HHousen/ParaSCI · Datasets at Hugging Face

sentence1 stringlengths 16 446	sentence2 stringlengths 14 436
to discover all taxonomic relations , i . e . if a pair of terms is not in the training set , it may become a negative example in the learning process , and will be classified as a non-taxonomic relation .	moreover , if a pair of terms is not contained in the training set , there is high possibility that it will become a negative example in the learning process , and will likely be recognized as a non-taxonomic relation .
each turn is represented as a 300-dimensional vector using the pretrained word2vec embedding model that is trained on google news .	each token is represented by its embedding obtained from a pretrained word embedding model trained on part of google news dataset .
for the neural models , we use 100-dimensional glove embeddings , pre-trained on wikipedia and gigaword .	for the classification task , we use pre-trained glove embedding vectors as lexical features .
the feature weights for the log-linear combination of the features are tuned using minimum error rate training on the devset in terms of bleu .	the optimisation of the feature weights of the model is done with minimum error rate training against the bleu evaluation metric .
during the last decade , statistical machine translation systems have evolved from the original word-based approach into phrase-based translation systems .	in the last decade , statistical machine translation has been advanced by expanding the basic unit of translation from word to phrase and grammar .
we also use mini-batch adagrad for optimization and apply dropout .	we use adagrad for deciding the feature update step .
in this paper we have described affecthor , the system which we submitted to the semeval-2018 affects in tweets .	in this paper we describe our submission to semeval-2018 task 1 : affects in tweets .
goldwater et al showed that modeling dependencies between adjacent words dramatically improves word segmentation accuracy .	goldwater et al used hierarchical dirichlet processes to induce contextual word models .
we also report the results using bleu and ter metrics .	we evaluate the translation quality using the case-sensitive bleu-4 metric .
we selected conditional random fields as the baseline model .	to this end , we use conditional random fields .
we tune weights by minimizing bleu loss on the dev set through mert and report bleu scores on the test set .	we tune model weights using minimum error rate training on the wmt 2008 test data .
it is implemented with another rnn with lstm cells with an attention mechanism and a softmax layer .	this system is a basic encoderdecoder with an attention mechanism .
we use a set of 318 english function words from the scikit-learn package .	we use the linearsvc classifier as implemented in scikit-learn package 17 with the default parameters .
experimental results indicate that this combination outperforms the unsupervised boosting method .	experimental results also show that all boosting methods outperform their corresponding methods without boosting .
pitler and nenkova use the entity grid to capture the coherence of a text for readability assessment .	the entity grid is applied to readability assessment by pitler and nenkova .
we used the pharaoh decoder for both the minimum error rate training and test dataset decoding .	we used minimum error rate training to tune the feature weights for maximum bleu on the development set .
we use skip-gram with negative sampling for obtaining the word embeddings .	we derive 100-dimensional word vectors using word2vec skip-gram model trained over the domain corpus .
we use stanford corenlp for feature generation .	we use stanford corenlp for pos tagging and lemmatization .
sentiment analysis in twitter is a particularly challenging task , because of the informal and “ creative ” writing style , with improper use of grammar , figurative language , misspellings and slang .	as sentiment analysis in twitter is a very recent subject , it is certain that more research and improvements are needed .
the exponential log-linear model weights of both the smt and re-scoring stages of our system were set by tuning the system on development data using the mert procedure by means of the publicly available zmert toolkit 1 .	the exponential log-linear model weights of our system are set by tuning the system on development data using the mert procedure by means of the publicly available zmert toolkit 1 .
we present a novel framework for word alignment that incorporates synonym knowledge collected from monolingual linguistic resources .	we proposed a novel framework that incorporates synonyms from monolingual linguistic resources in a word alignment generative model .
relation extraction is the task of automatically detecting occurrences of expressed relations between entities in a text and structuring the detected information in a tabularized form .	relation extraction is the task of finding relationships between two entities from text .
for decoding , we used moses with the default options .	we used the moses toolkit with its default settings .
we used the stanford parser to extract dependency features for each quote and response .	for ptb pos tags , we tagged the text with the stanford parser .
fomicheva and specia investigate bias in monolingual evaluation of mt and conclude reference bias to be a serious issue , with human annotators strongly biased by the reference translation provided .	following this intuition , fomicheva and specia carry out an investigation into bias in monolingual evaluation of mt and conclude that in a monolingual setting , human assessors of mt are strongly biased by the reference translation .
since our method is applicable to various monotone submodular objective functions and can find almost optimal solutions .	thus we obtain a fast greedy method for compressive summarization , which works with various monotone submodular objective functions and enjoys an approximation guarantee .
in this task , we use the 300-dimensional 840b glove word embeddings .	we use the glove vectors of 300 dimension to represent the input words .
