NOUN
ajst-905	48	6	do	do	AUX
ajst-905	48	7	not	not	PART
ajst-905	48	8	always	always	ADV
ajst-905	48	9	depend	depend	VERB
ajst-905	48	10	on	on	ADP
ajst-905	48	11	visual	visual	ADJ
ajst-905	48	12	information	information	NOUN
ajst-905	48	13	,	,	PUNCT
ajst-905	48	14	lu	lu	PROPN
ajst-905	48	15	et	et	PROPN
ajst-905	48	16	al	al	PROPN
ajst-905	48	17	.	.	PUNCT
ajst-905	49	1	[	[	X
ajst-905	49	2	10	10	NUM
ajst-905	49	3	]	]	PUNCT
ajst-905	49	4	proposed	propose	VERB
ajst-905	49	5	an	an	DET
ajst-905	49	6	adaptive	adaptive	ADJ
ajst-905	49	7	attention	attention	NOUN
ajst-905	49	8	model	model	NOUN
ajst-905	49	9	with	with	ADP
ajst-905	49	10	visual	visual	ADJ
ajst-905	49	11	sentry	sentry	NOUN
ajst-905	49	12	,	,	PUNCT
ajst-905	49	13	which	which	PRON
ajst-905	49	14	can	can	AUX
ajst-905	49	15	decide	decide	VERB
ajst-905	49	16	when	when	SCONJ
ajst-905	49	17	to	to	PART
ajst-905	49	18	pay	pay	VERB
ajst-905	49	19	attention	attention	NOUN
ajst-905	49	20	to	to	ADP
ajst-905	49	21	visual	visual	ADJ
ajst-905	49	22	information	information	NOUN
ajst-905	49	23	and	and	CCONJ
ajst-905	49	24	language	language	NOUN
ajst-905	49	25	model	model	NOUN
ajst-905	49	26	.	.	PUNCT
ajst-905	50	1	with	with	ADP
ajst-905	50	2	the	the	DET
ajst-905	50	3	development	development	NOUN
ajst-905	50	4	of	of	ADP
ajst-905	50	5	feature	feature	NOUN
ajst-905	50	6	extraction	extraction	NOUN
ajst-905	50	7	methods	method	NOUN
ajst-905	50	8	,	,	PUNCT
ajst-905	50	9	region	region	NOUN
ajst-905	50	10	features	feature	VERB
ajst-905	50	11	[	[	X
ajst-905	50	12	8	8	NUM
ajst-905	50	13	]	]	PUNCT
ajst-905	50	14	bring	bring	VERB
ajst-905	50	15	improved	improved	ADJ
ajst-905	50	16	performance	performance	NOUN
ajst-905	50	17	to	to	PART
ajst-905	50	18	image	image	VERB
ajst-905	50	19	subtitle	subtitle	NOUN
ajst-905	50	20	models	model	NOUN
ajst-905	50	21	.	.	PUNCT
ajst-905	51	1	herdade	herdade	PROPN
ajst-905	51	2	et	et	PROPN
ajst-905	51	3	al	al	PROPN
ajst-905	51	4	.	.	PUNCT
ajst-905	52	1	[	[	X
ajst-905	52	2	12	12	NUM
ajst-905	52	3	]	]	PUNCT
ajst-905	52	4	proposed	propose	VERB
ajst-905	52	5	to	to	PART
ajst-905	52	6	integrate	integrate	VERB
ajst-905	52	7	the	the	DET
ajst-905	52	8	geometric	geometric	ADJ
ajst-905	52	9	relationship	relationship	NOUN
ajst-905	52	10	information	information	NOUN
ajst-905	52	11	between	between	ADP
ajst-905	52	12	regional	regional	ADJ
ajst-905	52	13	features	feature	NOUN
ajst-905	52	14	into	into	ADP
ajst-905	52	15	the	the	DET
ajst-905	52	16	decoding	decode	VERB
ajst-905	52	17	process	process	NOUN
ajst-905	52	18	,	,	PUNCT
ajst-905	52	19	which	which	PRON
ajst-905	52	20	enhanced	enhance	VERB
ajst-905	52	21	the	the	DET
ajst-905	52	22	interaction	interaction	NOUN
ajst-905	52	23	115	115	NUM
ajst-905	52	24	between	between	ADP
ajst-905	52	25	regional	regional	ADJ
ajst-905	52	26	features	feature	NOUN
ajst-905	52	27	.	.	PUNCT
ajst-905	53	1	cornia	cornia	PROPN
ajst-905	53	2	et	et	PROPN
ajst-905	53	3	al	al	PROPN
ajst-905	53	4	.	.	PUNCT
ajst-905	54	1	[	[	X
ajst-905	54	2	14	14	NUM
ajst-905	54	3	]	]	PUNCT
ajst-905	54	4	proposed	propose	VERB
ajst-905	54	5	the	the	DET
ajst-905	54	6	meshed	mesh	VERB
ajst-905	54	7	-	-	PUNCT
ajst-905	54	8	memory	memory	NOUN
ajst-905	54	9	transformer	transformer	NOUN
ajst-905	54	10	model	model	NOUN
ajst-905	54	11	,	,	PUNCT
ajst-905	54	12	which	which	PRON
ajst-905	54	13	uses	use	VERB
ajst-905	54	14	region	region	NOUN
ajst-905	54	15	features	feature	VERB
ajst-905	54	16	to	to	PART
ajst-905	54	17	represent	represent	VERB
ajst-905	54	18	the	the	DET
ajst-905	54	19	images	image	NOUN
ajst-905	54	20	.	.	PUNCT
ajst-905	55	1	it	it	PRON
ajst-905	55	2	also	also	ADV
ajst-905	55	3	meshed	mesh	VERB
ajst-905	55	4	the	the	DET
ajst-905	55	5	encoder	encoder	NOUN
ajst-905	55	6	and	and	CCONJ
ajst-905	55	7	decoder	decoder	NOUN
ajst-905	55	8	connectivity	connectivity	NOUN
ajst-905	55	9	relationships	relationship	NOUN
ajst-905	55	10	.	.	PUNCT
ajst-905	56	1	although	although	SCONJ
ajst-905	56	2	the	the	DET
ajst-905	56	3	use	use	NOUN
ajst-905	56	4	of	of	ADP
ajst-905	56	5	region	region	NOUN
ajst-905	56	6	features	feature	NOUN
ajst-905	56	7	improves	improve	VERB
ajst-905	56	8	the	the	DET
ajst-905	56	9	performance	performance	NOUN
ajst-905	56	10	of	of	ADP
ajst-905	56	11	the	the	DET
ajst-905	56	12	image	image	NOUN
ajst-905	56	13	caption	caption	NOUN
ajst-905	56	14	models	model	NOUN
ajst-905	56	15	,	,	PUNCT
ajst-905	56	16	region	region	NOUN
ajst-905	56	17	features	feature	NOUN
ajst-905	56	18	still	still	ADV
ajst-905	56	19	have	have	VERB
ajst-905	56	20	some	some	DET
ajst-905	56	21	problems	problem	NOUN
ajst-905	56	22	such	such	ADJ
ajst-905	56	23	as	as	ADP
ajst-905	56	24	time	time	NOUN
ajst-905	56	25	-	-	PUNCT
ajst-905	56	26	consuming	consume	VERB
ajst-905	56	27	and	and	CCONJ
ajst-905	56	28	low	low	ADJ
ajst-905	56	29	coverage	coverage	NOUN
ajst-905	56	30	of	of	ADP
ajst-905	56	31	image	image	NOUN
ajst-905	56	32	details	detail	NOUN
ajst-905	56	33	.	.	PUNCT
ajst-905	57	1	however	however	ADV
ajst-905	57	2	,	,	PUNCT
ajst-905	57	3	these	these	PRON
ajst-905	57	4	are	be	AUX
ajst-905	57	5	precisely	precisely	ADV
ajst-905	57	6	the	the	DET
ajst-905	57	7	advantages	advantage	NOUN
ajst-905	57	8	of	of	ADP
ajst-905	57	9	the	the	DET
ajst-905	57	10	grid	grid	NOUN
ajst-905	57	11	features	feature	NOUN
ajst-905	57	12	of	of	ADP
ajst-905	57	13	traditional	traditional	ADJ
ajst-905	57	14	methods	method	NOUN
ajst-905	57	15	.	.	PUNCT
ajst-905	58	1	the	the	DET
ajst-905	58	2	rstnet	rstnet	NOUN
ajst-905	58	3	proposed	propose	VERB
ajst-905	58	4	by	by	ADP
ajst-905	58	5	zhang	zhang	PROPN
ajst-905	58	6	et	et	PROPN
ajst-905	58	7	al	al	PROPN
ajst-905	58	8	.	.	PUNCT
ajst-905	59	1	[	[	X
ajst-905	59	2	15	15	NUM
ajst-905	59	3	]	]	X
ajst-905	59	4	uses	use	VERB
ajst-905	59	5	grid	grid	NOUN
ajst-905	59	6	features	feature	NOUN
ajst-905	59	7	to	to	PART
ajst-905	59	8	represent	represent	VERB
ajst-905	59	9	the	the	DET
ajst-905	59	10	images	image	NOUN
ajst-905	59	11	and	and	CCONJ
ajst-905	59	12	achieves	achieve	VERB
ajst-905	59	13	good	good	ADJ
ajst-905	59	14	results	result	NOUN
ajst-905	59	15	.	.	PUNCT
ajst-905	60	1	on	on	ADP
ajst-905	60	2	the	the	DET
ajst-905	60	3	basis	basis	NOUN
ajst-905	60	4	of	of	ADP
ajst-905	60	5	previous	previous	ADJ
ajst-905	60	6	works	work	NOUN
ajst-905	60	7	,	,	PUNCT
ajst-905	60	8	we	we	PRON
ajst-905	60	9	try	try	VERB
ajst-905	60	10	to	to	PART
ajst-905	60	11	fuse	fuse	VERB
ajst-905	60	12	region	region	NOUN
ajst-905	60	13	features	feature	NOUN
ajst-905	60	14	and	and	CCONJ
ajst-905	60	15	grid	grid	NOUN
ajst-905	60	16	features	feature	NOUN
ajst-905	60	17	to	to	PART
ajst-905	60	18	enrich	enrich	VERB
ajst-905	60	19	the	the	DET
ajst-905	60	20	visual	visual	ADJ
ajst-905	60	21	feature	feature	NOUN
ajst-905	60	22	information	information	NOUN
ajst-905	60	23	and	and	CCONJ
ajst-905	60	24	give	give	VERB
ajst-905	60	25	consideration	consideration	NOUN
ajst-905	60	26	to	to	ADP
ajst-905	60	27	the	the	DET
ajst-905	60	28	subject	subject	ADJ
ajst-905	60	29	object	object	NOUN
ajst-905	60	30	and	and	CCONJ
ajst-905	60	31	detail	detail	NOUN
ajst-905	60	32	content	content	NOUN
ajst-905	60	33	of	of	ADP
ajst-905	60	34	the	the	DET
ajst-905	60	35	image	image	NOUN
ajst-905	60	36	,	,	PUNCT
ajst-905	60	37	so	so	SCONJ
ajst-905	60	38	as	as	SCONJ
ajst-905	60	39	to	to	PART
ajst-905	60	40	achieve	achieve	VERB
ajst-905	60	41	better	well	ADJ
ajst-905	60	42	results	result	NOUN
ajst-905	60	43	in	in	ADP
ajst-905	60	44	the	the	DET
ajst-905	60	45	final	final	ADJ
ajst-905	60	46	description	description	NOUN
ajst-905	60	47	.	.	PUNCT
ajst-905	61	1	figure	figure	NOUN
ajst-905	61	2	1	1	NUM
ajst-905	61	3	.	.	PUNCT
ajst-905	62	1	the	the	DET
ajst-905	62	2	overall	overall	ADJ
ajst-905	62	3	architecture	architecture	NOUN
ajst-905	62	4	of	of	ADP
ajst-905	62	5	fpdt	fpdt	ADJ
ajst-905	62	6	model	model	NOUN
ajst-905	62	7	.	.	PUNCT
ajst-905	63	1	the	the	DET
ajst-905	63	2	specific	specific	ADJ
ajst-905	63	3	structure	structure	NOUN
ajst-905	63	4	of	of	ADP
ajst-905	63	5	the	the	DET
ajst-905	63	6	encoder	encoder	NOUN
ajst-905	63	7	layer	layer	NOUN
ajst-905	63	8	is	be	AUX
ajst-905	63	9	shown	show	VERB
ajst-905	63	10	in	in	ADP
ajst-905	63	11	fig.2	fig.2	PROPN
ajst-905	63	12	.	.	PUNCT
ajst-905	64	1	3	3	X
ajst-905	64	2	.	.	PUNCT
ajst-905	64	3	model	model	NOUN
ajst-905	64	4	architecture	architecture	NOUN
ajst-905	64	5	3.1	3.1	NUM
ajst-905	64	6	.	.	PUNCT
ajst-905	65	1	overall	overall	ADJ
ajst-905	65	2	architecture	architecture	NOUN
ajst-905	65	3	the	the	DET
ajst-905	65	4	overall	overall	ADJ
ajst-905	65	5	structure	structure	NOUN
ajst-905	65	6	of	of	ADP
ajst-905	65	7	the	the	DET
ajst-905	65	8	model	model	NOUN
ajst-905	65	9	is	be	AUX
ajst-905	65	10	shown	show	VERB
ajst-905	65	11	in	in	ADP
ajst-905	65	12	figure	figure	NOUN
ajst-905	65	13	1	1	NUM
ajst-905	65	14	.	.	PUNCT
ajst-905	66	1	we	we	PRON
ajst-905	66	2	take	take	VERB
ajst-905	66	3	two	two	NUM
ajst-905	66	4	features	feature	NOUN
ajst-905	66	5	as	as	ADP
ajst-905	66	6	inputs	input	NOUN
ajst-905	66	7	to	to	ADP
ajst-905	66	8	the	the	DET
ajst-905	66	9	fpdt	fpdt	PROPN
ajst-905	66	10	model	model	NOUN
ajst-905	66	11	’s	’s	PART
ajst-905	66	12	encoder	encoder	NOUN
ajst-905	66	13	.	.	PUNCT
ajst-905	67	1	the	the	DET
ajst-905	67	2	structure	structure	NOUN
ajst-905	67	3	and	and	CCONJ
ajst-905	67	4	principles	principle	NOUN
ajst-905	67	5	of	of	ADP
ajst-905	67	6	each	each	DET
ajst-905	67	7	part	part	NOUN
ajst-905	67	8	are	be	AUX
ajst-905	67	9	described	describe	VERB
ajst-905	67	10	next	next	ADV
ajst-905	67	11	.	.	PUNCT
ajst-905	68	1	3.2	3.2	NUM
ajst-905	68	2	.	.	PUNCT
ajst-905	68	3	feature	feature	NOUN
ajst-905	68	4	-	-	PUNCT
ajst-905	68	5	fusion	fusion	NOUN
ajst-905	68	6	encoder	encoder	NOUN
ajst-905	68	7	encoder	encoder	NOUN
ajst-905	68	8	layer	layer	NOUN
ajst-905	68	9	.	.	PUNCT
ajst-905	69	1	the	the	DET
ajst-905	69	2	structure	structure	NOUN
ajst-905	69	3	of	of	ADP
ajst-905	69	4	the	the	DET
ajst-905	69	5	encoder	encoder	NOUN
ajst-905	69	6	layer	layer	NOUN
ajst-905	69	7	is	be	AUX
ajst-905	69	8	shown	show	VERB
ajst-905	69	9	in	in	ADP
ajst-905	69	10	figure	figure	NOUN
ajst-905	69	11	2	2	NUM
ajst-905	69	12	,	,	PUNCT
ajst-905	69	13	where	where	SCONJ
ajst-905	69	14	addnorm	addnorm	NOUN
ajst-905	69	15	represents	represent	VERB
ajst-905	69	16	the	the	DET
ajst-905	69	17	residual	residual	ADJ
ajst-905	69	18	join	join	NOUN
ajst-905	69	19	and	and	CCONJ
ajst-905	69	20	normalization	normalization	NOUN
ajst-905	69	21	operation	operation	NOUN
ajst-905	69	22	.	.	PUNCT
ajst-905	70	1	before	before	SCONJ
ajst-905	70	2	the	the	DET
ajst-905	70	3	features	feature	NOUN
ajst-905	70	4	are	be	AUX
ajst-905	70	5	fed	feed	VERB
ajst-905	70	6	into	into	ADP
ajst-905	70	7	the	the	DET
ajst-905	70	8	encoder	encoder	NOUN
ajst-905	70	9	layer	layer	NOUN
ajst-905	70	10	,	,	PUNCT
ajst-905	70	11	we	we	PRON
ajst-905	70	12	fuse	fuse	VERB
ajst-905	70	13	the	the	DET
ajst-905	70	14	region	region	NOUN
ajst-905	70	15	features	feature	VERB
ajst-905	70	16	and	and	CCONJ
ajst-905	70	17	grid	grid	NOUN
ajst-905	70	18	features	feature	NOUN
ajst-905	70	19	through	through	ADP
ajst-905	70	20	the	the	DET
ajst-905	70	21	feature	feature	NOUN
ajst-905	70	22	augment	augment	NOUN
ajst-905	70	23	module	module	NOUN
ajst-905	70	24	.	.	PUNCT
ajst-905	71	1	this	this	DET
ajst-905	71	2	module	module	NOUN
ajst-905	71	3	can	can	AUX
ajst-905	71	4	fuse	fuse	VERB
ajst-905	71	5	the	the	DET
ajst-905	71	6	related	relate	VERB
ajst-905	71	7	regional	regional	ADJ
ajst-905	71	8	features	feature	NOUN
ajst-905	71	9	and	and	CCONJ
ajst-905	71	10	grid	grid	NOUN
ajst-905	71	11	features	feature	NOUN
ajst-905	71	12	,	,	PUNCT
ajst-905	71	13	and	and	CCONJ
ajst-905	71	14	retain	retain	VERB
ajst-905	71	15	the	the	DET
ajst-905	71	16	non	non	ADJ
ajst-905	71	17	-	-	ADJ
ajst-905	71	18	related	related	ADJ
ajst-905	71	19	features	feature	NOUN
ajst-905	71	20	,	,	PUNCT
ajst-905	71	21	so	so	SCONJ
ajst-905	71	22	as	as	SCONJ
ajst-905	71	23	to	to	PART
ajst-905	71	24	reduce	reduce	VERB
ajst-905	71	25	noise	noise	NOUN
ajst-905	71	26	and	and	CCONJ
ajst-905	71	27	information	information	NOUN
ajst-905	71	28	repetition	repetition	NOUN
ajst-905	71	29	.	.	PUNCT
ajst-905	72	1	and	and	CCONJ
ajst-905	72	2	the	the	DET
ajst-905	72	3	complete	complete	ADJ
ajst-905	72	4	operator	operator	NOUN
ajst-905	72	5	is	be	AUX
ajst-905	72	6	defined	define	VERB
ajst-905	72	7	as	as	SCONJ
ajst-905	72	8	follows	follow	VERB
ajst-905	72	9	:	:	PUNCT
ajst-905	72	10	max(cos	max(cos	PROPN
ajst-905	72	11	(	(	PUNCT
ajst-905	72	12	,	,	PUNCT
ajst-905	72	13	)	)	PUNCT
ajst-905	72	14	,	,	PUNCT
ajst-905	72	15	)	)	PUNCT
ajst-905	73	1	ij	ij	INTJ
ajst-905	73	2	i	i	PRON
ajst-905	73	3	js	js	VERB
ajst-905	73	4	g	g	PROPN
ajst-905	73	5	r	r	NOUN
ajst-905	73	6	t	t	PROPN
ajst-905	73	7			PROPN
