Datasets:

jerichosiahaya
/

kg-extractor-triples

arxiv_id stringclasses 9 values	chunk int64 0 59	head stringlengths 2 44	relation stringclasses 11 values	tail stringlengths 3 52
1301.3781	3	neural network based language models	outperforms	N-gram models
1301.3781	6	neural network language model	based on	feedforward neural network
1301.3781	7	word vectors	applied to	NLP applications
1301.3781	8	log-bilinear model	uses	diagonal weight matrices
1301.3781	8	distributed representations of words	outperforms	LSA
1301.3781	10	feedforward neural network language model	has part	projection layer
1301.3781	11	NNLM architecture	has part	hidden layer
1301.3781	12	hierarchical softmax	uses	Huffman binary tree
1301.3781	15	RNN model	has part	hidden layer
1301.3781	16	DistBelief	instance of	large-scale distributed framework
1301.3781	16	mini-batch asynchronous gradient descent	uses	Adagrad
1301.3781	18	N-gram NNLM	trained on	distributed representations of words
1301.3781	19	bag-of-words model	based on	feedforward NNLM
1301.3781	20	CBOW	uses	continuous distributed representation of the context
1301.3781	34	feedforward NNLM	uses	DistBelief parallel training
1301.3781	35	NNLM vectors	outperforms	RNN
1301.3781	38	Skip-gram model	trained on	Google News data
1301.3781	49	word vectors	evaluated on	SemEval-2012 Task 2
1301.3781	50	neural network based word vectors	applied to	sentiment analysis
1301.3781	50	neural network based word vectors	applied to	paraphrase detection
1409.3215	0	Our method	uses	multilayered Long Short-Term Memory
1409.3215	1	LSTM	outperforms	phrase-based SMT system
1409.3215	1	LSTM	evaluated on	English to French translation task
1409.3215	5	Long Short-Term Memory	applied to	general sequence to sequence problems
1409.3215	5	second LSTM	instance of	recurrent neural network language model
1409.3215	7	novel differentiable attention mechanism	proposed by	Graves [10]
1409.3215	7	elegant variant of this idea	applied to	machine translation
1409.3215	8	ensemble of 5 deep LSTMs	evaluated on	English to French translation task
1409.3215	9	small-vocabulary neural network architecture	outperforms	phrase-based SMT system
1409.3215	11	Recurrent Neural Network	based on	feedforward neural networks
1409.3215	13	LSTM	applied to	problems with long range temporal dependencies
1409.3215	15	LSTM formulation	based on	Graves [10]
1409.3215	16	deep LSTMs	outperforms	shallow LSTMs
1409.3215	25	deep LSTMs	outperforms	shallow LSTMs
1409.3215	30	pure neural translation system	outperforms	phrase-based SMT baseline
1409.3215	33	LSTM	evaluated on	long sentences
1409.3215	40	LSTM	outperforms	baseline
1409.3215	41	RNN-Language Model (RNNLM)	applied to	MT task
1409.3215	42	Auli et al. [1]	uses	topic model
1409.3215	43	Kalchbrenner and Blunsom [18]	uses	convolutional neural networks
1409.3215	43	Bahdanau et al. [2]	uses	attention mechanism
1409.3215	44	Hermann et al. [12]	uses	feedforward networks
1409.3215	45	large deep LSTM	outperforms	standard SMT-based system
1409.3215	46	LSTMs	trained on	reversed dataset
1409.3215	47	relatively unoptimized approach	outperforms	SMT system
1502.03167	0	Batch Normalization	proposed by	Sergey Ioffe
1502.03167	1	Batch Normalization	outperforms	original model
1502.03167	2	momentum	instance of	SGD variants
1502.03167	2	Adagrad	instance of	SGD variants
1502.03167	8	Batch Normalization	uses	normalization step
1502.03167	10	ensemble of such networks	outperforms	best known results
1502.03167	22	Batch Normalizing Transform	applied to	activation
1502.03167	39	MNIST	instance of	dataset