ajst-905	74	1			INTJ
ajst-905	74	2	(	(	PUNCT
ajst-905	74	3	1	1	NUM
ajst-905	74	4	)	)	PUNCT
ajst-905	74	5	,	,	PUNCT
ajst-905	75	1	i	i	PRON
ajst-905	75	2	j	j	VERB
ajst-905	76	1	ij	ij	INTJ
ajst-905	77	1	i	i	PRON
ajst-905	77	2	j	j	PROPN
ajst-905	78	1	ij	ij	INTJ
ajst-905	78	2	g	g	NOUN
ajst-905	78	3	r	r	NOUN
ajst-905	78	4	s	s	PROPN
ajst-905	78	5	t	t	NOUN
ajst-905	78	6	p	p	NOUN
ajst-905	78	7	g	g	PROPN
ajst-905	78	8	r	r	NOUN
ajst-905	78	9	s	s	PROPN
ajst-905	78	10	t	t	NOUN
ajst-905	78	11			PROPN
ajst-905	78	12			PROPN
ajst-905	79	1			NUM
ajst-905	79	2			PRON
ajst-905	79	3	(	(	PUNCT
ajst-905	79	4	2	2	X
ajst-905	79	5	)	)	PUNCT
ajst-905	79	6	where	where	SCONJ
ajst-905	79	7			ADJ
ajst-905	79	8	in	in	ADP
ajst-905	79	9	eq	eq	PROPN
ajst-905	79	10	.	.	PUNCT
ajst-905	80	1	(	(	PUNCT
ajst-905	80	2	8)	8)	NUM
ajst-905	80	3	represents	represent	VERB
ajst-905	80	4	the	the	DET
ajst-905	80	5	addition	addition	NOUN
ajst-905	80	6	of	of	ADP
ajst-905	80	7	matrix	matrix	NOUN
ajst-905	80	8	elements	element	NOUN
ajst-905	80	9	.	.	PUNCT
ajst-905	81	1	the	the	DET
ajst-905	81	2	max	max	PROPN
ajst-905	81	3	(	(	PUNCT
ajst-905	81	4	)	)	PUNCT
ajst-905	81	5	is	be	AUX
ajst-905	81	6	the	the	DET
ajst-905	81	7	maximum	maximum	ADJ
ajst-905	81	8	function	function	NOUN
ajst-905	81	9	,	,	PUNCT
ajst-905	81	10	which	which	PRON
ajst-905	81	11	takes	take	VERB
ajst-905	81	12	the	the	DET
ajst-905	81	13	maximum	maximum	NOUN
ajst-905	81	14	between	between	ADP
ajst-905	81	15	the	the	DET
ajst-905	81	16	similarity	similarity	NOUN
ajst-905	81	17	score	score	NOUN
ajst-905	81	18	𝑆	𝑆	PROPN
ajst-905	81	19	and	and	CCONJ
ajst-905	81	20	𝑡	𝑡	PROPN
ajst-905	81	21	.	.	PUNCT
ajst-905	82	1	note	note	VERB
ajst-905	82	2	that	that	SCONJ
ajst-905	82	3	t	t	PROPN
ajst-905	82	4	is	be	AUX
ajst-905	82	5	a	a	DET
ajst-905	82	6	threshold	threshold	NOUN
ajst-905	82	7	value	value	NOUN
ajst-905	82	8	used	use	VERB
ajst-905	82	9	to	to	PART
ajst-905	82	10	determine	determine	VERB
ajst-905	82	11	whether	whether	SCONJ
ajst-905	82	12	two	two	NUM
ajst-905	82	13	features	feature	NOUN
ajst-905	82	14	represent	represent	VERB
ajst-905	82	15	the	the	DET
ajst-905	82	16	same	same	ADJ
ajst-905	82	17	region	region	NOUN
ajst-905	82	18	in	in	ADP
ajst-905	82	19	the	the	DET
ajst-905	82	20	image	image	NOUN
ajst-905	82	21	.	.	PUNCT
ajst-905	83	1	if	if	SCONJ
ajst-905	83	2	the	the	DET
ajst-905	83	3	similarity	similarity	NOUN
ajst-905	83	4	score	score	NOUN
ajst-905	83	5	𝑆	𝑆	PROPN
ajst-905	83	6	is	be	AUX
ajst-905	83	7	smaller	small	ADJ
ajst-905	83	8	than	than	ADP
ajst-905	83	9	threshold	threshold	NOUN
ajst-905	83	10	value	value	NOUN
ajst-905	83	11	𝑡	𝑡	PROPN
ajst-905	83	12	,	,	PUNCT
ajst-905	83	13	it	it	PRON
ajst-905	83	14	will	will	AUX
ajst-905	83	15	retain	retain	VERB
ajst-905	83	16	the	the	DET
ajst-905	83	17	region	region	NOUN
ajst-905	83	18	features	feature	NOUN
ajst-905	83	19	and	and	CCONJ
ajst-905	83	20	grid	grid	NOUN
ajst-905	83	21	features	feature	NOUN
ajst-905	83	22	.	.	PUNCT
ajst-905	84	1	the	the	DET
ajst-905	84	2	cos	cos	PROPN
ajst-905	84	3	represents	represent	VERB
ajst-905	84	4	the	the	DET
ajst-905	84	5	cosine	cosine	NOUN
ajst-905	84	6	similarity	similarity	NOUN
ajst-905	84	7	function	function	NOUN
ajst-905	84	8	which	which	PRON
ajst-905	84	9	used	use	VERB
ajst-905	84	10	to	to	PART
ajst-905	84	11	measure	measure	VERB
ajst-905	84	12	the	the	DET
ajst-905	84	13	correlation	correlation	NOUN
ajst-905	84	14	between	between	ADP
ajst-905	84	15	region	region	NOUN
ajst-905	84	16	features	feature	NOUN
ajst-905	84	17	and	and	CCONJ
ajst-905	84	18	grid	grid	NOUN
ajst-905	84	19	features	feature	NOUN
ajst-905	84	20	.	.	PUNCT
ajst-905	85	1	116	116	NUM
ajst-905	85	2	figure	figure	NOUN
ajst-905	85	3	2	2	NUM
ajst-905	85	4	.	.	PUNCT
ajst-905	86	1	the	the	DET
ajst-905	86	2	structure	structure	NOUN
ajst-905	86	3	of	of	ADP
ajst-905	86	4	the	the	DET
ajst-905	86	5	encoder	encoder	NOUN
ajst-905	86	6	layer	layer	NOUN
ajst-905	86	7	.	.	PUNCT
ajst-905	87	1	after	after	ADP
ajst-905	87	2	feature	feature	NOUN
ajst-905	87	3	augment	augment	NOUN
ajst-905	87	4	module	module	NOUN
ajst-905	87	5	,	,	PUNCT
ajst-905	87	6	we	we	PRON
ajst-905	87	7	fed	feed	VERB
ajst-905	87	8	the	the	DET
ajst-905	87	9	augmented	augment	VERB
ajst-905	87	10	features	feature	NOUN
ajst-905	87	11	into	into	ADP
ajst-905	87	12	the	the	DET
ajst-905	87	13	self	self	NOUN
ajst-905	87	14	-	-	PUNCT
ajst-905	87	15	attention	attention	NOUN
ajst-905	87	16	layer	layer	NOUN
ajst-905	87	17	and	and	CCONJ
ajst-905	87	18	position	position	NOUN
ajst-905	87	19	-	-	PUNCT
ajst-905	87	20	wise	wise	ADJ
ajst-905	87	21	feedforward	feedforward	NOUN
ajst-905	87	22	layer	layer	NOUN
ajst-905	87	23	,	,	PUNCT
ajst-905	87	24	and	and	CCONJ
ajst-905	87	25	its	its	PRON
ajst-905	87	26	whole	whole	ADJ
ajst-905	87	27	operation	operation	NOUN
ajst-905	87	28	is	be	AUX
ajst-905	87	29	defined	define	VERB
ajst-905	87	30	as	as	SCONJ
ajst-905	87	31	follows	follow	VERB
ajst-905	87	32	:	:	PUNCT
ajst-905	87	33	,	,	PUNCT
ajst-905	87	34	,	,	PUNCT
ajst-905	87	35	q	q	PROPN
ajst-905	87	36	k	k	X
ajst-905	87	37	vq	vq	PROPN
ajst-905	87	38	pw	pw	PROPN
ajst-905	88	1	k	k	PROPN
ajst-905	88	2	pw	pw	X
ajst-905	88	3	v	v	ADP
ajst-905	88	4	pw	pw	ADJ
ajst-905	89	1			ADV
ajst-905	89	2			NOUN
ajst-905	89	3	,	,	PUNCT
ajst-905	89	4	(	(	PUNCT
ajst-905	89	5	3	3	X
ajst-905	89	6	)	)	PUNCT
ajst-905	89	7	max	max	PROPN
ajst-905	89	8	(	(	PUNCT
ajst-905	89	9	)	)	PUNCT
ajst-905	90	1	t	t	PROPN
ajst-905	91	1	k	k	PROPN
ajst-905	92	1	qk	qk	PROPN
ajst-905	92	2	z	z	PROPN
ajst-905	92	3	soft	soft	PROPN
ajst-905	92	4	v	v	ADP
ajst-905	92	5	d	d	NOUN
ajst-905	92	6			PROPN
ajst-905	92	7	,	,	PUNCT
ajst-905	92	8	(	(	PUNCT
ajst-905	92	9	4	4	NUM
ajst-905	92	10	)	)	PUNCT
ajst-905	92	11	(	(	PUNCT
ajst-905	92	12	)	)	PUNCT
ajst-905	92	13	y	y	PROPN
ajst-905	92	14	addnorm	addnorm	PROPN
ajst-905	92	15	z	z	PROPN
ajst-905	92	16	,	,	PUNCT
ajst-905	92	17	(	(	PUNCT
ajst-905	92	18	5	5	X
ajst-905	92	19	)	)	SYM
ajst-905	92	20	1	1	NUM
ajst-905	92	21	1	1	NUM
ajst-905	92	22	2	2	NUM
ajst-905	92	23	2	2	NUM
ajst-905	92	24	(	(	PUNCT
ajst-905	92	25	(	(	PUNCT
ajst-905	92	26	)	)	PUNCT
ajst-905	92	27	)	)	PUNCT
ajst-905	92	28	tx	tx	PROPN
ajst-905	92	29	addnorm	addnorm	NOUN
ajst-905	92	30	wy	wy	PROPN
ajst-905	92	31	b	b	PROPN
ajst-905	92	32	w	w	PROPN
ajst-905	92	33	b	b	NOUN
ajst-905	92	34			PUNCT
ajst-905	92	35			PUNCT
ajst-905	92	36	,	,	PUNCT
ajst-905	92	37	(	(	PUNCT
ajst-905	92	38	6	6	NUM
ajst-905	92	39	)	)	PUNCT
ajst-905	92	40	where	where	SCONJ
ajst-905	92	41	the	the	DET
ajst-905	92	42	𝑊	𝑊	PROPN
ajst-905	92	43	,	,	PUNCT
ajst-905	92	44	𝑊	𝑊	PROPN
ajst-905	92	45	,	,	PUNCT
ajst-905	92	46	𝑊	𝑊	PROPN
ajst-905	92	47	represents	represent	VERB
ajst-905	92	48	the	the	DET
ajst-905	92	49	learnable	learnable	ADJ
ajst-905	92	50	weight	weight	NOUN
ajst-905	92	51	matrices	matrix	NOUN
ajst-905	92	52	and	and	CCONJ
ajst-905	92	53			PROPN
ajst-905	92	54	represents	represent	VERB
ajst-905	92	55	the	the	DET
ajst-905	92	56	relu	relu	NOUN
ajst-905	92	57	activation	activation	NOUN
ajst-905	92	58	function	function	NOUN
ajst-905	92	59	.	.	PUNCT
ajst-905	93	1	the	the	DET
ajst-905	93	2	𝑏	𝑏	PROPN
ajst-905	93	3	and	and	CCONJ
ajst-905	93	4	𝑏	𝑏	PROPN
ajst-905	93	5	are	be	AUX
ajst-905	93	6	the	the	DET
ajst-905	93	7	bias	bias	NOUN
ajst-905	93	8	terms	term	NOUN
ajst-905	93	9	,	,	PUNCT
ajst-905	93	10	𝑊	𝑊	PROPN
ajst-905	93	11	and	and	CCONJ
ajst-905	93	12	𝑊	𝑊	PROPN
ajst-905	93	13	are	be	AUX
ajst-905	93	14	the	the	DET
ajst-905	93	15	learnable	learnable	ADJ
ajst-905	93	16	weight	weight	NOUN
ajst-905	93	17	matrices	matrix	NOUN
ajst-905	93	18	.	.	PUNCT
ajst-905	94	1	full	full	ADJ
ajst-905	94	2	encoder	encoder	NOUN
ajst-905	94	3	.	.	PUNCT
ajst-905	95	1	in	in	ADP
ajst-905	95	2	the	the	DET
ajst-905	95	3	above	above	ADJ
ajst-905	95	4	definition	definition	NOUN
ajst-905	95	5	,	,	PUNCT
ajst-905	95	6	we	we	PRON
ajst-905	95	7	use	use	VERB
ajst-905	95	8	3	3	NUM
ajst-905	95	9	encoder	encoder	NOUN
ajst-905	95	10	layers	layer	NOUN
ajst-905	95	11	and	and	CCONJ
ajst-905	95	12	multi	multi	ADJ
ajst-905	95	13	-	-	ADJ
ajst-905	95	14	head	head	ADJ
ajst-905	95	15	attention	attention	NOUN
ajst-905	95	16	in	in	ADP
ajst-905	95	17	the	the	DET
ajst-905	95	18	specific	specific	ADJ
ajst-905	95	19	implementation	implementation	NOUN
ajst-905	95	20	.	.	PUNCT
ajst-905	96	1	through	through	ADP
ajst-905	96	2	the	the	DET
ajst-905	96	3	feature	feature	NOUN
ajst-905	96	4	augment	augment	NOUN
ajst-905	96	5	operation	operation	NOUN
ajst-905	96	6	to	to	PART
ajst-905	96	7	fuse	fuse	VERB
ajst-905	96	8	the	the	DET
ajst-905	96	9	region	region	NOUN
ajst-905	96	10	features	feature	NOUN
ajst-905	96	11	and	and	CCONJ
ajst-905	96	12	grid	grid	NOUN
ajst-905	96	13	features	feature	NOUN
ajst-905	96	14	.	.	PUNCT
ajst-905	97	1	so	so	ADV
ajst-905	97	2	that	that	SCONJ
ajst-905	97	3	the	the	DET
ajst-905	97	4	fpdt	fpdt	ADJ
ajst-905	97	5	model	model	NOUN
ajst-905	97	6	can	can	AUX
ajst-905	97	7	obtain	obtain	VERB
ajst-905	97	8	good	good	ADJ
ajst-905	97	9	fine	fine	ADV
ajst-905	97	10	-	-	PUNCT
ajst-905	97	11	grained	grain	VERB
ajst-905	97	12	details	detail	NOUN
ajst-905	97	13	and	and	CCONJ
ajst-905	97	14	sufficient	sufficient	ADJ
ajst-905	97	15	image	image	NOUN
ajst-905	97	16	information	information	NOUN
ajst-905	97	17	to	to	PART
ajst-905	97	18	ensure	ensure	VERB
ajst-905	97	19	the	the	DET
ajst-905	97	20	decoding	decode	VERB
ajst-905	97	21	effect	effect	NOUN
ajst-905	97	22	.	.	PUNCT
ajst-905	98	1	3.3	3.3	NUM
ajst-905	98	2	.	.	PUNCT
ajst-905	99	1	parallel	parallel	ADJ
ajst-905	99	2	input	input	NOUN
ajst-905	99	3	decoder	decoder	NOUN
ajst-905	99	4	figure	figure	NOUN
ajst-905	99	5	3	3	NUM
ajst-905	99	6	.	.	PUNCT
ajst-905	100	1	(	(	PUNCT
ajst-905	100	2	a	a	X
ajst-905	100	3	)	)	PUNCT
ajst-905	100	4	the	the	DET
ajst-905	100	5	structure	structure	NOUN
ajst-905	100	6	of	of	ADP
ajst-905	100	7	decoder	decoder	NOUN
ajst-905	100	8	layer	layer	NOUN
ajst-905	100	9	117	117	NUM
ajst-905	100	10	the	the	DET
ajst-905	100	11	encoder	encoder	NOUN
ajst-905	100	12	focuses	focus	VERB
ajst-905	100	13	on	on	ADP
ajst-905	100	14	understanding	understanding	NOUN
ajst-905	100	15	and	and	CCONJ
ajst-905	100	16	encoding	encode	VERB
ajst-905	100	17	the	the	DET
ajst-905	100	18	information	information	NOUN
ajst-905	100	19	of	of	ADP
ajst-905	100	20	significant	significant	ADJ
ajst-905	100	21	regions	region	NOUN
ajst-905	100	22	,	,	PUNCT
ajst-905	100	23	while	while	SCONJ
ajst-905	100	24	the	the	DET
ajst-905	100	25	decoder	decoder	NOUN
ajst-905	100	26	focuses	focus	VERB
ajst-905	100	27	on	on	ADP
ajst-905	100	28	generating	generate	VERB
ajst-905	100	29	the	the	DET
ajst-905	100	30	final	final	ADJ
ajst-905	100	31	caption	caption	NOUN
ajst-905	100	32	result	result	NOUN
ajst-905	100	33	based	base	VERB
ajst-905	100	34	on	on	ADP
ajst-905	100	35	the	the	DET
ajst-905	100	36	output	output	NOUN
ajst-905	100	37	of	of	ADP
ajst-905	100	38	the	the	DET
ajst-905	100	39	encoder	encoder	NOUN
ajst-905	100	40	.	.	PUNCT
ajst-905	101	1	in	in	ADP
ajst-905	101	2	decoding	decode	VERB
ajst-905	101	3	stage	stage	NOUN
ajst-905	101	4	,	,	PUNCT
ajst-905	101	5	the	the	DET
ajst-905	101	6	effect	effect	NOUN
ajst-905	101	7	of	of	ADP
ajst-905	101	8	final	final	ADJ
ajst-905	101	9	image	image	NOUN
ajst-905	101	10	description	description	NOUN
ajst-905	101	11	largely	largely	ADV
ajst-905	101	12	depends	depend	VERB
ajst-905	101	13	on	on	ADP
ajst-905	101	14	the	the	DET
ajst-905	101	15	information	information	NOUN
ajst-905	101	16	passed	pass	VERB
ajst-905	101	17	in	in	ADP
ajst-905	101	18	the	the	DET
ajst-905	101	19	decoding	decode	VERB
ajst-905	101	20	stage	stage	NOUN
ajst-905	101	21	.	.	PUNCT
ajst-905	102	1	in	in	ADP
ajst-905	102	2	order	order	NOUN
ajst-905	102	3	to	to	PART
ajst-905	102	4	make	make	VERB
ajst-905	102	5	the	the	DET
ajst-905	102	6	decoder	decoder	NOUN
ajst-905	102	7	decode	decode	VERB
ajst-905	102	8	better	well	ADV
ajst-905	102	9	,	,	PUNCT
ajst-905	102	10	we	we	PRON
ajst-905	102	11	try	try	VERB
ajst-905	102	12	to	to	PART
ajst-905	102	13	enhance	enhance	VERB
ajst-905	102	14	the	the	DET
ajst-905	102	15	input	input	NOUN
ajst-905	102	16	encoding	encode	VERB
ajst-905	102	17	vector	vector	NOUN
ajst-905	102	18	.	.	PUNCT
ajst-905	103	1	the	the	DET
ajst-905	103	2	structure	structure	NOUN
ajst-905	103	3	of	of	ADP
ajst-905	103	4	decoder	decoder	NOUN
ajst-905	103	5	layer	layer	NOUN
ajst-905	103	6	is	be	AUX
ajst-905	103	7	shown	show	VERB
ajst-905	103	8	in	in	ADP
ajst-905	103	9	figure	figure	NOUN
ajst-905	103	10	3	3	NUM
ajst-905	103	11	.	.	PUNCT
ajst-905	103	12	decoder	decoder	NOUN
ajst-905	103	13	layer	layer	NOUN
ajst-905	103	14	.	.	PUNCT
ajst-905	104	1	as	as	SCONJ
ajst-905	104	2	shown	show	VERB
ajst-905	104	3	in	in	ADP
ajst-905	104	4	the	the	DET
ajst-905	104	5	figure	figure	NOUN
ajst-905	104	6	3	3	NUM
ajst-905	104	7	,	,	PUNCT
ajst-905	104	8	the	the	DET
ajst-905	104	9	word	word	NOUN
ajst-905	104	10	sequence	sequence	NOUN
ajst-905	104	11	is	be	AUX
ajst-905	104	12	processed	process	VERB
ajst-905	104	13	by	by	ADP
ajst-905	104	14	word	word	NOUN
ajst-905	104	15	embedding	embed	VERB
ajst-905	104	16	and	and	CCONJ
ajst-905	104	17	the	the	DET
ajst-905	104	18	position	position	NOUN
ajst-905	104	19	encoding	encoding	NOUN
ajst-905	104	20	.	.	PUNCT
ajst-905	105	1	we	we	PRON
ajst-905	105	2	use	use	VERB
ajst-905	105	3	sinusoidal	sinusoidal	NOUN
ajst-905	105	4	positional	positional	ADJ
ajst-905	105	5	encoding	encoding	NOUN
ajst-905	105	6	[	[	X
ajst-905	105	7	7	7	NUM
ajst-905	105	8	]	]	PUNCT
ajst-905	105	9	to	to	PART
ajst-905	105	10	represents	represent	VERB
ajst-905	105	11	the	the	DET
ajst-905	105	12	words	word	NOUN
ajst-905	105	13	’	'	PUNCT
ajst-905	105	14	positions	position	NOUN
ajst-905	105	15	in	in	ADP
ajst-905	105	16	the	the	DET
ajst-905	105	17	sequence	sequence	NOUN
ajst-905	105	18	.	.	PUNCT
ajst-905	106	1	in	in	ADP
ajst-905	106	2	order	order	NOUN
ajst-905	106	3	to	to	PART
ajst-905	106	4	improve	improve	VERB
ajst-905	106	5	the	the	DET
ajst-905	106	6	final	final	ADJ
ajst-905	106	7	decoding	decode	VERB
ajst-905	106	8	effect	effect	NOUN
ajst-905	106	9	,	,	PUNCT
ajst-905	106	10	we	we	PRON
ajst-905	106	11	introduce	introduce	VERB
ajst-905	106	12	the	the	DET
ajst-905	106	13	output	output	NOUN
ajst-905	106	14	𝑋	𝑋	NOUN
ajst-905	106	15	of	of	ADP
ajst-905	106	16	different	different	ADJ
ajst-905	106	17	encoder	encoder	NOUN
ajst-905	106	18	layers	layer	NOUN
ajst-905	106	19	𝑡	𝑡	VERB
ajst-905	106	20	into	into	ADP
ajst-905	106	21	the	the	DET
ajst-905	106	22	decoder	decoder	NOUN
ajst-905	106	23	layer	layer	NOUN
ajst-905	106	24	𝑡	𝑡	PROPN
ajst-905	106	25	as	as	ADP
ajst-905	106	26	additional	additional	ADJ
ajst-905	106	27	visual	visual	ADJ
ajst-905	106	28	information	information	NOUN
ajst-905	106	29	.	.	PUNCT
ajst-905	107	1	the	the	DET
ajst-905	107	2	𝑋	𝑋	PROPN
ajst-905	107	3	is	be	AUX
ajst-905	107	4	concatenated	concatenate	VERB
ajst-905	107	5	with	with	ADP
ajst-905	107	6	the	the	DET
ajst-905	107	7	visual	visual	ADJ
ajst-905	107	8	features	feature	NOUN
ajst-905	107	9	𝑉	𝑉	PROPN
ajst-905	107	10	as	as	ADP
ajst-905	107	11	a	a	DET
ajst-905	107	12	parallel	parallel	ADJ
ajst-905	107	13	input	input	NOUN
ajst-905	107	14	to	to	ADP
ajst-905	107	15	the	the	DET
ajst-905	107	16	decoder	decoder	NOUN
ajst-905	107	17	layer	layer	NOUN
ajst-905	107	18	.	.	PUNCT
ajst-905	108	1	the	the	DET
ajst-905	108	2	operators	operator	NOUN
ajst-905	108	3	of	of	ADP
ajst-905	108	4	parallel	parallel	ADJ
ajst-905	108	5	input	input	NOUN
ajst-905	108	6	cross	cross	NOUN
ajst-905	108	7	attention	attention	NOUN
ajst-905	108	8	are	be	AUX
ajst-905	108	9	defined	define	VERB
ajst-905	108	10	as	as	SCONJ
ajst-905	108	11	follows	follow	VERB
ajst-905	108	12	:	:	PUNCT
ajst-905	108	13			ADJ
ajst-905	108	14			PROPN
ajst-905	108	15	,	,	PUNCT
ajst-905	108	16	tz	tz	PROPN
ajst-905	108	17	v	v	NOUN
ajst-905	108	18	x	x	PUNCT
ajst-905	108	19	(	(	PUNCT
ajst-905	108	20	7	7	X
ajst-905	108	21	)	)	PUNCT
ajst-905	108	22			NOUN
ajst-905	108	23			PROPN
ajst-905	108	24	,	,	PUNCT
ajst-905	108	25	,	,	PUNCT
ajst-905	108	26	q	q	PROPN
ajst-905	108	27	k	k	NOUN
ajst-905	108	28	vs	vs	ADP
ajst-905	108	29	attention	attention	NOUN
ajst-905	108	30	w	w	PROPN
ajst-905	108	31	y	y	PROPN
ajst-905	108	32	w	w	PROPN
ajst-905	108	33	z	z	PROPN
ajst-905	108	34	w	w	PROPN
ajst-905	108	35	z	z	PROPN
ajst-905	108	36	(	(	PUNCT
ajst-905	108	37	8)	8)	NUM
ajst-905	108	38	where	where	SCONJ
ajst-905	108	39	the	the	PRON
ajst-905	108	40	[	[	PUNCT
ajst-905	108	41	,	,	PUNCT
ajst-905	108	42	]	]	PUNCT
ajst-905	108	43	in	in	ADP
ajst-905	108	44	eq	eq	ADP
ajst-905	108	45	.	.	PUNCT
ajst-905	109	1	(	(	PUNCT
ajst-905	109	2	7	7	X
ajst-905	109	3	)	)	PUNCT
ajst-905	109	4	represents	represent	VERB
ajst-905	109	5	concatenation	concatenation	NOUN
ajst-905	109	6	operation	operation	NOUN
ajst-905	109	7	.	.	PUNCT
ajst-905	110	1	in	in	ADP
ajst-905	110	2	eq	eq	ADP
ajst-905	110	3	.	.	PUNCT
ajst-905	110	4	(	(	PUNCT
ajst-905	110	5	8)	8)	NUM
ajst-905	110	6	,	,	PUNCT
ajst-905	110	7	𝑊	𝑊	PROPN
ajst-905	110	8	and	and	CCONJ
ajst-905	110	9	𝑊	𝑊	PROPN
ajst-905	110	10	𝑅	𝑅	ADJ
ajst-905	110	11			NOUN
ajst-905	110	12	are	be	AUX
ajst-905	110	13	learnable	learnable	ADJ
ajst-905	110	14	weight	weight	NOUN
ajst-905	110	15	matrices	matrix	NOUN
ajst-905	110	16	.	.	PUNCT
ajst-905	111	1	full	full	ADJ
ajst-905	111	2	decoder	decoder	NOUN
ajst-905	111	3	.	.	PUNCT
ajst-905	112	1	in	in	ADP
ajst-905	112	2	the	the	DET
ajst-905	112	3	above	above	ADJ
ajst-905	112	4	definition	definition	NOUN
ajst-905	112	5	,	,	PUNCT
ajst-905	112	6	we	we	PRON
ajst-905	112	7	use	use	VERB
ajst-905	112	8	3	3	NUM
ajst-905	112	9	decoder	decoder	NOUN
ajst-905	112	10	layers	layer	NOUN
ajst-905	112	11	which	which	PRON
ajst-905	112	12	are	be	AUX
ajst-905	112	13	stacked	stack	VERB
ajst-905	112	14	in	in	ADP
ajst-905	112	15	sequence	sequence	NOUN
ajst-905	112	16	in	in	ADP
ajst-905	112	17	the	the	DET
ajst-905	112	18	specific	specific	ADJ
ajst-905	112	19	implementation	implementation	NOUN
ajst-905	112	20	.	.	PUNCT
ajst-905	113	1	3.4	3.4	NUM
ajst-905	113	2	.	.	PUNCT
ajst-905	113	3	training	training	NOUN
ajst-905	113	4	details	detail	NOUN
ajst-905	113	5	refer	refer	VERB
ajst-905	113	6	to	to	ADP
ajst-905	113	7	the	the	DET
ajst-905	113	8	training	training	NOUN
ajst-905	113	9	methods	method	NOUN
ajst-905	113	10	commonly	commonly	ADV
ajst-905	113	11	used	use	VERB
ajst-905	113	12	in	in	ADP
ajst-905	113	13	image	image	NOUN
ajst-905	113	14	caption	caption	NOUN
ajst-905	113	15	tasks	task	NOUN
ajst-905	113	16	,	,	PUNCT
ajst-905	113	17	we	we	PRON
ajst-905	113	18	optimize	optimize	VERB
ajst-905	113	19	the	the	DET
ajst-905	113	20	fpdt	fpdt	ADJ
ajst-905	113	21	model	model	NOUN
ajst-905	113	22	with	with	ADP
ajst-905	113	23	the	the	DET
ajst-905	113	24	crossentropy	crossentropy	NOUN
ajst-905	113	25	loss	loss	NOUN
ajst-905	113	26	(	(	PUNCT
ajst-905	113	27	xe	xe	PROPN
ajst-905	113	28	):	):	PUNCT
ajst-905	113	29			PROPN
ajst-905	113	30			SYM
ajst-905	113	31			NOUN
ajst-905	113	32			X
ajst-905	114	1			PROPN
ajst-905	114	2	*	*	PUNCT
ajst-905	114	3	*	*	PUNCT
ajst-905	114	4	1	1	NUM
ajst-905	114	5	:	:	SYM
ajst-905	114	6	1	1	NUM
ajst-905	114	7	1	1	NUM
ajst-905	114	8	log	log	NOUN
ajst-905	114	9	|	|	ADV
ajst-905	114	10	t	t	NOUN
ajst-905	114	11	t	t	NOUN
ajst-905	114	12	txe	txe	NOUN
ajst-905	114	13	t	t	PROPN
ajst-905	114	14	l	l	NOUN
ajst-905	115	1	p	p	X
ajst-905	115	2	y	y	PROPN
ajst-905	115	3	y	y	NUM
ajst-905	115	4			PROPN
ajst-905	115	5			NUM
ajst-905	115	6			NUM
ajst-905	116	1			X
ajst-905	116	2	(	(	PUNCT
ajst-905	116	3	9	9	NUM
ajst-905	116	4	)	)	PUNCT
ajst-905	116	5	where	where	SCONJ
ajst-905	116	6	the	the	DET
ajst-905	116	7	𝑦	𝑦	NOUN
ajst-905	116	8	:	:	PUNCT
ajst-905	116	9	∗	∗	NOUN
ajst-905	116	10	is	be	AUX
ajst-905	116	11	ground	ground	NOUN
ajst-905	116	12	truth	truth	NOUN
ajst-905	116	13	sequence	sequence	NOUN
ajst-905	116	14	and	and	CCONJ
ajst-905	116	15	the	the	DET
ajst-905	116	16			NOUN
ajst-905	116	17	is	be	AUX
ajst-905	116	18	the	the	DET
ajst-905	116	19	parameters	parameter	NOUN
ajst-905	116	20	of	of	ADP
ajst-905	116	21	our	our	PRON
ajst-905	116	22	model	model	NOUN
ajst-905	116	23	which	which	PRON
ajst-905	116	24	needs	need	VERB
ajst-905	116	25	to	to	PART
ajst-905	116	26	be	be	AUX
ajst-905	116	27	optimized	optimize	VERB
ajst-905	116	28	.	.	PUNCT
ajst-905	117	1	after	after	ADP
ajst-905	117	2	that	that	PRON
ajst-905	117	3	,	,	PUNCT
ajst-905	117	4	we	we	PRON
ajst-905	117	5	continually	continually	ADV
ajst-905	117	6	optimize	optimize	VERB
ajst-905	117	7	the	the	DET
ajst-905	117	8	non	non	ADJ
ajst-905	117	9	-	-	ADJ
ajst-905	117	10	differentiable	differentiable	ADJ
ajst-905	117	11	cider	cider	NOUN
ajst-905	117	12	-	-	PUNCT
ajst-905	117	13	d	d	NOUN
ajst-905	117	14	score	score	NOUN
ajst-905	117	15	by	by	ADP
ajst-905	117	16	self	self	NOUN
ajst-905	117	17	-	-	PUNCT
ajst-905	117	18	critical	critical	ADJ
ajst-905	117	19	sequence	sequence	NOUN
ajst-905	117	20	training	training	NOUN
ajst-905	117	21	(	(	PUNCT
ajst-905	117	22	scst	scst	NOUN
ajst-905	117	23	):	):	PUNCT
ajst-905	117	24			NOUN
ajst-905	117	25			SYM
ajst-905	117	26			NOUN
ajst-905	118	1			NOUN
ajst-905	119	1			PROPN
ajst-905	120	1			PROPN
ajst-905	121	1	1	1	NOUN
ajst-905	121	2	:	:	PUNCT
ajst-905	121	3	1	1	NUM
ajst-905	121	4	:	:	SYM
ajst-905	121	5	1	1	NUM
ajst-905	121	6	1	1	NUM
ajst-905	121	7	log	log	NOUN
ajst-905	121	8	k	k	PROPN
ajst-905	122	1	i	i	PRON
ajst-905	122	2	i	i	PRON
ajst-905	123	1	t	t	X
ajst-905	123	2	trl	trl	PROPN
ajst-905	124	1	i	i	X
ajst-905	124	2	l	l	NOUN
ajst-905	124	3	r	r	NOUN
ajst-905	124	4	y	y	PROPN
ajst-905	124	5	b	b	PROPN
ajst-905	124	6	p	p	X
ajst-905	124	7	y	y	PROPN
ajst-905	124	8	k	k	NOUN
ajst-905	124	9			PROPN
ajst-905	125	1			PROPN
ajst-905	126	1			ADJ
ajst-905	126	2			ADJ
ajst-905	126	3			PROPN
ajst-905	126	4			PROPN
ajst-905	126	5			X
ajst-905	126	6			PROPN
ajst-905	126	7			NOUN
ajst-905	126	8	(	(	PUNCT
ajst-905	126	9	10	10	NUM
ajst-905	126	10	)	)	PUNCT
ajst-905	126	11	where	where	SCONJ
ajst-905	126	12	the	the	DET
ajst-905	126	13	k	k	PROPN
ajst-905	126	14	is	be	AUX
ajst-905	126	15	the	the	DET
ajst-905	126	16	beam	beam	NOUN
ajst-905	126	17	size	size	NOUN
ajst-905	126	18	,	,	PUNCT
ajst-905	126	19	𝑦	𝑦	NOUN
ajst-905	126	20	:	:	PUNCT
ajst-905	126	21	is	be	AUX
ajst-905	126	22	the	the	DET
ajst-905	126	23	𝑖-th	𝑖-th	PROPN
ajst-905	126	24	sentence	sentence	NOUN
ajst-905	126	25	in	in	ADP
ajst-905	126	26	the	the	DET
ajst-905	126	27	beam	beam	NOUN
ajst-905	126	28	.	.	PUNCT
ajst-905	127	1	𝑟	𝑟	X
ajst-905	127	2	(	(	PUNCT
ajst-905	127	3	)	)	PUNCT
ajst-905	127	4	is	be	AUX
ajst-905	127	5	the	the	DET
ajst-905	127	6	cider	cider	NOUN
ajst-905	127	7	-	-	PUNCT
ajst-905	127	8	d	d	NOUN
ajst-905	127	9	function	function	NOUN
ajst-905	127	10	,	,	PUNCT
ajst-905	127	11	and	and	CCONJ
ajst-905	127	12	𝑏	𝑏	NOUN
ajst-905	127	13			X
ajst-905	127	14	𝑟	𝑟	X
ajst-905	127	15	𝑦	𝑦	NOUN
ajst-905	127	16	:	:	PUNCT
ajst-905	127	17	/𝑘	/𝑘	PUNCT
ajst-905	127	18	is	be	AUX
ajst-905	127	19	the	the	DET
ajst-905	127	20	baseline	baseline	NOUN
ajst-905	127	21	.	.	PUNCT
ajst-905	128	1	4	4	X
ajst-905	128	2	.	.	X
ajst-905	128	3	experiment	experiment	NOUN
ajst-905	128	4	and	and	CCONJ
ajst-905	128	5	analysis	analysis	NOUN
ajst-905	128	6	4.1	4.1	NUM
ajst-905	128	7	.	.	PUNCT
ajst-905	128	8	dataset	dataset	VERB
ajst-905	128	9	the	the	DET
ajst-905	128	10	data	datum	NOUN
ajst-905	128	11	set	set	VERB
ajst-905	128	12	used	use	VERB
ajst-905	128	13	in	in	ADP
ajst-905	128	14	our	our	PRON
ajst-905	128	15	experiment	experiment	NOUN
ajst-905	128	16	is	be	AUX
ajst-905	128	17	mscoco	mscoco	ADJ
ajst-905	128	18	.	.	PUNCT
ajst-905	129	1	mscoco	mscoco	PROPN
ajst-905	129	2	is	be	AUX
ajst-905	129	3	a	a	DET
ajst-905	129	4	dataset	dataset	NOUN
ajst-905	129	5	provided	provide	VERB
ajst-905	129	6	by	by	ADP
ajst-905	129	7	the	the	DET
ajst-905	129	8	microsoft	microsoft	PROPN
ajst-905	129	9	team	team	NOUN
ajst-905	129	10	.	.	PUNCT
ajst-905	130	1	it	it	PRON
ajst-905	130	2	contains	contain	VERB
ajst-905	130	3	more	more	ADJ
ajst-905	130	4	than	than	ADP
ajst-905	130	5	120,000	120,000	NUM
ajst-905	130	6	images	image	NOUN
ajst-905	130	7	.	.	PUNCT
ajst-905	131	1	each	each	DET
ajst-905	131	2	image	image	NOUN
ajst-905	131	3	has	have	VERB
ajst-905	131	4	5	5	NUM
ajst-905	131	5	different	different	ADJ
ajst-905	131	6	annotations	annotation	NOUN
ajst-905	131	7	.	.	PUNCT
ajst-905	132	1	in	in	ADP
ajst-905	132	2	addition	addition	NOUN
ajst-905	132	3	to	to	ADP
ajst-905	132	4	this	this	PRON
ajst-905	132	5	,	,	PUNCT
ajst-905	132	6	karpathy	karpathy	PROPN
ajst-905	132	7	et	et	PROPN
ajst-905	132	8	al	al	PROPN
ajst-905	132	9	.	.	PROPN
ajst-905	132	10	provide	provide	VERB
ajst-905	132	11	another	another	DET
ajst-905	132	12	split	split	NOUN
ajst-905	132	13	where	where	SCONJ
ajst-905	132	14	5,000	5,000	NUM
ajst-905	132	15	images	image	NOUN
ajst-905	132	16	are	be	AUX
ajst-905	132	17	used	use	VERB
ajst-905	132	18	for	for	ADP
ajst-905	132	19	evaluation	evaluation	NOUN
ajst-905	132	20	,	,	PUNCT
ajst-905	132	21	5,000	5,000	NUM
ajst-905	132	22	for	for	ADP
ajst-905	132	23	testing	testing	NOUN
ajst-905	132	24	and	and	CCONJ
ajst-905	132	25	the	the	DET
ajst-905	132	26	rest	rest	NOUN
ajst-905	132	27	for	for	ADP
ajst-905	132	28	training	training	NOUN
ajst-905	132	29	.	.	PUNCT
ajst-905	133	1	4.2	4.2	NUM
ajst-905	133	2	.	.	PUNCT
ajst-905	134	1	experimental	experimental	ADJ
ajst-905	134	2	settings	setting	NOUN
ajst-905	134	3	evaluation	evaluation	NOUN
ajst-905	134	4	metrics	metric	NOUN
ajst-905	134	5	.	.	PUNCT
ajst-905	135	1	following	follow	VERB
ajst-905	135	2	the	the	DET