1502.03167	40	Batch Normalization	applied to	each hidden layer
1502.03167	42	Inception	trained on	ImageNet
1502.03167	42	Convolutional layers	uses	ReLU
1502.03167	43	Stochastic Gradient Descent	uses	momentum
1502.03167	49	BN-x5	based on	Inception
1502.03167	50	BN-x30	based on	BN-x5
1502.03167	50	BN-Baseline	outperforms	Inception
1502.03167	50	BN-x5	outperforms	Inception
1502.03167	51	Deep Image ensemble	evaluated on	ImageNet Large Scale Visual Recognition Competition
1502.03167	52	our ensemble	uses	Dropout
1502.03167	55	Batch Normalization	uses	mini-batch
1502.03167	59	our method	applied to	Recurrent Neural Networks
1512.03385	0	ImageNet	instance of	dataset
1512.03385	1	COCO	instance of	object detection dataset
1512.03385	2	Deep networks	uses	low/mid/highlevel features
1512.03385	4	SGD	uses	backpropagation
1512.03385	5	added layers	instance of	identity mapping
1512.03385	6	deep residual learning framework	applied to	degradation problem
1512.03385	7	shortcut connections	instance of	identity mapping
1512.03385	7	Caffe	instance of	common libraries
1512.03385	10	VLAD	uses	residual vectors
1512.03385	10	Fisher Vector	instance of	probabilistic version
1512.03385	10	Multigrid method	uses	subproblems
1512.03385	11	hierarchical basis preconditioning	uses	variables that represent residual vectors
1512.03385	12	highway networks	uses	gating functions
1512.03385	13	H(x)	instance of	underlying mapping
1512.03385	17	ReLU	instance of	nonlinearity
1512.03385	19	element-wise addition	applied to	feature maps
1512.03385	20	plain baselines	based on	VGG nets
1512.03385	20	downsampling	uses	convolutional layers
1512.03385	21	our model	outperforms	VGG nets
1512.03385	23	identity shortcuts	applied to	same dimensions
1512.03385	24	projection shortcut	uses	convolutions
1512.03385	24	implementation for ImageNet	based on	practice in [21, 41]
1512.03385	25	batch normalization	applied to	each convolution
1512.03385	25	SGD	uses	mini-batch size of 256
1512.03385	26	ImageNet 2012	instance of	dataset
1512.03385	31	baseline architectures	based on	plain nets
1512.03385	32	34-layer ResNet	outperforms	18-layer ResNet
1512.03385	32	identity mapping	applied to	all shortcuts
1512.03385	33	ResNet-50	uses	projections
1512.03385	39	50-layer ResNet	outperforms	2-layer block
1512.03385	40	152-layer ResNet	outperforms	VGG-16/19 nets
1512.03385	40	50/101/152-layer ResNets	outperforms	34-layer ones
1512.03385	41	CIFAR-10 dataset	has part	50k training images
1512.03385	43	identity shortcuts	applied to	all cases
1512.03385	51	maxout	instance of	regularization

End of preview. Expand in Data Studio

kg-extractor triples (synthetic)

Relation triples labeled by an LLM (Claude) from nine arXiv papers, for training and evaluating small relation-extraction models such as REBEL. Paper text is not included; only triples and their chunk index.

arxiv_id: source paper (e.g. 1706.03762 = Attention Is All You Need)
chunk: index of the ~120-word chunk the triple came from (re-create chunks with the kg-extractor pipeline on the same PDF)
head, tail: entity strings, verbatim from the chunk
relation: one of: instance of, subclass of, part of, has part, uses, based on, outperforms, trained on, evaluated on, proposed by, applied to (older labels were mapped into this set)

Known issues

Labels are single-annotator (an LLM), not human gold. Expect noise.
Older mappings are judgment calls; at least one mapping is wrong (replaces → outperforms, e.g. "50-layer ResNet" / "2-layer block").
Sparse: only clearly stated facts were labeled.

Downloads last month: 29