ajst-905	135	3	standard	standard	ADJ
ajst-905	135	4	evaluation	evaluation	NOUN
ajst-905	135	5	protocol	protocol	NOUN
ajst-905	135	6	,	,	PUNCT
ajst-905	135	7	we	we	PRON
ajst-905	135	8	use	use	VERB
ajst-905	135	9	the	the	DET
ajst-905	135	10	full	full	ADJ
ajst-905	135	11	set	set	NOUN
ajst-905	135	12	of	of	ADP
ajst-905	135	13	captioning	caption	VERB
ajst-905	135	14	metrics	metric	NOUN
ajst-905	135	15	to	to	PART
ajst-905	135	16	evaluate	evaluate	VERB
ajst-905	135	17	the	the	DET
ajst-905	135	18	quality	quality	NOUN
ajst-905	135	19	of	of	ADP
ajst-905	135	20	our	our	PRON
ajst-905	135	21	ieat	ieat	NOUN
ajst-905	135	22	model	model	NOUN
ajst-905	135	23	,	,	PUNCT
ajst-905	135	24	including	include	VERB
ajst-905	135	25	bleu	bleu	NOUN
ajst-905	135	26	[	[	X
ajst-905	135	27	22	22	NUM
ajst-905	135	28	]	]	PUNCT
ajst-905	135	29	,	,	PUNCT
ajst-905	135	30	meteor	meteor	NOUN
ajst-905	136	1	[	[	X
ajst-905	136	2	33	33	NUM
ajst-905	136	3	]	]	PUNCT
ajst-905	136	4	,	,	PUNCT
ajst-905	136	5	rouge	rouge	NOUN
ajst-905	136	6	[	[	X
ajst-905	136	7	34	34	NUM
ajst-905	136	8	]	]	PUNCT
ajst-905	136	9	,	,	PUNCT
ajst-905	136	10	cider	cider	NOUN
ajst-905	136	11	[	[	NOUN
ajst-905	136	12	35	35	NUM
ajst-905	136	13	]	]	PUNCT
ajst-905	136	14	,	,	PUNCT
ajst-905	136	15	and	and	CCONJ
ajst-905	136	16	spice	spice	NOUN
ajst-905	137	1	[	[	X
ajst-905	137	2	36	36	NUM
ajst-905	137	3	]	]	PUNCT
ajst-905	137	4	.	.	PUNCT
ajst-905	138	1	implementation	implementation	NOUN
ajst-905	138	2	details	detail	NOUN
ajst-905	138	3	.	.	PUNCT
ajst-905	139	1	in	in	ADP
ajst-905	139	2	terms	term	NOUN
ajst-905	139	3	of	of	ADP
ajst-905	139	4	the	the	DET
ajst-905	139	5	extraction	extraction	NOUN
ajst-905	139	6	of	of	ADP
ajst-905	139	7	visual	visual	ADJ
ajst-905	139	8	features	feature	NOUN
ajst-905	139	9	,	,	PUNCT
ajst-905	139	10	we	we	PRON
ajst-905	139	11	utilize	utilize	VERB
ajst-905	139	12	the	the	DET
ajst-905	139	13	region	region	NOUN
ajst-905	139	14	features	feature	NOUN
ajst-905	139	15	and	and	CCONJ
ajst-905	139	16	the	the	DET
ajst-905	139	17	grid	grid	NOUN
ajst-905	139	18	features	feature	NOUN
ajst-905	139	19	provided	provide	VERB
ajst-905	139	20	by	by	ADP
ajst-905	139	21	jiang	jiang	PROPN
ajst-905	139	22	et	et	PROPN
ajst-905	139	23	al	al	PROPN
ajst-905	139	24	.	.	PUNCT
ajst-905	140	1	[	[	X
ajst-905	140	2	9	9	NUM
ajst-905	140	3	]	]	PUNCT
ajst-905	140	4	where	where	SCONJ
ajst-905	140	5	the	the	DET
ajst-905	140	6	grid	grid	NOUN
ajst-905	140	7	size	size	NOUN
ajst-905	140	8	is	be	AUX
ajst-905	140	9	7	7	NUM
ajst-905	140	10	×	×	NOUN
ajst-905	140	11	7	7	NUM
ajst-905	140	12	and	and	CCONJ
ajst-905	140	13	the	the	DET
ajst-905	140	14	dimension	dimension	NOUN
ajst-905	140	15	of	of	ADP
ajst-905	140	16	image	image	NOUN
ajst-905	140	17	features	feature	NOUN
ajst-905	140	18	is	be	AUX
ajst-905	140	19	2048	2048	NUM
ajst-905	140	20	.	.	PUNCT
ajst-905	141	1	in	in	ADP
ajst-905	141	2	our	our	PRON
ajst-905	141	3	implementation	implementation	NOUN
ajst-905	141	4	,	,	PUNCT
ajst-905	141	5	we	we	PRON
ajst-905	141	6	set	set	VERB
ajst-905	141	7	the	the	DET
ajst-905	141	8	number	number	NOUN
ajst-905	141	9	of	of	ADP
ajst-905	141	10	heads	head	NOUN
ajst-905	141	11	equal	equal	ADJ
ajst-905	141	12	to	to	ADP
ajst-905	141	13	8	8	NUM
ajst-905	141	14	,	,	PUNCT
ajst-905	141	15	the	the	DET
ajst-905	141	16	batch	batch	NOUN
ajst-905	141	17	size	size	NOUN
ajst-905	141	18	equal	equal	ADJ
ajst-905	141	19	to	to	ADP
ajst-905	141	20	50	50	NUM
ajst-905	141	21	and	and	CCONJ
ajst-905	141	22	the	the	DET
ajst-905	141	23	number	number	NOUN
ajst-905	141	24	of	of	ADP
ajst-905	141	25	memory	memory	NOUN
ajst-905	141	26	slots	slot	NOUN
ajst-905	141	27	is	be	AUX
ajst-905	141	28	40	40	NUM
ajst-905	141	29	.	.	PUNCT
ajst-905	142	1	the	the	DET
ajst-905	142	2	𝑑	𝑑	PROPN
ajst-905	142	3	of	of	ADP
ajst-905	142	4	each	each	DET
ajst-905	142	5	layer	layer	NOUN
ajst-905	142	6	is	be	AUX
ajst-905	142	7	512	512	NUM
ajst-905	142	8	and	and	CCONJ
ajst-905	142	9	the	the	DET
ajst-905	142	10	inner	inner	ADJ
ajst-905	142	11	dimension	dimension	NOUN
ajst-905	142	12	of	of	ADP
ajst-905	142	13	ffn	ffn	PROPN
ajst-905	142	14	is	be	AUX
ajst-905	142	15	2048	2048	NUM
ajst-905	142	16	.	.	PUNCT
ajst-905	143	1	the	the	DET
ajst-905	143	2	threshold	threshold	NOUN
ajst-905	143	3	value	value	NOUN
ajst-905	143	4	𝑡	𝑡	PROPN
ajst-905	143	5	is	be	AUX
ajst-905	143	6	set	set	VERB
ajst-905	143	7	to	to	ADP
ajst-905	143	8	0.6	0.6	NUM
ajst-905	143	9	.	.	PUNCT
ajst-905	144	1	the	the	DET
ajst-905	144	2	dropout	dropout	NOUN
ajst-905	144	3	probability	probability	NOUN
ajst-905	144	4	is	be	AUX
ajst-905	144	5	set	set	VERB
ajst-905	144	6	to	to	ADP
ajst-905	144	7	0.1	0.1	NUM
ajst-905	144	8	and	and	CCONJ
ajst-905	144	9	we	we	PRON
ajst-905	144	10	set	set	VERB
ajst-905	144	11	the	the	DET
ajst-905	144	12	beam	beam	NOUN
ajst-905	144	13	size	size	NOUN
ajst-905	144	14	to	to	ADP
ajst-905	144	15	5	5	NUM
ajst-905	144	16	.	.	PUNCT
ajst-905	145	1	we	we	PRON
ajst-905	145	2	adopt	adopt	VERB
ajst-905	145	3	adam	adam	NOUN
ajst-905	145	4	optimizer	optimizer	NOUN
ajst-905	145	5	to	to	PART
ajst-905	145	6	train	train	VERB
ajst-905	145	7	our	our	PRON
ajst-905	145	8	model	model	NOUN
ajst-905	145	9	.	.	PUNCT
ajst-905	146	1	in	in	ADP
ajst-905	146	2	the	the	DET
ajst-905	146	3	xe	xe	PROPN
ajst-905	146	4	pre	pre	ADJ
ajst-905	146	5	-	-	ADJ
ajst-905	146	6	training	training	ADJ
ajst-905	146	7	stage	stage	NOUN
ajst-905	146	8	,	,	PUNCT
ajst-905	146	9	our	our	PRON
ajst-905	146	10	learning	learning	NOUN
ajst-905	146	11	rate	rate	NOUN
ajst-905	146	12	strategy	strategy	NOUN
ajst-905	146	13	is	be	AUX
ajst-905	146	14	defined	define	VERB
ajst-905	146	15	as	as	SCONJ
ajst-905	146	16	follows	follow	VERB
ajst-905	146	17	:	:	PUNCT
ajst-905	147	1	_	_	PUNCT
ajst-905	147	2	*	*	PUNCT
ajst-905	147	3	/	/	SYM
ajst-905	147	4	3	3	NUM
ajst-905	147	5	,	,	PUNCT
ajst-905	147	6	3	3	NUM
ajst-905	147	7	_	_	NUM
ajst-905	147	8	,	,	PUNCT
ajst-905	147	9	4	4	NUM
ajst-905	147	10	10	10	NUM
ajst-905	147	11	_	_	PUNCT
ajst-905	148	1	_	_	PUNCT
ajst-905	149	1	*	*	PUNCT
ajst-905	149	2	0.2	0.2	NUM
ajst-905	149	3	,	,	PUNCT
ajst-905	149	4	11	11	NUM
ajst-905	149	5	12	12	NUM
ajst-905	149	6	_	_	PUNCT
ajst-905	150	1	*	*	PUNCT
ajst-905	150	2	0.2	0.2	NUM
ajst-905	150	3	*	*	SYM
ajst-905	150	4	0.2	0.2	NUM
ajst-905	150	5	,	,	PUNCT
ajst-905	150	6	13	13	NUM
ajst-905	150	7	base	base	NOUN
ajst-905	150	8	lr	lr	X
ajst-905	150	9	e	e	NOUN
ajst-905	150	10	e	e	NOUN
ajst-905	150	11	base	base	NOUN
ajst-905	150	12	lr	lr	X
ajst-905	150	13	e	e	PROPN
ajst-905	150	14	lambda	lambda	NOUN
ajst-905	150	15	lr	lr	PROPN
ajst-905	150	16	base	base	NOUN
ajst-905	150	17	lr	lr	X
ajst-905	150	18	e	e	PROPN
ajst-905	150	19	base	base	NOUN
ajst-905	150	20	lr	lr	X
ajst-905	150	21	e	e	NOUN
ajst-905	150	22			NOUN
ajst-905	150	23			PRON
ajst-905	150	24			NOUN
ajst-905	150	25			ADV
ajst-905	150	26			NUM
ajst-905	150	27			NOUN
ajst-905	150	28			ADV
ajst-905	150	29			NUM
ajst-905	150	30			NOUN
ajst-905	150	31	(	(	PUNCT
ajst-905	150	32	11	11	NUM
ajst-905	150	33	)	)	PUNCT
ajst-905	150	34	where	where	SCONJ
ajst-905	150	35	the	the	DET
ajst-905	150	36	𝑏𝑎𝑠𝑒_𝑙𝑟	𝑏𝑎𝑠𝑒_𝑙𝑟	NOUN
ajst-905	150	37	is	be	AUX
ajst-905	150	38	set	set	VERB
ajst-905	150	39	to	to	ADP
ajst-905	150	40	110	110	NUM
ajst-905	150	41	and	and	CCONJ
ajst-905	150	42	the	the	DET
ajst-905	150	43	𝑒	𝑒	PROPN
ajst-905	150	44	represents	represent	VERB
ajst-905	150	45	the	the	DET
ajst-905	150	46	current	current	ADJ
ajst-905	150	47	epoch	epoch	NOUN
ajst-905	150	48	number	number	NOUN
ajst-905	150	49	.	.	PUNCT
ajst-905	151	1	during	during	ADP
ajst-905	151	2	the	the	DET
ajst-905	151	3	xe	xe	PROPN
ajst-905	151	4	pre	pre	ADJ
ajst-905	151	5	-	-	ADJ
ajst-905	151	6	training	training	ADJ
ajst-905	151	7	stage	stage	NOUN
ajst-905	151	8	,	,	PUNCT
ajst-905	151	9	if	if	SCONJ
ajst-905	151	10	the	the	DET
ajst-905	151	11	cider	cider	NOUN
ajst-905	151	12	value	value	NOUN
ajst-905	151	13	drops	drop	VERB
ajst-905	151	14	for	for	ADP
ajst-905	151	15	5	5	NUM
ajst-905	151	16	consecutive	consecutive	ADJ
ajst-905	151	17	epochs	epoch	NOUN
ajst-905	151	18	,	,	PUNCT
ajst-905	151	19	we	we	PRON
ajst-905	151	20	will	will	AUX
ajst-905	151	21	switch	switch	VERB
ajst-905	151	22	it	it	PRON
ajst-905	151	23	to	to	ADP
ajst-905	151	24	self	self	NOUN
ajst-905	151	25	-	-	PUNCT
ajst-905	151	26	critical	critical	ADJ
ajst-905	151	27	sequence	sequence	NOUN
ajst-905	151	28	training	training	NOUN
ajst-905	151	29	and	and	CCONJ
ajst-905	151	30	the	the	DET
ajst-905	151	31	learning	learning	NOUN
ajst-905	151	32	rate	rate	NOUN
ajst-905	151	33	is	be	AUX
ajst-905	151	34	set	set	VERB
ajst-905	151	35	to	to	ADP
ajst-905	151	36	a	a	DET
ajst-905	151	37	fixed	fix	VERB
ajst-905	151	38	value	value	NOUN
ajst-905	151	39	of	of	ADP
ajst-905	151	40	510	510	NUM
ajst-905	151	41	.	.	PUNCT
ajst-905	152	1	the	the	DET
ajst-905	152	2	training	training	NOUN
ajst-905	152	3	process	process	NOUN
ajst-905	152	4	will	will	AUX
ajst-905	152	5	stop	stop	VERB
ajst-905	152	6	when	when	SCONJ
ajst-905	152	7	the	the	DET
ajst-905	152	8	cider	cider	NOUN
ajst-905	152	9	value	value	NOUN
ajst-905	152	10	drops	drop	VERB
ajst-905	152	11	for	for	ADP
ajst-905	152	12	5	5	NUM
ajst-905	152	13	consecutive	consecutive	ADJ
ajst-905	152	14	epochs	epoch	NOUN
ajst-905	152	15	in	in	ADP
ajst-905	152	16	self	self	NOUN
ajst-905	152	17	-	-	PUNCT
ajst-905	152	18	critical	critical	ADJ
ajst-905	152	19	sequence	sequence	NOUN
ajst-905	152	20	training	training	NOUN
ajst-905	152	21	.	.	PUNCT
ajst-905	153	1	4.3	4.3	NUM
ajst-905	153	2	.	.	PUNCT
ajst-905	153	3	comparisons	comparison	NOUN
ajst-905	153	4	with	with	ADP
ajst-905	153	5	state	state	NOUN
ajst-905	153	6	-	-	PUNCT
ajst-905	153	7	of	of	ADP
ajst-905	153	8	-	-	PUNCT
ajst-905	153	9	the	the	DET
ajst-905	153	10	-	-	PUNCT
ajst-905	153	11	art	art	NOUN
ajst-905	153	12	methods	method	NOUN
ajst-905	153	13	offline	offline	ADJ
ajst-905	153	14	evaluation	evaluation	NOUN
ajst-905	153	15	.	.	PUNCT
ajst-905	154	1	we	we	PRON
ajst-905	154	2	report	report	VERB
ajst-905	154	3	the	the	DET
ajst-905	154	4	performance	performance	NOUN
ajst-905	154	5	of	of	ADP
ajst-905	154	6	the	the	DET
ajst-905	154	7	fpdt	fpdt	ADJ
ajst-905	154	8	model	model	NOUN
ajst-905	154	9	with	with	ADP
ajst-905	154	10	models	model	NOUN
ajst-905	154	11	that	that	PRON
ajst-905	154	12	have	have	AUX
ajst-905	154	13	achieved	achieve	VERB
ajst-905	154	14	cutting	cutting	NOUN
ajst-905	154	15	-	-	PUNCT
ajst-905	154	16	edge	edge	NOUN
ajst-905	154	17	results	result	NOUN
ajst-905	154	18	in	in	ADP
ajst-905	154	19	recent	recent	ADJ
ajst-905	154	20	years	year	NOUN
ajst-905	154	21	in	in	ADP
ajst-905	154	22	the	the	DET
ajst-905	154	23	same	same	ADJ
ajst-905	154	24	experimental	experimental	ADJ
ajst-905	154	25	settings	setting	NOUN
ajst-905	154	26	,	,	PUNCT
ajst-905	154	27	including	include	VERB
ajst-905	154	28	scst	scst	NOUN
ajst-905	154	29	[	[	X
ajst-905	154	30	16	16	NUM
ajst-905	154	31	]	]	PUNCT
ajst-905	154	32	,	,	PUNCT
ajst-905	154	33	up	up	ADP
ajst-905	154	34	-	-	PUNCT
ajst-905	154	35	down	down	ADP
ajst-905	154	36	[	[	X
ajst-905	154	37	1	1	NUM
ajst-905	154	38	]	]	PUNCT
ajst-905	154	39	,	,	PUNCT
ajst-905	154	40	rfnet	rfnet	NOUN
ajst-905	155	1	[	[	X
ajst-905	155	2	17	17	NUM
ajst-905	155	3	]	]	PUNCT
ajst-905	155	4	,	,	PUNCT
ajst-905	155	5	gcn	gcn	NOUN
ajst-905	155	6	-	-	NOUN
ajst-905	155	7	lstm	lstm	NOUN
ajst-905	155	8	[	[	X
ajst-905	155	9	18	18	NUM
ajst-905	155	10	]	]	PUNCT
ajst-905	155	11	,	,	PUNCT
ajst-905	155	12	sgae	sgae	NOUN
ajst-905	156	1	[	[	X
ajst-905	156	2	19	19	NUM
ajst-905	156	3	]	]	PUNCT
ajst-905	156	4	,	,	PUNCT
ajst-905	156	5	ort	ort	NOUN
ajst-905	156	6	[	[	X
ajst-905	156	7	20	20	NUM
ajst-905	156	8	]	]	PUNCT
ajst-905	156	9	,	,	PUNCT
ajst-905	156	10	aoanet	aoanet	NOUN
ajst-905	156	11	[	[	X
ajst-905	156	12	21	21	NUM
ajst-905	156	13	]	]	PUNCT
ajst-905	156	14	,	,	PUNCT
ajst-905	156	15	𝑀	𝑀	PROPN
ajst-905	156	16	transformer	transformer	NOUN
ajst-905	156	17	[	[	X
ajst-905	156	18	23	23	NUM
ajst-905	156	19	]	]	PUNCT
ajst-905	156	20	,	,	PUNCT
ajst-905	156	21	x	x	NOUN
ajst-905	156	22	-	-	NOUN
ajst-905	156	23	transformer	transformer	NOUN
ajst-905	156	24	[	[	X
ajst-905	156	25	22	22	NUM
ajst-905	156	26	]	]	PUNCT
ajst-905	156	27	.	.	PUNCT
ajst-905	157	1	the	the	DET
ajst-905	157	2	results	result	NOUN
ajst-905	157	3	are	be	AUX
ajst-905	157	4	shown	show	VERB
ajst-905	157	5	in	in	ADP
ajst-905	157	6	table	table	NOUN
ajst-905	157	7	2	2	NUM
ajst-905	157	8	.	.	SYM
ajst-905	157	9	118	118	NUM
ajst-905	157	10	table	table	NOUN
ajst-905	157	11	1	1	NUM
ajst-905	157	12	.	.	PUNCT
ajst-905	157	13	comparison	comparison	NOUN
ajst-905	157	14	with	with	ADP
ajst-905	157	15	the	the	DET
ajst-905	157	16	state	state	NOUN
ajst-905	157	17	of	of	ADP
ajst-905	157	18	art	art	NOUN
ajst-905	157	19	models	model	NOUN
ajst-905	157	20	published	publish	VERB
ajst-905	157	21	on	on	ADP
ajst-905	157	22	the	the	DET
ajst-905	157	23	mscoco	mscoco	ADJ
ajst-905	157	24	online	online	ADJ
ajst-905	157	25	leaderboard	leaderboard	NOUN
ajst-905	157	26	.	.	PUNCT
ajst-905	158	1	c5	c5	PROPN
ajst-905	158	2	means	mean	VERB
ajst-905	158	3	there	there	PRON
ajst-905	158	4	are	be	VERB
ajst-905	158	5	5	5	NUM
ajst-905	158	6	standard	standard	ADJ
ajst-905	158	7	captions	caption	NOUN
ajst-905	158	8	per	per	ADP
ajst-905	158	9	image	image	NOUN
ajst-905	158	10	while	while	SCONJ
ajst-905	158	11	c40	c40	NOUN
ajst-905	158	12	means	mean	VERB
ajst-905	158	13	there	there	PRON
ajst-905	158	14	are	be	VERB
ajst-905	158	15	40	40	NUM
ajst-905	158	16	captions	caption	NOUN
ajst-905	158	17	per	per	ADP
ajst-905	158	18	image	image	NOUN
ajst-905	158	19	.	.	PUNCT
ajst-905	159	1	b-1	b-1	PROPN
ajst-905	159	2	b-2	b-2	PROPN
ajst-905	159	3	b-3	b-3	PROPN
ajst-905	159	4	b-4	b-4	PROPN
ajst-905	159	5	m	m	PROPN
ajst-905	159	6	r	r	NOUN
ajst-905	159	7	c	c	PROPN
ajst-905	159	8	c5	c5	PROPN
ajst-905	159	9	c40	c40	PROPN
ajst-905	159	10	c5	c5	PROPN
ajst-905	159	11	c40	c40	PROPN
ajst-905	159	12	c5	c5	PROPN
ajst-905	159	13	c40	c40	PROPN
ajst-905	159	14	c5	c5	PROPN
ajst-905	159	15	c40	c40	PROPN
ajst-905	159	16	c5	c5	PROPN
ajst-905	159	17	c40	c40	PROPN
ajst-905	159	18	c5	c5	PROPN
ajst-905	159	19	c40	c40	PROPN
ajst-905	159	20	c5	c5	PROPN
ajst-905	159	21	c40	c40	PROPN
ajst-905	159	22	scst	scst	PROPN
ajst-905	159	23	78.1	78.1	NUM
ajst-905	159	24	93.7	93.7	NUM
ajst-905	159	25	61.9	61.9	NUM
ajst-905	159	26	86.0	86.0	NUM
ajst-905	159	27	47.0	47.0	NUM
ajst-905	159	28	75.9	75.9	NUM
ajst-905	159	29	35.2	35.2	NUM
ajst-905	159	30	64.5	64.5	NUM
ajst-905	159	31	27.0	27.0	NUM
ajst-905	159	32	35.5	35.5	NUM
ajst-905	159	33	56.3	56.3	NUM
ajst-905	159	34	70.7	70.7	NUM
ajst-905	159	35	114.7	114.7	NUM
ajst-905	159	36	116.7	116.7	NUM
ajst-905	159	37	up	up	ADP
ajst-905	159	38	-	-	PUNCT
ajst-905	159	39	down	down	NOUN
ajst-905	159	40	80.2	80.2	NUM
ajst-905	159	41	95.2	95.2	NUM
ajst-905	159	42	64.1	64.1	NUM
ajst-905	159	43	88.8	88.8	NUM
ajst-905	159	44	49.1	49.1	NUM
ajst-905	159	45	79.4	79.4	NUM
ajst-905	159	46	36.9	36.9	NUM
ajst-905	159	47	68.5	68.5	NUM
ajst-905	159	48	27.6	27.6	NUM
ajst-905	159	49	36.7	36.7	NUM
ajst-905	159	50	57.1	57.1	NUM
ajst-905	159	51	72.4	72.4	NUM
ajst-905	159	52	117.9	117.9	NUM
ajst-905	159	53	120.5	120.5	NUM
ajst-905	159	54	rfnet	rfnet	NOUN
ajst-905	159	55	80.4	80.4	NUM
ajst-905	159	56	95.0	95.0	NUM
ajst-905	159	57	64.9	64.9	NUM
ajst-905	159	58	89.3	89.3	NUM
ajst-905	159	59	50.1	50.1	NUM
ajst-905	159	60	80.1	80.1	NUM
ajst-905	159	61	38.0	38.0	NUM
ajst-905	159	62	69.2	69.2	NUM
ajst-905	159	63	28.2	28.2	NUM
ajst-905	159	64	37.2	37.2	NUM
ajst-905	159	65	58.2	58.2	NUM
ajst-905	159	66	73.1	73.1	NUM
ajst-905	159	67	122.9	122.9	NUM
ajst-905	159	68	125.1	125.1	NUM
ajst-905	159	69	gcn	gcn	NOUN
ajst-905	159	70	-	-	PUNCT
ajst-905	159	71	lstm	lstm	ADJ
ajst-905	159	72	80.8	80.8	NUM
ajst-905	159	73	95.9	95.9	NUM
ajst-905	159	74	65.5	65.5	NUM
ajst-905	159	75	89.3	89.3	NUM
ajst-905	159	76	50.8	50.8	NUM
ajst-905	159	77	80.3	80.3	NUM
ajst-905	159	78	38.7	38.7	NUM
ajst-905	159	79	69.7	69.7	NUM
ajst-905	159	80	28.5	28.5	NUM
ajst-905	159	81	37.6	37.6	NUM
ajst-905	159	82	58.5	58.5	NUM
ajst-905	159	83	73.4	73.4	NUM
ajst-905	159	84	125.3	125.3	NUM
ajst-905	159	85	126.5	126.5	NUM
ajst-905	159	86	sgae	sgae	NOUN
ajst-905	159	87	81.0	81.0	NUM
ajst-905	159	88	95.3	95.3	NUM
ajst-905	159	89	65.6	65.6	NUM
ajst-905	159	90	89.5	89.5	NUM
ajst-905	159	91	50.7	50.7	NUM
ajst-905	159	92	80.4	80.4	NUM
ajst-905	159	93	38.5	38.5	NUM
ajst-905	159	94	69.7	69.7	NUM
ajst-905	159	95	28.2	28.2	NUM
ajst-905	159	96	37.2	37.2	NUM
ajst-905	159	97	58.6	58.6	NUM
ajst-905	159	98	73.6	73.6	NUM
ajst-905	159	99	123.8	123.8	NUM
ajst-905	159	100	126.5	126.5	NUM
ajst-905	159	101	aoanet	aoanet	NOUN
ajst-905	159	102	81.0	81.0	NUM
ajst-905	159	103	95.0	95.0	NUM
ajst-905	159	104	65.8	65.8	NUM
ajst-905	159	105	89.6	89.6	NUM
ajst-905	159	106	51.4	51.4	NUM
ajst-905	159	107	81.3	81.3	NUM
ajst-905	159	108	39.4	39.4	NUM
ajst-905	159	109	71.2	71.2	NUM
ajst-905	159	110	29.1	29.1	NUM
ajst-905	159	111	38.5	38.5	NUM
ajst-905	159	112	58.9	58.9	NUM
ajst-905	159	113	74.5	74.5	NUM
ajst-905	159	114	126.9	126.9	NUM
ajst-905	159	115	129.6	129.6	NUM
ajst-905	159	116	𝑀	𝑀	PROPN
ajst-905	159	117	transformer	transformer	NOUN
ajst-905	159	118	81.6	81.6	NUM
ajst-905	159	119	96.0	96.0	NUM
ajst-905	159	120	66.4	66.4	NUM
ajst-905	159	121	90.8	90.8	NUM
ajst-905	159	122	51.8	51.8	NUM
ajst-905	159	123	82.7	82.7	NUM
ajst-905	159	124	39.7	39.7	NUM
ajst-905	159	125	72.8	72.8	NUM
ajst-905	159	126	29.4	29.4	NUM
ajst-905	159	127	39.0	39.0	NUM
ajst-905	159	128	59.2	59.2	NUM
ajst-905	159	129	74.8	74.8	NUM
ajst-905	159	130	129.3	129.3	NUM
ajst-905	159	131	132.1	132.1	NUM
ajst-905	159	132	x	x	NOUN
ajst-905	159	133	-	-	NOUN
ajst-905	159	134	transformer	transformer	ADJ
ajst-905	159	135	81.3	81.3	NUM
ajst-905	159	136	95.4	95.4	NUM
ajst-905	159	137	66.3	66.3	NUM
ajst-905	159	138	90.0	90.0	NUM
ajst-905	159	139	51.9	51.9	NUM
ajst-905	159	140	81.7	81.7	NUM
ajst-905	159	141	39.9	39.9	NUM
ajst-905	159	142	71.8	71.8	NUM
ajst-905	159	143	29.5	29.5	NUM
ajst-905	159	144	39.0	39.0	NUM
ajst-905	159	145	59.3	59.3	NUM
ajst-905	159	146	74.9	74.9	NUM
ajst-905	159	147	129.3	129.3	NUM
ajst-905	159	148	131.4	131.4	NUM
ajst-905	159	149	fpdt	fpdt	NOUN
ajst-905	159	150	81.5	81.5	NUM
ajst-905	159	151	96.2	96.2	NUM
ajst-905	159	152	66.5	66.5	NUM
ajst-905	159	153	91.0	91.0	NUM
ajst-905	159	154	51.7	51.7	NUM
ajst-905	159	155	81.9	81.9	NUM
ajst-905	159	156	40.3	40.3	NUM
ajst-905	159	157	72.9	72.9	NUM
ajst-905	159	158	29.6	29.6	NUM
ajst-905	159	159	39.2	39.2	NUM
ajst-905	159	160	59.3	59.3	NUM
ajst-905	159	161	75.1	75.1	NUM
ajst-905	159	162	129.5	129.5	NUM
ajst-905	159	163	131.8	131.8	NUM
ajst-905	159	164	table	table	NOUN
ajst-905	159	165	2	2	NUM
ajst-905	159	166	.	.	PUNCT
ajst-905	159	167	comparison	comparison	NOUN
ajst-905	159	168	with	with	ADP
ajst-905	159	169	the	the	DET
ajst-905	159	170	state	state	NOUN
ajst-905	159	171	of	of	ADP
ajst-905	159	172	the	the	DET
ajst-905	159	173	art	art	NOUN
ajst-905	159	174	on	on	ADP
ajst-905	159	175	the	the	DET
ajst-905	159	176	“	"	PUNCT
ajst-905	159	177	karpathy	karpathy	ADJ
ajst-905	159	178	”	"	PUNCT
ajst-905	159	179	test	test	NOUN
ajst-905	159	180	split	split	NOUN
ajst-905	159	181	,	,	PUNCT
ajst-905	159	182	in	in	ADP
ajst-905	159	183	single	single	ADJ
ajst-905	159	184	-	-	PUNCT
ajst-905	159	185	model	model	NOUN
ajst-905	159	186	setting	setting	NOUN
ajst-905	159	187	.	.	PUNCT
ajst-905	160	1	the	the	DET
ajst-905	160	2	b-1	b-1	PROPN
ajst-905	160	3	,	,	PUNCT
ajst-905	160	4	b-4	b-4	PROPN
ajst-905	160	5	,	,	PUNCT
ajst-905	160	6	m	m	PROPN
ajst-905	160	7	,	,	PUNCT
ajst-905	160	8	r	r	NOUN
ajst-905	160	9	,	,	PUNCT
ajst-905	160	10	c	c	NOUN
ajst-905	160	11	,	,	PUNCT
ajst-905	160	12	s	s	VERB
ajst-905	160	13	are	be	AUX
ajst-905	160	14	the	the	DET
ajst-905	160	15	abbreviation	abbreviation	NOUN
ajst-905	160	16	for	for	ADP
ajst-905	160	17	bleu-1	bleu-1	PROPN
ajst-905	160	18	,	,	PUNCT
ajst-905	160	19	bleu-4	bleu-4	PRON
ajst-905	160	20	,	,	PUNCT
ajst-905	160	21	meteor	meteor	NOUN
ajst-905	160	22	,	,	PUNCT
ajst-905	160	23	rouge	rouge	NOUN
ajst-905	160	24	,	,	PUNCT
ajst-905	160	25	cider	cider	NOUN
ajst-905	160	26	and	and	CCONJ
ajst-905	160	27	spice	spice	NOUN
ajst-905	160	28	scores	score	NOUN
ajst-905	160	29	,	,	PUNCT
ajst-905	160	30	and	and	CCONJ
ajst-905	160	31	all	all	DET
ajst-905	160	32	values	value	NOUN
ajst-905	160	33	are	be	AUX
ajst-905	160	34	reported	report	VERB
ajst-905	160	35	as	as	ADP
ajst-905	160	36	a	a	DET
ajst-905	160	37	percentage	percentage	NOUN
ajst-905	160	38	.	.	PUNCT
ajst-905	161	1	b-1	b-1	PROPN
ajst-905	161	2	b-4	b-4	PROPN
ajst-905	161	3	m	m	PROPN
ajst-905	162	1	r	r	NOUN
ajst-905	162	2	c	c	NOUN
ajst-905	162	3	s	s	NOUN
ajst-905	162	4	scst	scst	NOUN
ajst-905	162	5	34.2	34.2	NUM
ajst-905	162	6	26.7	26.7	NUM
ajst-905	162	7	57.7	57.7	NUM
ajst-905	162	8	114.0	114.0	NUM
ajst-905	162	9	up	up	ADP
ajst-905	162	10	-	-	PUNCT
ajst-905	162	11	down	down	ADP
ajst-905	162	12	79.8	79.8	NUM
ajst-905	162	13	36.3	36.3	NUM
ajst-905	162	14	27.7	27.7	NUM
ajst-905	162	15	56.9	56.9	NUM
ajst-905	162	16	120.1	120.1	NUM
ajst-905	162	17	21.4	21.4	NUM
ajst-905	162	18	rfnet	rfnet	NOUN
ajst-905	162	19	79.1	79.1	NUM
ajst-905	162	20	36.5	36.5	NUM
ajst-905	162	21	27.7	27.7	NUM
ajst-905	162	22	57.3	57.3	NUM
ajst-905	162	23	121.9	121.9	NUM
ajst-905	162	24	21.2	21.2	NUM
ajst-905	162	25	gcn	gcn	NOUN
ajst-905	162	26	-	-	PUNCT
ajst-905	162	27	lstm	lstm	ADJ
ajst-905	162	28	80.5	80.5	NUM
ajst-905	162	29	38.2	38.2	NUM
ajst-905	162	30	28.5	28.5	NUM
ajst-905	162	31	58.3	58.3	NUM
ajst-905	162	32	127.6	127.6	NUM
ajst-905	162	33	22.0	22.0	NUM
ajst-905	162	34	sgae	sgae	NOUN
ajst-905	162	35	80.8	80.8	NUM
ajst-905	162	36	38.4	38.4	NUM
ajst-905	162	37	28.4	28.4	NUM
ajst-905	162	38	58.6	58.6	NUM
ajst-905	162	39	127.8	127.8	NUM
ajst-905	162	40	22.1	22.1	NUM
ajst-905	162	41	ort	ort	NOUN
ajst-905	162	42	80.5	80.5	NUM
ajst-905	162	43	38.6	38.6	NUM
ajst-905	162	44	28.7	28.7	NUM
ajst-905	162	45	58.4	58.4	NUM
ajst-905	162	46	128.3	128.3	NUM
ajst-905	162	47	22.6	22.6	NUM
ajst-905	162	48	aoanet	aoanet	NOUN
ajst-905	162	49	80.2	80.2	NUM
ajst-905	162	50	38.9	38.9	NUM
ajst-905	162	51	29.2	29.2	NUM
ajst-905	162	52	58.8	58.8	NUM
ajst-905	162	53	129.8	129.8	NUM
ajst-905	162	54	22.4	22.4	NUM
ajst-905	162	55	𝑀	𝑀	PROPN
ajst-905	162	56	transformer	transformer	NOUN
ajst-905	162	57	80.8	80.8	NUM
ajst-905	162	58	39.1	39.1	NUM
ajst-905	162	59	29.2	29.2	NUM
ajst-905	162	60	58.6	58.6	NUM
ajst-905	162	61	131.2	131.2	NUM
ajst-905	162	62	22.6	22.6	NUM
ajst-905	162	63	x	x	NOUN
ajst-905	162	64	-	-	NOUN
ajst-905	162	65	transformer	transformer	NOUN
ajst-905	162	66	80.9	80.9	NUM
ajst-905	162	67	39.7	39.7	NUM
ajst-905	162	68	29.5	29.5	NUM
ajst-905	162	69	59.1	59.1	NUM
ajst-905	162	70	132.8	132.8	NUM
ajst-905	162	71	23.4	23.4	NUM
ajst-905	162	72	fpdt	fpdt	NOUN
ajst-905	162	73	81.1	81.1	NUM
ajst-905	162	74	39.7	39.7	NUM
ajst-905	162	75	29.6	29.6	NUM
ajst-905	162	76	59.2	59.2	NUM
ajst-905	162	77	133.2	133.2	NUM
ajst-905	162	78	23.1	23.1	NUM
ajst-905	162	79	table	table	NOUN
ajst-905	162	80	3	3	NUM
ajst-905	162	81	.	.	PUNCT
ajst-905	162	82	comparison	comparison	NOUN
ajst-905	162	83	with	with	ADP
ajst-905	162	84	the	the	DET
ajst-905	162	85	state	state	NOUN
ajst-905	162	86	of	of	ADP
ajst-905	162	87	art	art	NOUN
ajst-905	162	88	models	model	NOUN
ajst-905	162	89	on	on	ADP
ajst-905	162	90	the	the	DET
ajst-905	162	91	“	"	PUNCT
ajst-905	162	92	karpathy	karpathy	ADJ
ajst-905	162	93	”	"	PUNCT
ajst-905	162	94	test	test	NOUN
ajst-905	162	95	split	split	NOUN
ajst-905	162	96	,	,	PUNCT
ajst-905	162	97	4	4	NUM
ajst-905	162	98	models	model	NOUN
ajst-905	162	99	are	be	AUX
ajst-905	162	100	used	use	VERB
ajst-905	162	101	for	for	ADP
ajst-905	162	102	ensemble	ensemble	ADJ
ajst-905	162	103	.	.	PUNCT
ajst-905	163	1	b-1	b-1	PROPN
ajst-905	163	2	b-4	b-4	PROPN
ajst-905	163	3	m	m	PROPN
ajst-905	163	4	r	r	NOUN
ajst-905	163	5	c	c	NOUN
ajst-905	163	6	s	s	NOUN
ajst-905	163	7	scst	scst	NOUN
ajst-905	163	8	35.4	35.4	NUM
ajst-905	163	9	27.1	27.1	NUM
ajst-905	163	10	56.6	56.6	NUM
ajst-905	163	11	117.5	117.5	NUM
ajst-905	163	12	gcn	gcn	NOUN
ajst-905	163	13	-	-	PUNCT
ajst-905	163	14	lstm	lstm	ADJ
ajst-905	163	15	80.9	80.9	NUM
ajst-905	163	16	38.3	38.3	NUM
ajst-905	163	17	28.6	28.6	NUM
ajst-905	163	18	58.5	58.5	NUM
ajst-905	163	19	128.7	128.7	NUM
ajst-905	163	20	22.1	22.1	NUM
ajst-905	163	21	sgae	sgae	NOUN
ajst-905	163	22	81.1	81.1	NUM
ajst-905	163	23	39.0	39.0	NUM
ajst-905	163	24	28.4	28.4	NUM
ajst-905	163	25	58.9	58.9	NUM
ajst-905	163	26	129.1	129.1	NUM
ajst-905	163	27	22.2	22.2	NUM
ajst-905	163	28	aoanet	aoanet	NOUN
ajst-905	163	29	81.6	81.6	NUM
ajst-905	163	30	40.2	40.2	NUM
ajst-905	163	31	29.3	29.3	NUM
ajst-905	163	32	59.4	59.4	NUM
ajst-905	163	33	132.0	132.0	NUM
ajst-905	163	34	22.8	22.8	NUM
ajst-905	163	35	𝑀	𝑀	PROPN
ajst-905	163	36	transformer	transformer	NOUN
ajst-905	163	37	82.0	82.0	NUM
ajst-905	163	38	40.5	40.5	NUM
ajst-905	163	39	29.7	29.7	NUM
ajst-905	163	40	59.5	59.5	NUM
ajst-905	163	41	134.5	134.5	NUM
ajst-905	163	42	23.5	23.5	NUM
ajst-905	163	43	x	x	NOUN
ajst-905	163	44	-	-	NOUN
ajst-905	163	45	transformer	transformer	NOUN
ajst-905	163	46	81.7	81.7	NUM
ajst-905	163	47	40.7	40.7	NUM
ajst-905	163	48	29.9	29.9	NUM
ajst-905	163	49	59.7	59.7	NUM
ajst-905	163	50	135.3	135.3	NUM
ajst-905	163	51	23.8	23.8	NUM
ajst-905	163	52	fpdt	fpdt	NOUN
ajst-905	163	53	81.8	81.8	NUM
ajst-905	163	54	40.9	40.9	NUM
ajst-905	163	55	29.9	29.9	NUM
ajst-905	163	56	59.8	59.8	NUM
ajst-905	163	57	135.6	135.6	NUM
ajst-905	163	58	23.5	23.5	NUM
ajst-905	163	59	as	as	SCONJ
ajst-905	163	60	shown	show	VERB
ajst-905	163	61	in	in	ADP
ajst-905	163	62	the	the	DET
ajst-905	163	63	table	table	NOUN
ajst-905	163	64	2	2	NUM
ajst-905	163	65	,	,	PUNCT
ajst-905	163	66	except	except	SCONJ
ajst-905	163	67	the	the	DET
ajst-905	163	68	spice	spice	NOUN
ajst-905	163	69	score	score	NOUN
ajst-905	163	70	,	,	PUNCT
ajst-905	163	71	our	our	PRON
ajst-905	163	72	model	model	NOUN
ajst-905	163	73	fpdt	fpdt	NOUN
ajst-905	163	74	outperforms	outperform	VERB
ajst-905	163	75	the	the	DET
ajst-905	163	76	other	other	ADJ
ajst-905	163	77	models	model	NOUN
ajst-905	163	78	in	in	ADP
ajst-905	163	79	blue-1，blue-4	blue-1，blue-4	NOUN
ajst-905	163	80	,	,	PUNCT
ajst-905	163	81	meteor	meteor	NOUN
ajst-905	163	82	,	,	PUNCT
ajst-905	163	83	rouge	rouge	NOUN
ajst-905	163	84	and	and	CCONJ
ajst-905	163	85	cider	cider	NOUN
ajst-905	163	86	scores	score	NOUN
ajst-905	163	87	.	.	PUNCT
ajst-905	164	1	and	and	CCONJ
ajst-905	164	2	in	in	ADP
ajst-905	164	3	cider	cider	NOUN
ajst-905	164	4	score	score	NOUN
ajst-905	164	5	our	our	PRON
ajst-905	164	6	model	model	NOUN
ajst-905	164	7	achieves	achieve	VERB
ajst-905	164	8	a	a	DET
ajst-905	164	9	133.2	133.2	NUM
ajst-905	164	10	%	%	NOUN
ajst-905	164	11	performance	performance	NOUN
ajst-905	164	12	ahead	ahead	ADV
ajst-905	164	13	of	of	ADP
ajst-905	164	14	the	the	DET
ajst-905	164	15	xtransformer	xtransformer	PROPN
ajst-905	164	16	model	model	PROPN
ajst-905	164	17	,	,	PUNCT
ajst-905	164	18	the	the	DET
ajst-905	164	19	best	good	ADJ
ajst-905	164	20	model	model	NOUN
ajst-905	164	21	in	in	ADP
ajst-905	164	22	the	the	DET
ajst-905	164	23	nine	nine	NUM
ajst-905	164	24	models	model	NOUN
ajst-905	164	25	,	,	PUNCT
ajst-905	164	26	by	by	ADP
ajst-905	164	27	about	about	ADV
ajst-905	164	28	1	1	NUM
ajst-905	164	29	%	%	NOUN
ajst-905	164	30	.	.	PUNCT
ajst-905	165	1	the	the	DET
ajst-905	165	2	outstanding	outstanding	ADJ
ajst-905	165	3	performance	performance	NOUN
ajst-905	165	4	in	in	ADP
ajst-905	165	5	the	the	DET
ajst-905	165	6	scores	score	NOUN
ajst-905	165	7	verifies	verifie	NOUN
ajst-905	165	8	the	the	DET
ajst-905	165	9	effectiveness	effectiveness	NOUN
ajst-905	165	10	of	of	ADP
ajst-905	165	11	our	our	PRON
ajst-905	165	12	proposed	propose	VERB
ajst-905	165	13	model	model	NOUN
ajst-905	165	14	.	.	PUNCT
ajst-905	166	1	in	in	ADP
ajst-905	166	2	table	table	NOUN
ajst-905	166	3	3	3	NUM
ajst-905	166	4	,	,	PUNCT
ajst-905	166	5	we	we	PRON
ajst-905	166	6	report	report	VERB
ajst-905	166	7	the	the	DET
ajst-905	166	8	fpdt	fpdt	NOUN
ajst-905	166	9	’s	’s	PART
ajst-905	166	10	performance	performance	NOUN
ajst-905	166	11	compared	compare	VERB
ajst-905	166	12	with	with	ADP
ajst-905	166	13	other	other	ADJ
ajst-905	166	14	models	model	NOUN
ajst-905	166	15	,	,	PUNCT
ajst-905	166	16	using	use	VERB
ajst-905	166	17	ensembles	ensemble	NOUN
ajst-905	166	18	of	of	ADP
ajst-905	166	19	four	four	NUM
ajst-905	166	20	models	model	NOUN
ajst-905	166	21	.	.	PUNCT
ajst-905	167	1	our	our	PRON
ajst-905	167	2	ensembled	ensembled	ADJ
ajst-905	167	3	model	model	NOUN
ajst-905	167	4	has	have	AUX
ajst-905	167	5	improved	improve	VERB
ajst-905	167	6	to	to	ADP
ajst-905	167	7	some	some	DET
ajst-905	167	8	extent	extent	NOUN
ajst-905	167	9	in	in	ADP
ajst-905	167	10	all	all	DET
ajst-905	167	11	scores	score	NOUN
ajst-905	167	12	.	.	PUNCT
ajst-905	168	1	except	except	SCONJ
ajst-905	168	2	for	for	ADP
ajst-905	168	3	the	the	DET
ajst-905	168	4	spice	spice	NOUN
ajst-905	168	5	score	score	NOUN
ajst-905	168	6	,	,	PUNCT
ajst-905	168	7	which	which	PRON
ajst-905	168	8	is	be	AUX
ajst-905	168	9	slightly	slightly	ADV
ajst-905	168	10	lower	low	ADJ
ajst-905	168	11	than	than	ADP
ajst-905	168	12	the	the	DET
ajst-905	168	13	x	x	ADJ
ajst-905	168	14	-	-	ADJ
ajst-905	168	15	transformer	transformer	ADJ
ajst-905	168	16	model	model	NOUN
ajst-905	168	17	,	,	PUNCT
ajst-905	168	18	the	the	DET
ajst-905	168	19	other	other	ADJ
ajst-905	168	20	five	five	NUM
ajst-905	168	21	scores	score	NOUN
ajst-905	168	22	are	be	AUX
ajst-905	168	23	still	still	ADV
ajst-905	168	24	the	the	DET
ajst-905	168	25	best	good	ADJ
ajst-905	168	26	.	.	PUNCT
ajst-905	169	1	online	online	ADJ
ajst-905	169	2	evaluation	evaluation	NOUN
ajst-905	169	3	.	.	PUNCT
ajst-905	170	1	we	we	PRON
ajst-905	170	2	also	also	ADV
ajst-905	170	3	report	report	VERB
ajst-905	170	4	the	the	DET
ajst-905	170	5	performance	performance	NOUN
ajst-905	170	6	of	of	ADP
ajst-905	170	7	the	the	DET
ajst-905	170	8	model	model	NOUN
ajst-905	170	9	on	on	ADP
ajst-905	170	10	the	the	DET
ajst-905	170	11	online	online	PROPN
ajst-905	170	12	coco	coco	PROPN
ajst-905	170	13	test	test	PROPN
ajst-905	170	14	server	server	NOUN
ajst-905	170	15	.	.	PUNCT
ajst-905	171	1	the	the	DET
ajst-905	171	2	coco	coco	PROPN
ajst-905	171	3	test	test	NOUN
ajst-905	171	4	split	split	VERB
ajst-905	171	5	used	use	VERB
ajst-905	171	6	in	in	ADP
ajst-905	171	7	the	the	DET
ajst-905	171	8	online	online	ADJ
ajst-905	171	9	test	test	NOUN
ajst-905	171	10	is	be	AUX
ajst-905	171	11	not	not	PART
ajst-905	171	12	publicly	publicly	ADV
ajst-905	171	13	available	available	ADJ
ajst-905	171	14	.	.	PUNCT
ajst-905	172	1	the	the	DET
ajst-905	172	2	results	result	NOUN
ajst-905	172	3	are	be	AUX
ajst-905	172	4	shown	show	VERB
ajst-905	172	5	in	in	ADP
ajst-905	172	6	table	table	NOUN
ajst-905	172	7	1	1	NUM
ajst-905	172	8	.	.	PUNCT
ajst-905	172	9	from	from	ADP
ajst-905	172	10	the	the	DET
ajst-905	172	11	test	test	NOUN
ajst-905	172	12	results	result	NOUN
ajst-905	172	13	,	,	PUNCT
ajst-905	172	14	compared	compare	VERB
ajst-905	172	15	to	to	ADP
ajst-905	172	16	x	x	NOUN
ajst-905	172	17	-	-	NOUN
ajst-905	172	18	transformer	transformer	NOUN
ajst-905	172	19	,	,	PUNCT
ajst-905	172	20	the	the	DET
ajst-905	172	21	second	second	ADJ
ajst-905	172	22	model	model	NOUN
ajst-905	172	23	in	in	ADP
ajst-905	172	24	the	the	DET
ajst-905	172	25	ranking	ranking	NOUN
ajst-905	172	26	,	,	PUNCT
ajst-905	172	27	our	our	PRON
ajst-905	172	28	model	model	NOUN
ajst-905	172	29	has	have	AUX
ajst-905	172	30	improved	improve	VERB
ajst-905	172	31	in	in	ADP
ajst-905	172	32	most	most	ADJ
ajst-905	172	33	evaluation	evaluation	NOUN
ajst-905	172	34	indicators	indicator	NOUN
ajst-905	172	35	.	.	PUNCT
ajst-905	173	1	qualitative	qualitative	ADJ
ajst-905	173	2	results	result	NOUN
ajst-905	173	3	and	and	CCONJ
ajst-905	173	4	visualization	visualization	NOUN
ajst-905	173	5	.	.	PUNCT
ajst-905	174	1	to	to	PART
ajst-905	174	2	make	make	VERB
ajst-905	174	3	the	the	DET
ajst-905	174	4	comparison	comparison	NOUN
ajst-905	174	5	of	of	ADP
ajst-905	174	6	model	model	NOUN
ajst-905	174	7	effects	effect	NOUN
ajst-905	174	8	more	more	ADV
ajst-905	174	9	intuitive	intuitive	ADJ
ajst-905	174	10	,	,	PUNCT
ajst-905	174	11	we	we	PRON
ajst-905	174	12	showed	show	VERB
ajst-905	174	13	four	four	NUM
ajst-905	174	14	images	image	NOUN
ajst-905	174	15	of	of	ADP
ajst-905	174	16	the	the	DET
ajst-905	174	17	test	test	NOUN
ajst-905	174	18	results	result	NOUN
ajst-905	174	19	.	.	PUNCT
ajst-905	175	1	we	we	PRON
ajst-905	175	2	compared	compare	VERB
ajst-905	175	3	the	the	DET
ajst-905	175	4	results	result	NOUN
ajst-905	175	5	with	with	ADP
ajst-905	175	6	the	the	DET
ajst-905	175	7	results	result	NOUN
ajst-905	175	8	of	of	ADP
ajst-905	175	9	different	different	ADJ
ajst-905	175	10	models	model	NOUN
ajst-905	175	11	.	.	PUNCT
ajst-905	176	1	119	119	NUM
ajst-905	176	2	figure	figure	NOUN
ajst-905	176	3	4	4	NUM
ajst-905	176	4	.	.	PUNCT
ajst-905	177	1	the	the	DET
ajst-905	177	2	visualization	visualization	NOUN
ajst-905	177	3	results	result	NOUN
ajst-905	177	4	compared	compare	VERB
ajst-905	177	5	with	with	ADP
ajst-905	177	6	different	different	ADJ
ajst-905	177	7	models	model	NOUN
ajst-905	177	8	.	.	PUNCT
ajst-905	178	1	as	as	SCONJ
ajst-905	178	2	we	we	PRON
ajst-905	178	3	can	can	AUX
ajst-905	178	4	see	see	VERB
ajst-905	178	5	,	,	PUNCT
ajst-905	178	6	our	our	PRON
ajst-905	178	7	model	model	NOUN
ajst-905	178	8	is	be	AUX
ajst-905	178	9	better	well	ADJ
ajst-905	178	10	than	than	ADP
ajst-905	178	11	the	the	DET
ajst-905	178	12	other	other	ADJ
ajst-905	178	13	three	three	NUM
ajst-905	178	14	models	model	NOUN
ajst-905	178	15	in	in	ADP
ajst-905	178	16	terms	term	NOUN
ajst-905	178	17	of	of	ADP
ajst-905	178	18	coverage	coverage	NOUN
ajst-905	178	19	of	of	ADP
ajst-905	178	20	image	image	NOUN
ajst-905	178	21	content	content	NOUN
ajst-905	178	22	and	and	CCONJ
ajst-905	178	23	the	the	DET
ajst-905	178	24	finegrainedness	finegrainedness	NOUN
ajst-905	178	25	of	of	ADP
ajst-905	178	26	the	the	DET
ajst-905	178	27	image	image	NOUN
ajst-905	178	28	caption	caption	NOUN
ajst-905	178	29	.	.	PUNCT
ajst-905	179	1	for	for	ADP
ajst-905	179	2	example	example	NOUN
ajst-905	179	3	,	,	PUNCT
ajst-905	179	4	in	in	ADP
ajst-905	179	5	the	the	DET
ajst-905	179	6	fourth	fourth	ADJ
ajst-905	179	7	image	image	NOUN
ajst-905	179	8	,	,	PUNCT
ajst-905	179	9	fpdt	fpdt	NOUN
ajst-905	179	10	gives	give	VERB
ajst-905	179	11	a	a	DET
ajst-905	179	12	good	good	ADJ
ajst-905	179	13	description	description	NOUN
ajst-905	179	14	of	of	ADP
ajst-905	179	15	the	the	DET
ajst-905	179	16	image	image	NOUN
ajst-905	179	17	content	content	NOUN
ajst-905	179	18	,	,	PUNCT
ajst-905	179	19	such	such	ADJ
ajst-905	179	20	as	as	ADP
ajst-905	179	21	photoing	photoing	NOUN
ajst-905	179	22	,	,	PUNCT
ajst-905	179	23	brushing	brushing	NOUN
ajst-905	179	24	teeth	tooth	NOUN
ajst-905	179	25	,	,	PUNCT
ajst-905	179	26	bathroom	bathroom	NOUN
ajst-905	179	27	,	,	PUNCT
ajst-905	179	28	etc	etc	X
ajst-905	179	29	.	.	X
ajst-905	179	30	other	other	ADJ
ajst-905	179	31	models	model	NOUN
ajst-905	179	32	do	do	AUX
ajst-905	179	33	n't	not	PART
ajst-905	179	34	do	do	VERB
ajst-905	179	35	a	a	DET
ajst-905	179	36	very	very	ADV
ajst-905	179	37	good	good	ADJ
ajst-905	179	38	job	job	NOUN
ajst-905	179	39	of	of	ADP
ajst-905	179	40	describing	describe	VERB
ajst-905	179	41	it	it	PRON
ajst-905	179	42	.	.	PUNCT
ajst-905	180	1	5	5	X
ajst-905	180	2	.	.	X
ajst-905	180	3	conclusion	conclusion	NOUN
ajst-905	180	4	in	in	ADP
ajst-905	180	5	this	this	DET
ajst-905	180	6	paper	paper	NOUN
ajst-905	180	7	,	,	PUNCT
ajst-905	180	8	we	we	PRON
ajst-905	180	9	propose	propose	VERB
ajst-905	180	10	a	a	DET
ajst-905	180	11	new	new	ADJ
ajst-905	180	12	transformer	transformer	NOUN
ajst-905	180	13	-	-	PUNCT
ajst-905	180	14	based	base	VERB
ajst-905	180	15	model	model	NOUN
ajst-905	180	16	,	,	PUNCT
ajst-905	180	17	fpdt	fpdt	ADJ
ajst-905	180	18	,	,	PUNCT
ajst-905	180	19	to	to	PART
ajst-905	180	20	improve	improve	VERB
ajst-905	180	21	the	the	DET
ajst-905	180	22	model	model	NOUN
ajst-905	180	23	by	by	ADP
ajst-905	180	24	fusing	fuse	VERB
ajst-905	180	25	grid	grid	NOUN
ajst-905	180	26	features	feature	NOUN
ajst-905	180	27	and	and	CCONJ
ajst-905	180	28	region	region	NOUN
ajst-905	180	29	features	feature	NOUN
ajst-905	180	30	.	.	PUNCT
ajst-905	181	1	it	it	PRON
ajst-905	181	2	also	also	ADV
ajst-905	181	3	achieves	achieve	VERB
ajst-905	181	4	better	well	ADV
ajst-905	181	5	decoding	decode	VERB
ajst-905	181	6	effect	effect	NOUN
ajst-905	181	7	by	by	ADP
ajst-905	181	8	improving	improve	VERB
ajst-905	181	9	the	the	DET
ajst-905	181	10	input	input	NOUN
ajst-905	181	11	information	information	NOUN
ajst-905	181	12	in	in	ADP
ajst-905	181	13	the	the	DET
ajst-905	181	14	decoding	decode	VERB
ajst-905	181	15	phase	phase	NOUN
ajst-905	181	16	which	which	PRON
ajst-905	181	17	allows	allow	VERB
ajst-905	181	18	the	the	DET
ajst-905	181	19	model	model	NOUN
ajst-905	181	20	have	have	VERB
ajst-905	181	21	more	more	ADJ
ajst-905	181	22	information	information	NOUN
ajst-905	181	23	about	about	ADP
ajst-905	181	24	the	the	DET
ajst-905	181	25	test	test	NOUN
ajst-905	181	26	images	image	NOUN
ajst-905	181	27	.	.	PUNCT
ajst-905	182	1	the	the	DET
ajst-905	182	2	final	final	ADJ
ajst-905	182	3	test	test	NOUN
ajst-905	182	4	results	result	NOUN
ajst-905	182	5	demonstrate	demonstrate	VERB
ajst-905	182	6	that	that	SCONJ
ajst-905	182	7	our	our	PRON
ajst-905	182	8	proposed	propose	VERB
ajst-905	182	9	module	module	NOUN
ajst-905	182	10	can	can	AUX
ajst-905	182	11	effectively	effectively	ADV
ajst-905	182	12	improve	improve	VERB
ajst-905	182	13	the	the	DET
ajst-905	182	14	model	model	NOUN
ajst-905	182	15	’s	’s	PART
ajst-905	182	16	performance	performance	NOUN
ajst-905	182	17	.	.	PUNCT
ajst-905	183	1	references	reference	NOUN
ajst-905	183	2	[	[	X
ajst-905	183	3	1	1	NUM
ajst-905	183	4	]	]	X
ajst-905	183	5	anderson	anderson	PROPN
ajst-905	183	6	,	,	PUNCT
ajst-905	183	7	peter	peter	PROPN
ajst-905	183	8	,	,	PUNCT
ajst-905	183	9	et	et	PROPN
ajst-905	183	10	al	al	PROPN
ajst-905	183	11	.	.	PUNCT
ajst-905	184	1	"	"	PUNCT
ajst-905	184	2	bottom	bottom	ADJ
ajst-905	184	3	-	-	PUNCT
ajst-905	184	4	up	up	NOUN
ajst-905	184	5	and	and	CCONJ
ajst-905	184	6	top	top	ADJ
ajst-905	184	7	-	-	PUNCT
ajst-905	184	8	down	down	ADP
ajst-905	184	9	attention	attention	NOUN
ajst-905	184	10	for	for	ADP
ajst-905	184	11	image	image	NOUN
ajst-905	184	12	captioning	captioning	NOUN
ajst-905	184	13	and	and	CCONJ
ajst-905	184	14	visual	visual	ADJ
ajst-905	184	15	question	question	NOUN
ajst-905	184	16	answering	answer	VERB
ajst-905	184	17	.	.	PUNCT
ajst-905	184	18	"	"	PUNCT
ajst-905	185	1	proceedings	proceeding	NOUN
ajst-905	185	2	of	of	ADP
ajst-905	185	3	the	the	DET
ajst-905	185	4	ieee	ieee	NOUN
ajst-905	185	5	conference	conference	NOUN
ajst-905	185	6	on	on	ADP
ajst-905	185	7	computer	computer	NOUN
ajst-905	185	8	vision	vision	NOUN
ajst-905	185	9	and	and	CCONJ
ajst-905	185	10	pattern	pattern	NOUN
ajst-905	185	11	recognition	recognition	NOUN
ajst-905	185	12	.	.	PUNCT
ajst-905	186	1	2018	2018	NUM
ajst-905	186	2	.	.	PUNCT
ajst-905	187	1	[	[	X
ajst-905	187	2	2	2	NUM
ajst-905	187	3	]	]	SYM
ajst-905	187	4	cornia	cornia	PROPN
ajst-905	187	5	,	,	PUNCT
ajst-905	187	6	marcella	marcella	PROPN
ajst-905	187	7	,	,	PUNCT
ajst-905	187	8	lorenzo	lorenzo	NOUN
ajst-905	187	9	baraldi	baraldi	NOUN
ajst-905	187	10	,	,	PUNCT
ajst-905	187	11	and	and	CCONJ
ajst-905	187	12	rita	rita	PROPN
ajst-905	187	13	cucchiara	cucchiara	PROPN
ajst-905	187	14	.	.	PUNCT
ajst-905	188	1	"	"	PUNCT
ajst-905	188	2	show	show	NOUN
ajst-905	188	3	,	,	PUNCT
ajst-905	188	4	control	control	NOUN
ajst-905	188	5	and	and	CCONJ
ajst-905	188	6	tell	tell	VERB
ajst-905	188	7	:	:	PUNCT
ajst-905	188	8	a	a	DET
ajst-905	188	9	framework	framework	NOUN
ajst-905	188	10	for	for	ADP
ajst-905	188	11	generating	generate	VERB
ajst-905	188	12	controllable	controllable	ADJ
ajst-905	188	13	and	and	CCONJ
ajst-905	188	14	grounded	grounded	ADJ
ajst-905	188	15	captions	caption	NOUN
ajst-905	188	16	.	.	PUNCT
ajst-905	188	17	"	"	PUNCT
ajst-905	188	18	proceedings	proceeding	NOUN
ajst-905	188	19	of	of	ADP
ajst-905	188	20	the	the	DET
ajst-905	188	21	ieee	ieee	NOUN
ajst-905	188	22	/	/	SYM
ajst-905	188	23	cvf	cvf	NOUN
ajst-905	188	24	conference	conference	NOUN
ajst-905	188	25	on	on	ADP
ajst-905	188	26	computer	computer	NOUN
ajst-905	188	27	vision	vision	NOUN
ajst-905	188	28	and	and	CCONJ
ajst-905	188	29	pattern	pattern	NOUN
ajst-905	188	30	recognition	recognition	NOUN
ajst-905	188	31	.	.	PUNCT
ajst-905	189	1	2019	2019	NUM
ajst-905	189	2	.	.	PUNCT
ajst-905	190	1	[	[	X
ajst-905	190	2	3	3	X
ajst-905	190	3	]	]	X
ajst-905	190	4	junhua	junhua	PROPN
ajst-905	190	5	mao	mao	PROPN
ajst-905	190	6	,	,	PUNCT
ajst-905	190	7	wei	wei	PROPN
ajst-905	190	8	xu	xu	PROPN
ajst-905	190	9	,	,	PUNCT
ajst-905	190	10	yi	yi	PROPN
ajst-905	190	11	y	y	PROPN
ajst-905	190	12	ang	ang	PROPN
ajst-905	190	13	,	,	PUNCT
ajst-905	190	14	jiang	jiang	PROPN
ajst-905	190	15	wang	wang	PROPN
ajst-905	190	16	,	,	PUNCT
ajst-905	190	17	and	and	CCONJ
ajst-905	190	18	alan	alan	PROPN
ajst-905	190	19	ly	ly	VERB
ajst-905	190	20	uille	uille	PROPN
ajst-905	190	21	.	.	PUNCT
ajst-905	191	1	explain	explain	VERB
ajst-905	191	2	images	image	NOUN
ajst-905	191	3	with	with	ADP
ajst-905	191	4	multimodal	multimodal	NOUN
ajst-905	191	5	recurrent	recurrent	ADJ
ajst-905	191	6	neural	neural	ADJ
ajst-905	191	7	networks	network	NOUN
ajst-905	191	8	.	.	PUNCT
ajst-905	192	1	arxiv	arxiv	PROPN
ajst-905	192	2	preprint	preprint	PROPN
ajst-905	192	3	arxiv:1410.1090	arxiv:1410.1090	PROPN
ajst-905	192	4	,	,	PUNCT
ajst-905	192	5	2014	2014	NUM
ajst-905	192	6	.	.	PUNCT
ajst-905	193	1	1	1	NUM
ajst-905	193	2	,	,	PUNCT
ajst-905	193	3	2	2	NUM
ajst-905	194	1	[	[	SYM
ajst-905	194	2	4	4	NUM
ajst-905	194	3	]	]	PUNCT
ajst-905	194	4	vinyals	vinyal	NOUN
ajst-905	194	5	,	,	PUNCT
ajst-905	194	6	oriol	oriol	PROPN
ajst-905	194	7	,	,	PUNCT
ajst-905	194	8	et	et	PROPN
ajst-905	194	9	al	al	PROPN
ajst-905	194	10	.	.	PUNCT
ajst-905	195	1	"	"	PUNCT
ajst-905	195	2	show	show	VERB
ajst-905	195	3	and	and	CCONJ
ajst-905	195	4	tell	tell	VERB
ajst-905	195	5	:	:	PUNCT
ajst-905	195	6	a	a	DET
ajst-905	195	7	neural	neural	ADJ
ajst-905	195	8	image	image	NOUN
ajst-905	195	9	caption	caption	NOUN
ajst-905	195	10	generator	generator	NOUN
ajst-905	195	11	.	.	PUNCT
ajst-905	195	12	"	"	PUNCT
ajst-905	196	1	proceedings	proceeding	NOUN
ajst-905	196	2	of	of	ADP
ajst-905	196	3	the	the	DET
ajst-905	196	4	ieee	ieee	NOUN
ajst-905	196	5	conference	conference	NOUN
ajst-905	196	6	on	on	ADP
ajst-905	196	7	computer	computer	NOUN
ajst-905	196	8	vision	vision	NOUN
ajst-905	196	9	and	and	CCONJ
ajst-905	196	10	pattern	pattern	NOUN
ajst-905	196	11	recognition	recognition	NOUN
ajst-905	196	12	.	.	PUNCT
ajst-905	197	1	2015	2015	NUM
ajst-905	197	2	.	.	PUNCT
ajst-905	198	1	[	[	X
ajst-905	198	2	5	5	NUM
ajst-905	198	3	]	]	PUNCT
ajst-905	198	4	vinyals	vinyal	NOUN
ajst-905	198	5	,	,	PUNCT
ajst-905	198	6	oriol	oriol	PROPN
ajst-905	198	7	,	,	PUNCT
ajst-905	198	8	et	et	PROPN
ajst-905	198	9	al	al	PROPN
ajst-905	198	10	.	.	PUNCT
ajst-905	199	1	"	"	PUNCT
ajst-905	199	2	show	show	VERB
ajst-905	199	3	and	and	CCONJ
ajst-905	199	4	tell	tell	VERB
ajst-905	199	5	:	:	PUNCT
ajst-905	199	6	lessons	lesson	NOUN
ajst-905	199	7	learned	learn	VERB
ajst-905	199	8	from	from	ADP
ajst-905	199	9	the	the	DET
ajst-905	199	10	2015	2015	NUM
ajst-905	199	11	mscoco	mscoco	NOUN
ajst-905	199	12	image	image	NOUN
ajst-905	199	13	captioning	captioning	NOUN
ajst-905	199	14	challenge	challenge	NOUN
ajst-905	199	15	.	.	PUNCT
ajst-905	199	16	"	"	PUNCT
ajst-905	200	1	ieee	ieee	NOUN
ajst-905	200	2	transactions	transaction	NOUN
ajst-905	200	3	on	on	ADP
ajst-905	200	4	pattern	pattern	NOUN
ajst-905	200	5	analysis	analysis	NOUN
ajst-905	200	6	and	and	CCONJ
ajst-905	200	7	machine	machine	NOUN
ajst-905	200	8	intelligence	intelligence	NOUN
ajst-905	200	9	39.4	39.4	NUM
ajst-905	200	10	(	(	PUNCT
ajst-905	200	11	2016	2016	NUM
ajst-905	200	12	):	):	PUNCT
ajst-905	200	13	652663	652663	NUM
ajst-905	200	14	.	.	PUNCT
ajst-905	201	1	[	[	X
ajst-905	201	2	6	6	NUM
ajst-905	201	3	]	]	SYM
ajst-905	201	4	xu	xu	PROPN
ajst-905	201	5	,	,	PUNCT
ajst-905	201	6	kelvin	kelvin	PROPN
ajst-905	201	7	,	,	PUNCT
ajst-905	201	8	et	et	PROPN
ajst-905	201	9	al	al	PROPN
ajst-905	201	10	.	.	PUNCT
ajst-905	202	1	"	"	PUNCT
ajst-905	202	2	show	show	NOUN
ajst-905	202	3	,	,	PUNCT
ajst-905	202	4	attend	attend	VERB
ajst-905	202	5	and	and	CCONJ
ajst-905	202	6	tell	tell	VERB
ajst-905	202	7	:	:	PUNCT
ajst-905	202	8	neural	neural	ADJ
ajst-905	202	9	image	image	NOUN
ajst-905	202	10	caption	caption	NOUN
ajst-905	202	11	generation	generation	NOUN
ajst-905	202	12	with	with	ADP
ajst-905	202	13	visual	visual	ADJ
ajst-905	202	14	attention	attention	NOUN
ajst-905	202	15	.	.	PUNCT
ajst-905	202	16	"	"	PUNCT
ajst-905	203	1	international	international	ADJ
ajst-905	203	2	conference	conference	NOUN
ajst-905	203	3	on	on	ADP
ajst-905	203	4	machine	machine	NOUN
ajst-905	203	5	learning	learning	NOUN
ajst-905	203	6	.	.	PUNCT
ajst-905	204	1	pmlr	pmlr	NOUN
ajst-905	204	2	,	,	PUNCT
ajst-905	204	3	2015	2015	NUM
ajst-905	204	4	.	.	PUNCT
ajst-905	205	1	[	[	X
ajst-905	205	2	7	7	NUM
ajst-905	205	3	]	]	X
ajst-905	205	4	vaswani	vaswani	NOUN
ajst-905	205	5	,	,	PUNCT
ajst-905	205	6	ashish	ashish	PROPN
ajst-905	205	7	,	,	PUNCT
ajst-905	205	8	et	et	PROPN
ajst-905	205	9	al	al	PROPN
ajst-905	205	10	.	.	PUNCT
ajst-905	206	1	"	"	PUNCT
ajst-905	206	2	attention	attention	NOUN
ajst-905	206	3	is	be	AUX
ajst-905	206	4	all	all	PRON
ajst-905	206	5	you	you	PRON
ajst-905	206	6	need	need	VERB
ajst-905	206	7	.	.	PUNCT
ajst-905	206	8	"	"	PUNCT
ajst-905	207	1	advances	advance	NOUN
ajst-905	207	2	in	in	ADP
ajst-905	207	3	neural	neural	ADJ
ajst-905	207	4	information	information	NOUN
ajst-905	207	5	processing	processing	NOUN
ajst-905	207	6	systems	system	NOUN
ajst-905	207	7	30	30	NUM
ajst-905	207	8	(	(	PUNCT
ajst-905	207	9	2017	2017	NUM
ajst-905	207	10	)	)	PUNCT
ajst-905	207	11	.	.	PUNCT
ajst-905	208	1	[	[	X
ajst-905	208	2	8	8	NUM
ajst-905	208	3	]	]	X
ajst-905	208	4	ren	ren	PROPN
ajst-905	208	5	,	,	PUNCT
ajst-905	208	6	shaoqing	shaoqing	NOUN
ajst-905	208	7	,	,	PUNCT
ajst-905	208	8	et	et	PROPN
ajst-905	208	9	al	al	PROPN
ajst-905	208	10	.	.	PUNCT
ajst-905	209	1	"	"	PUNCT
ajst-905	209	2	faster	fast	ADJ
ajst-905	209	3	r	r	NOUN
ajst-905	209	4	-	-	PUNCT
ajst-905	209	5	cnn	cnn	NOUN
ajst-905	209	6	:	:	PUNCT
ajst-905	209	7	towards	towards	ADP
ajst-905	209	8	real	real	ADJ
ajst-905	209	9	-	-	PUNCT
ajst-905	209	10	time	time	NOUN
ajst-905	209	11	object	object	NOUN
ajst-905	209	12	detection	detection	NOUN
ajst-905	209	13	with	with	ADP
ajst-905	209	14	region	region	NOUN
ajst-905	209	15	proposal	proposal	NOUN
ajst-905	209	16	networks	network	NOUN
ajst-905	209	17	.	.	PUNCT
ajst-905	209	18	"	"	PUNCT
ajst-905	210	1	advances	advance	NOUN
ajst-905	210	2	in	in	ADP
ajst-905	210	3	neural	neural	ADJ
ajst-905	210	4	information	information	NOUN
ajst-905	210	5	processing	processing	NOUN
ajst-905	210	6	systems	system	NOUN
ajst-905	210	7	28	28	NUM
ajst-905	210	8	(	(	PUNCT
ajst-905	210	9	2015	2015	NUM
ajst-905	210	10	)	)	PUNCT
ajst-905	210	11	.	.	PUNCT
ajst-905	211	1	[	[	X
ajst-905	211	2	9	9	NUM
ajst-905	211	3	]	]	X
ajst-905	211	4	jiang	jiang	PROPN
ajst-905	211	5	,	,	PUNCT
ajst-905	211	6	huaizu	huaizu	VERB
ajst-905	211	7	,	,	PUNCT
ajst-905	211	8	et	et	PROPN
ajst-905	211	9	al	al	PROPN
ajst-905	211	10	.	.	PUNCT
ajst-905	212	1	"	"	PUNCT
ajst-905	212	2	in	in	ADP
ajst-905	212	3	defense	defense	NOUN
ajst-905	212	4	of	of	ADP
ajst-905	212	5	grid	grid	NOUN
ajst-905	212	6	features	feature	NOUN
ajst-905	212	7	for	for	ADP
ajst-905	212	8	visual	visual	ADJ
ajst-905	212	9	question	question	NOUN
ajst-905	212	10	answering	answer	VERB
ajst-905	212	11	.	.	PUNCT
ajst-905	212	12	"	"	PUNCT
ajst-905	213	1	proceedings	proceeding	NOUN
ajst-905	213	2	of	of	ADP
ajst-905	213	3	the	the	DET
ajst-905	213	4	ieee	ieee	NOUN
ajst-905	213	5	/	/	SYM
ajst-905	213	6	cvf	cvf	NOUN
ajst-905	213	7	conference	conference	NOUN
ajst-905	213	8	on	on	ADP
ajst-905	213	9	computer	computer	NOUN
ajst-905	213	10	vision	vision	NOUN
ajst-905	213	11	and	and	CCONJ
ajst-905	213	12	pattern	pattern	NOUN
ajst-905	213	13	recognition	recognition	NOUN
ajst-905	213	14	.	.	PUNCT
ajst-905	214	1	2020	2020	NUM
ajst-905	214	2	.	.	PUNCT
ajst-905	215	1	[	[	X
ajst-905	215	2	10	10	NUM
ajst-905	215	3	]	]	X
ajst-905	215	4	lu	lu	PROPN
ajst-905	215	5	,	,	PUNCT
ajst-905	215	6	jiasen	jiasen	PROPN
ajst-905	215	7	,	,	PUNCT
ajst-905	215	8	et	et	PROPN
ajst-905	215	9	al	al	PROPN
ajst-905	215	10	.	.	PUNCT
ajst-905	215	11	"	"	PUNCT
ajst-905	215	12	knowing	know	VERB
ajst-905	215	13	when	when	SCONJ
ajst-905	215	14	to	to	PART
ajst-905	215	15	look	look	VERB
ajst-905	215	16	:	:	PUNCT
ajst-905	215	17	adaptive	adaptive	ADJ
ajst-905	215	18	attention	attention	NOUN
ajst-905	215	19	via	via	ADP
ajst-905	215	20	a	a	DET
ajst-905	215	21	visual	visual	ADJ
ajst-905	215	22	sentinel	sentinel	NOUN
ajst-905	215	23	for	for	ADP
ajst-905	215	24	image	image	NOUN
ajst-905	215	25	captioning	captioning	NOUN
ajst-905	215	26	.	.	PUNCT
ajst-905	215	27	"	"	PUNCT
ajst-905	216	1	proceedings	proceeding	NOUN
ajst-905	216	2	of	of	ADP
ajst-905	216	3	the	the	DET
ajst-905	216	4	ieee	ieee	NOUN
ajst-905	216	5	conference	conference	NOUN
ajst-905	216	6	on	on	ADP
ajst-905	216	7	computer	computer	NOUN
ajst-905	216	8	vision	vision	NOUN
ajst-905	216	9	and	and	CCONJ
ajst-905	216	10	pattern	pattern	NOUN
ajst-905	216	11	recognition	recognition	NOUN
ajst-905	216	12	.	.	PUNCT
ajst-905	217	1	2017	2017	NUM
ajst-905	217	2	.	.	PUNCT
ajst-905	218	1	[	[	X
ajst-905	218	2	11	11	NUM
ajst-905	218	3	]	]	X
ajst-905	218	4	chen	chen	PROPN
ajst-905	218	5	,	,	PUNCT
ajst-905	218	6	long	long	ADV
ajst-905	218	7	,	,	PUNCT
ajst-905	218	8	et	et	PROPN
ajst-905	218	9	al	al	PROPN
ajst-905	218	10	.	.	PUNCT
ajst-905	219	1	"	"	PUNCT
ajst-905	219	2	sca	sca	PROPN
ajst-905	219	3	-	-	PUNCT
ajst-905	219	4	cnn	cnn	PROPN
ajst-905	219	5	:	:	PUNCT
ajst-905	219	6	spatial	spatial	ADJ
ajst-905	219	7	and	and	CCONJ
ajst-905	219	8	channel	channel	NOUN
ajst-905	219	9	-	-	PUNCT
ajst-905	219	10	wise	wise	ADJ
ajst-905	219	11	attention	attention	NOUN
ajst-905	219	12	in	in	ADP
ajst-905	219	13	convolutional	convolutional	ADJ
ajst-905	219	14	networks	network	NOUN
ajst-905	219	15	for	for	ADP
ajst-905	219	16	image	image	NOUN
ajst-905	219	17	captioning	captioning	NOUN
ajst-905	219	18	.	.	PUNCT
ajst-905	219	19	"	"	PUNCT
ajst-905	220	1	proceedings	proceeding	NOUN
ajst-905	220	2	of	of	ADP
ajst-905	220	3	the	the	DET
ajst-905	220	4	ieee	ieee	NOUN
ajst-905	220	5	conference	conference	NOUN
ajst-905	220	6	on	on	ADP
ajst-905	220	7	computer	computer	NOUN
ajst-905	220	8	vision	vision	NOUN
ajst-905	220	9	and	and	CCONJ
ajst-905	220	10	pattern	pattern	NOUN
ajst-905	220	11	recognition	recognition	NOUN
ajst-905	220	12	.	.	PUNCT
ajst-905	221	1	2017	2017	NUM
ajst-905	221	2	.	.	PUNCT
ajst-905	222	1	[	[	X
ajst-905	222	2	12	12	NUM
ajst-905	222	3	]	]	X
ajst-905	222	4	herdade	herdade	PROPN
ajst-905	222	5	,	,	PUNCT
ajst-905	222	6	simao	simao	PROPN
ajst-905	222	7	,	,	PUNCT
ajst-905	222	8	et	et	PROPN
ajst-905	222	9	al	al	PROPN
ajst-905	222	10	.	.	PUNCT
ajst-905	222	11	"	"	PUNCT
ajst-905	222	12	image	image	NOUN
ajst-905	222	13	captioning	captioning	NOUN
ajst-905	222	14	:	:	PUNCT
ajst-905	222	15	transforming	transform	VERB
ajst-905	222	16	objects	object	NOUN
ajst-905	222	17	into	into	ADP
ajst-905	222	18	words	word	NOUN
ajst-905	222	19	.	.	PUNCT
ajst-905	222	20	"	"	PUNCT
ajst-905	223	1	advances	advance	NOUN
ajst-905	223	2	in	in	ADP
ajst-905	223	3	neural	neural	ADJ
ajst-905	223	4	information	information	NOUN
ajst-905	223	5	processing	processing	NOUN
ajst-905	223	6	systems	system	NOUN
ajst-905	223	7	32	32	NUM
ajst-905	223	8	(	(	PUNCT
ajst-905	223	9	2019	2019	NUM
ajst-905	223	10	)	)	PUNCT
ajst-905	223	11	.	.	PUNCT
ajst-905	224	1	[	[	X
ajst-905	224	2	13	13	NUM
ajst-905	224	3	]	]	SYM
ajst-905	224	4	mao	mao	PROPN
ajst-905	224	5	,	,	PUNCT
ajst-905	224	6	junhua	junhua	PROPN
ajst-905	224	7	,	,	PUNCT
ajst-905	224	8	et	et	PROPN
ajst-905	224	9	al	al	PROPN
ajst-905	224	10	.	.	PUNCT
ajst-905	225	1	"	"	PUNCT
ajst-905	225	2	deep	deep	ADJ
ajst-905	225	3	captioning	captioning	NOUN
ajst-905	225	4	with	with	ADP
ajst-905	225	5	multimodal	multimodal	ADJ
ajst-905	225	6	recurrent	recurrent	ADJ
ajst-905	225	7	neural	neural	ADJ
ajst-905	225	8	networks	network	NOUN
ajst-905	225	9	(	(	PUNCT
ajst-905	225	10	m	m	NOUN
ajst-905	225	11	-	-	PUNCT
ajst-905	225	12	rnn	rnn	PROPN
ajst-905	225	13	)	)	PUNCT
ajst-905	225	14	.	.	PUNCT
ajst-905	225	15	"	"	PUNCT
ajst-905	226	1	arxiv	arxiv	PROPN
ajst-905	226	2	preprint	preprint	VERB
ajst-905	226	3	arxiv:1412.6632	arxiv:1412.6632	NOUN
ajst-905	226	4	(	(	PUNCT
ajst-905	226	5	2014	2014	NUM
ajst-905	226	6	)	)	PUNCT
ajst-905	226	7	.	.	PUNCT
ajst-905	227	1	[	[	X
ajst-905	227	2	14	14	NUM
ajst-905	227	3	]	]	SYM
ajst-905	227	4	cornia	cornia	PROPN
ajst-905	227	5	,	,	PUNCT
ajst-905	227	6	marcella	marcella	PROPN
ajst-905	227	7	,	,	PUNCT
ajst-905	227	8	et	et	PROPN
ajst-905	227	9	al	al	PROPN
ajst-905	227	10	.	.	PUNCT
ajst-905	228	1	"	"	PUNCT
ajst-905	228	2	meshed	mesh	VERB
ajst-905	228	3	-	-	PUNCT
ajst-905	228	4	memory	memory	NOUN
ajst-905	228	5	transformer	transformer	NOUN
ajst-905	228	6	for	for	ADP
ajst-905	228	7	image	image	NOUN
ajst-905	228	8	captioning	captioning	NOUN
ajst-905	228	9	.	.	PUNCT
ajst-905	228	10	"	"	PUNCT
ajst-905	229	1	proceedings	proceeding	NOUN
ajst-905	229	2	of	of	ADP
ajst-905	229	3	the	the	DET
ajst-905	229	4	ieee	ieee	NOUN
ajst-905	229	5	/	/	SYM
ajst-905	229	6	cvf	cvf	NOUN
ajst-905	229	7	conference	conference	NOUN
ajst-905	229	8	on	on	ADP
ajst-905	229	9	computer	computer	NOUN
ajst-905	229	10	vision	vision	NOUN
ajst-905	229	11	and	and	CCONJ
ajst-905	229	12	pattern	pattern	NOUN
ajst-905	229	13	recognition	recognition	NOUN
ajst-905	229	14	.	.	PUNCT
ajst-905	230	1	2020	2020	NUM
ajst-905	230	2	.	.	PUNCT
ajst-905	231	1	[	[	X
ajst-905	231	2	15	15	NUM
ajst-905	231	3	]	]	X
ajst-905	231	4	zhang	zhang	PROPN
ajst-905	231	5	,	,	PUNCT
ajst-905	231	6	xuying	xuying	NOUN
ajst-905	231	7	,	,	PUNCT
ajst-905	231	8	et	et	PROPN
ajst-905	231	9	al	al	PROPN
ajst-905	231	10	.	.	PUNCT
ajst-905	232	1	"	"	PUNCT
ajst-905	232	2	rstnet	rstnet	NOUN
ajst-905	232	3	:	:	PUNCT
ajst-905	232	4	captioning	caption	VERB
ajst-905	232	5	with	with	ADP
ajst-905	232	6	adaptive	adaptive	ADJ
ajst-905	232	7	attention	attention	NOUN
ajst-905	232	8	on	on	ADP
ajst-905	232	9	visual	visual	ADJ
ajst-905	232	10	and	and	CCONJ
ajst-905	232	11	non	non	ADJ
ajst-905	232	12	-	-	ADJ
ajst-905	232	13	visual	visual	ADJ
ajst-905	232	14	words	word	NOUN
ajst-905	232	15	.	.	PUNCT
ajst-905	232	16	"	"	PUNCT
ajst-905	233	1	proceedings	proceeding	NOUN
ajst-905	233	2	of	of	ADP
ajst-905	233	3	the	the	DET
ajst-905	233	4	ieee	ieee	NOUN
ajst-905	233	5	/	/	SYM
ajst-905	233	6	cvf	cvf	NOUN
ajst-905	233	7	conference	conference	NOUN
ajst-905	233	8	on	on	ADP
ajst-905	233	9	computer	computer	NOUN
ajst-905	233	10	vision	vision	NOUN
ajst-905	233	11	and	and	CCONJ
ajst-905	233	12	pattern	pattern	NOUN
ajst-905	233	13	recognition	recognition	NOUN
ajst-905	233	14	.	.	PUNCT
ajst-905	234	1	2021	2021	NUM
ajst-905	234	2	.	.	PUNCT
ajst-905	235	1	[	[	X
ajst-905	235	2	16	16	NUM
ajst-905	235	3	]	]	X
ajst-905	235	4	rennie	rennie	PROPN
ajst-905	235	5	,	,	PUNCT
ajst-905	235	6	steven	steven	PROPN
ajst-905	235	7	j.	j.	PROPN
ajst-905	235	8	,	,	PUNCT
ajst-905	235	9	et	et	PROPN
ajst-905	235	10	al	al	PROPN
ajst-905	235	11	.	.	PUNCT
ajst-905	236	1	"	"	PUNCT
ajst-905	236	2	self	self	NOUN
ajst-905	236	3	-	-	PUNCT
ajst-905	236	4	critical	critical	ADJ
ajst-905	236	5	sequence	sequence	NOUN
ajst-905	236	6	training	training	NOUN
ajst-905	236	7	for	for	ADP
ajst-905	236	8	image	image	NOUN
ajst-905	236	9	captioning	captioning	NOUN
ajst-905	236	10	.	.	PUNCT
ajst-905	236	11	"	"	PUNCT
ajst-905	237	1	proceedings	proceeding	NOUN
ajst-905	237	2	of	of	ADP
ajst-905	237	3	the	the	DET
ajst-905	237	4	ieee	ieee	NOUN
ajst-905	237	5	conference	conference	NOUN
ajst-905	237	6	on	on	ADP
ajst-905	237	7	computer	computer	NOUN
ajst-905	237	8	vision	vision	NOUN
ajst-905	237	9	and	and	CCONJ
ajst-905	237	10	pattern	pattern	NOUN
ajst-905	237	11	recognition	recognition	NOUN
ajst-905	237	12	.	.	PUNCT
ajst-905	238	1	2017	2017	NUM
ajst-905	238	2	.	.	PUNCT
ajst-905	239	1	120	120	NUM
ajst-905	240	1	[	[	SYM
ajst-905	240	2	17	17	NUM
ajst-905	240	3	]	]	X
ajst-905	240	4	jiang	jiang	PROPN
ajst-905	240	5	,	,	PUNCT
ajst-905	240	6	wenhao	wenhao	PROPN
ajst-905	240	7	,	,	PUNCT
ajst-905	240	8	et	et	PROPN
ajst-905	240	9	al	al	PROPN
ajst-905	240	10	.	.	PUNCT
ajst-905	241	1	"	"	PUNCT
ajst-905	241	2	recurrent	recurrent	ADJ
ajst-905	241	3	fusion	fusion	NOUN
ajst-905	241	4	network	network	NOUN
ajst-905	241	5	for	for	ADP
ajst-905	241	6	image	image	NOUN
ajst-905	241	7	captioning	captioning	NOUN
ajst-905	241	8	.	.	PUNCT
ajst-905	241	9	"	"	PUNCT
ajst-905	242	1	proceedings	proceeding	NOUN
ajst-905	242	2	of	of	ADP
ajst-905	242	3	the	the	DET
ajst-905	242	4	european	european	PROPN
ajst-905	242	5	conference	conference	PROPN
ajst-905	242	6	on	on	ADP
ajst-905	242	7	computer	computer	NOUN
ajst-905	242	8	vision	vision	NOUN
ajst-905	242	9	(	(	PUNCT
ajst-905	242	10	eccv	eccv	ADV
ajst-905	242	11	)	)	PUNCT
ajst-905	242	12	.	.	PUNCT
ajst-905	243	1	2018	2018	NUM
ajst-905	243	2	.	.	PUNCT
ajst-905	244	1	[	[	X
ajst-905	244	2	18	18	NUM
ajst-905	244	3	]	]	SYM
ajst-905	244	4	yao	yao	PROPN
ajst-905	244	5	,	,	PUNCT
ajst-905	244	6	ting	ting	PROPN
ajst-905	244	7	,	,	PUNCT
ajst-905	244	8	et	et	PROPN
ajst-905	244	9	al	al	PROPN
ajst-905	244	10	.	.	PUNCT
ajst-905	244	11	"	"	PUNCT
ajst-905	244	12	exploring	explore	VERB
ajst-905	244	13	visual	visual	ADJ
ajst-905	244	14	relationship	relationship	NOUN
ajst-905	244	15	for	for	ADP
ajst-905	244	16	image	image	NOUN
ajst-905	244	17	captioning	captioning	NOUN
ajst-905	244	18	.	.	PUNCT
ajst-905	244	19	"	"	PUNCT
ajst-905	245	1	proceedings	proceeding	NOUN
ajst-905	245	2	of	of	ADP
ajst-905	245	3	the	the	DET
ajst-905	245	4	european	european	PROPN
ajst-905	245	5	conference	conference	PROPN
ajst-905	245	6	on	on	ADP
ajst-905	245	7	computer	computer	NOUN
ajst-905	245	8	vision	vision	NOUN
ajst-905	245	9	(	(	PUNCT
ajst-905	245	10	eccv	eccv	ADV
ajst-905	245	11	)	)	PUNCT
ajst-905	245	12	.	.	PUNCT
ajst-905	246	1	2018	2018	NUM
ajst-905	246	2	.	.	PUNCT
ajst-905	247	1	[	[	X
ajst-905	247	2	19	19	NUM
ajst-905	247	3	]	]	SYM
ajst-905	247	4	yang	yang	PROPN
ajst-905	247	5	,	,	PUNCT
ajst-905	247	6	xu	xu	PROPN
ajst-905	247	7	,	,	PUNCT
ajst-905	247	8	et	et	PROPN
ajst-905	247	9	al	al	PROPN
ajst-905	247	10	.	.	PUNCT
ajst-905	248	1	"	"	PUNCT
ajst-905	248	2	auto	auto	NOUN
ajst-905	248	3	-	-	PUNCT
ajst-905	248	4	encoding	encode	VERB
ajst-905	248	5	scene	scene	NOUN
ajst-905	248	6	graphs	graph	NOUN
ajst-905	248	7	for	for	ADP
ajst-905	248	8	image	image	NOUN
ajst-905	248	9	captioning	captioning	NOUN
ajst-905	248	10	.	.	PUNCT
ajst-905	248	11	"	"	PUNCT
ajst-905	249	1	proceedings	proceeding	NOUN
ajst-905	249	2	of	of	ADP
ajst-905	249	3	the	the	DET
ajst-905	249	4	ieee	ieee	NOUN
ajst-905	249	5	/	/	SYM
ajst-905	249	6	cvf	cvf	NOUN
ajst-905	249	7	conference	conference	NOUN
ajst-905	249	8	on	on	ADP
ajst-905	249	9	computer	computer	NOUN
ajst-905	249	10	vision	vision	NOUN
ajst-905	249	11	and	and	CCONJ
ajst-905	249	12	pattern	pattern	NOUN
ajst-905	249	13	recognition	recognition	NOUN
ajst-905	249	14	.	.	PUNCT
ajst-905	250	1	2019	2019	NUM
ajst-905	250	2	.	.	PUNCT
ajst-905	251	1	[	[	X
ajst-905	251	2	20	20	NUM
ajst-905	251	3	]	]	X
ajst-905	251	4	herdade	herdade	PROPN
ajst-905	251	5	,	,	PUNCT
ajst-905	251	6	simao	simao	PROPN
ajst-905	251	7	,	,	PUNCT
ajst-905	251	8	et	et	PROPN
ajst-905	251	9	al	al	PROPN
ajst-905	251	10	.	.	PUNCT
ajst-905	251	11	"	"	PUNCT
ajst-905	251	12	image	image	NOUN
ajst-905	251	13	captioning	captioning	NOUN
ajst-905	251	14	:	:	PUNCT
ajst-905	251	15	transforming	transform	VERB
ajst-905	251	16	objects	object	NOUN
ajst-905	251	17	into	into	ADP
ajst-905	251	18	words	word	NOUN
ajst-905	251	19	.	.	PUNCT
ajst-905	251	20	"	"	PUNCT
ajst-905	252	1	advances	advance	NOUN
ajst-905	252	2	in	in	ADP
ajst-905	252	3	neural	neural	ADJ
ajst-905	252	4	information	information	NOUN
ajst-905	252	5	processing	processing	NOUN
ajst-905	252	6	systems	system	NOUN
ajst-905	252	7	32	32	NUM
ajst-905	252	8	(	(	PUNCT
ajst-905	252	9	2019	2019	NUM
ajst-905	252	10	)	)	PUNCT
ajst-905	252	11	.	.	PUNCT
ajst-905	253	1	[	[	X
ajst-905	253	2	21	21	NUM
ajst-905	253	3	]	]	X
ajst-905	253	4	huang	huang	PROPN
ajst-905	253	5	,	,	PUNCT
ajst-905	253	6	lun	lun	PROPN
ajst-905	253	7	,	,	PUNCT
ajst-905	253	8	et	et	PROPN
ajst-905	253	9	al	al	PROPN
ajst-905	253	10	.	.	PUNCT
ajst-905	254	1	"	"	PUNCT
ajst-905	254	2	attention	attention	NOUN
ajst-905	254	3	on	on	ADP
ajst-905	254	4	attention	attention	NOUN
ajst-905	254	5	for	for	ADP
ajst-905	254	6	image	image	NOUN
ajst-905	254	7	captioning	captioning	NOUN
ajst-905	254	8	.	.	PUNCT
ajst-905	254	9	"	"	PUNCT
ajst-905	255	1	proceedings	proceeding	NOUN
ajst-905	255	2	of	of	ADP
ajst-905	255	3	the	the	DET
ajst-905	255	4	ieee	ieee	NOUN
ajst-905	255	5	/	/	SYM
ajst-905	255	6	cvf	cvf	NOUN
ajst-905	255	7	international	international	ADJ
ajst-905	255	8	conference	conference	NOUN
ajst-905	255	9	on	on	ADP
ajst-905	255	10	computer	computer	NOUN
ajst-905	255	11	vision	vision	NOUN
ajst-905	255	12	.	.	PUNCT
ajst-905	256	1	2019	2019	NUM
ajst-905	256	2	.	.	PUNCT
ajst-905	257	1	[	[	X
ajst-905	257	2	22	22	NUM
ajst-905	257	3	]	]	SYM
ajst-905	257	4	pan	pan	PROPN
ajst-905	257	5	,	,	PUNCT
ajst-905	257	6	yingwei	yingwei	NOUN
ajst-905	257	7	,	,	PUNCT
ajst-905	257	8	et	et	PROPN
ajst-905	257	9	al	al	PROPN
ajst-905	257	10	.	.	PUNCT
ajst-905	258	1	"	"	PUNCT
ajst-905	258	2	x	x	X
ajst-905	258	3	-	-	ADJ
ajst-905	258	4	linear	linear	ADJ
ajst-905	258	5	attention	attention	NOUN
ajst-905	258	6	networks	network	NOUN
ajst-905	258	7	for	for	ADP
ajst-905	258	8	image	image	NOUN
ajst-905	258	9	captioning	captioning	NOUN
ajst-905	258	10	.	.	PUNCT
ajst-905	258	11	"	"	PUNCT
ajst-905	259	1	proceedings	proceeding	NOUN
ajst-905	259	2	of	of	ADP
ajst-905	259	3	the	the	DET
ajst-905	259	4	ieee	ieee	NOUN
ajst-905	259	5	/	/	SYM
ajst-905	259	6	cvf	cvf	NOUN
ajst-905	259	7	conference	conference	NOUN
ajst-905	259	8	on	on	ADP
ajst-905	259	9	computer	computer	NOUN
ajst-905	259	10	vision	vision	NOUN
ajst-905	259	11	and	and	CCONJ
ajst-905	259	12	pattern	pattern	NOUN
ajst-905	259	13	recognition	recognition	NOUN
ajst-905	259	14	.	.	PUNCT
ajst-905	260	1	2020	2020	NUM
ajst-905	260	2	.	.	PUNCT
ajst-905	261	1	[	[	X
ajst-905	261	2	23	23	NUM
ajst-905	261	3	]	]	PUNCT
ajst-905	261	4	cornia	cornia	PROPN
ajst-905	261	5	,	,	PUNCT
ajst-905	261	6	marcella	marcella	PROPN
ajst-905	261	7	,	,	PUNCT
ajst-905	261	8	et	et	PROPN
ajst-905	261	9	al	al	PROPN
ajst-905	261	10	.	.	PUNCT
ajst-905	262	1	"	"	PUNCT
ajst-905	262	2	meshed	mesh	VERB
ajst-905	262	3	-	-	PUNCT
ajst-905	262	4	memory	memory	NOUN
ajst-905	262	5	transformer	transformer	NOUN
ajst-905	262	6	for	for	ADP
ajst-905	262	7	image	image	NOUN
ajst-905	262	8	captioning	captioning	NOUN
ajst-905	262	9	.	.	PUNCT
ajst-905	262	10	"	"	PUNCT
ajst-905	263	1	proceedings	proceeding	NOUN
ajst-905	263	2	of	of	ADP
ajst-905	263	3	the	the	DET
ajst-905	263	4	ieee	ieee	NOUN
ajst-905	263	5	/	/	SYM
ajst-905	263	6	cvf	cvf	NOUN
ajst-905	263	7	conference	conference	NOUN
ajst-905	263	8	on	on	ADP
ajst-905	263	9	computer	computer	NOUN
ajst-905	263	10	vision	vision	NOUN
ajst-905	263	11	and	and	CCONJ
ajst-905	263	12	pattern	pattern	NOUN
ajst-905	263	13	recognition	recognition	NOUN
ajst-905	263	14	.	.	PUNCT
ajst-905	264	1	2020	2020	NUM
ajst-905	264	2	.	.	PUNCT