in the dataset is accompanied by a single judgement .	each caption in the dataset is accompanied by a single judgement .
verbnet is a very large lexicon of verbs in english that extends levin with explicitly stated syntactic and semantic information .	verbnet is a verb lexicon with syntactic and semantic information for english verbs , referring to levin verb classes to construct the lexical entries .
for a fair comparison to our model , we used word2vec , that pretrain word embeddings at a token level .	for our experiments reported here , we obtained word vectors using the word2vec tool and the text8 corpus .
a 4-gram language model is trained on the monolingual data by srilm toolkit .	we use a fourgram language model with modified kneser-ney smoothing as implemented in the srilm toolkit .
the idea of distant supervision has widely used in the task of relation extraction .	distant supervision has been successfully used for the problem of relation extraction .
we consider natural language generation from the abstract meaning representation .	the latter category is exemplified by abstract meaning representation .
we estimated 5-gram language models using the sri toolkit with modified kneser-ney smoothing .	we also use a 4-gram language model trained using srilm with kneser-ney smoothing .
burstein et al employ this idea for evaluating coherence in student essays .	burstein et al , used it for an educational purpose , and used it to predict the readability of essays .
mccarthy et al provided a partial solution by describing a method to predict the predominant sense , or the most frequent sense , of a word in a corpus .	mccarthy et al propose a method for automatically identifying the predominant sense in a given domain .
coreference resolution is the next step on the way towards discourse understanding .	coreference resolution is the problem of partitioning a sequence of noun phrases ( or mentions ) , as they occur in a natural language text , into a set of referential entities .
previous applications of recursive neural networks to supervised relation extraction are based on constituency-based parsers .	recursive neural network and convolutional neural network have proven powerful in relation classification .
the task we study is part of the news evaluation campaign conducted in 2009 .	this task is part of the news evaluation campaign conducted in 2009 .
these models were implemented using the package scikit-learn .	the two baseline methods were implemented using scikit-learn in python .
in this paper , we demonstrate that significant gains can instead be achieved by using a more constrained , linguistically motivated grammar .	here , we extend the first approach , and show that better lexical generalization provides significant performance gains .
that achieves this by learning a situated model of meaning from an unlabeled video corpus .	this paired corpus is used to train a situated model of meaning that significantly improves video retrieval performance .
collobert et al propose a multi-task learning framework with dnn for various nlp tasks , including part-of-speech tagging , chunking , named entity recognition , and semantic role labelling .	for example , collobert et al effectively used a multilayer neural network for chunking , part-ofspeech tagging , ner and semantic role labelling .
mcdonald and pereira presented a graph-based parser that can generate graphs in which a word may depend on multiple heads , and evaluated it on the danish treebank .	mcdonald and pereira presented a graph-based parser that can generate dependency graphs in which a word may depend on multiple heads .
we use the stanford rule-based system for coreference resolution .	we included the stanford coreference resolution system in our model for this reason .
conditional random fields are popular models for many nlp tasks .	conditional random field is one of the most effective approaches used in ner tasks .
conditional random fields are undirected graphical models trained to maximize a conditional probability .	crfs are undirected graphical models trained to maximize a conditional probability .
coherence is the property of a good human-authored text that makes it easier to read and understand than a randomly-ordered collection of sentences .	coherence is a central aspect in natural language processing of multi-sentence texts .
to train our model we use markov chain monte carlo sampling .	we use gibbs sampling , a markov chain monte carlo method , to sample from the posterior .
the srilm toolkit was used for training the language models using kneser-ney smoothing .	the language models are 4-grams with modified kneser-ney smoothing which have been trained with the srilm toolkit .
sarcasm is a pervasive phenomenon in social media , permitting the concise communication of meaning , affect and attitude .	sarcasm is defined as ‘ the use of irony to mock or convey contempt ’ 1 .
we used the sri language modeling toolkit to train a fivegram model with modified kneser-ney smoothing .	we used srilm -sri language modeling toolkit to train several character models .
that , we believe , brings us a step further in understanding the benefits of multi-task learning .	this , we believe , is an important step toward understanding how mtl works .
we train a 4-gram language model on the xinhua portion of the gigaword corpus using the sri language toolkit with modified kneser-ney smoothing .	for both languages , we used the srilm toolkit to train a 5-gram language model using all monolingual data provided .
a task that is similar to ours is the task of keywords to question generation that has been addressed recently in zheng et al .	a task that is similar to ours is the task of keywords-to-question generation that has been addressed recently in zheng et al .
we estimated 5-gram language models using the sri toolkit with modified kneser-ney smoothing .	we used 5-gram models , estimated using the sri language modeling toolkit with modified kneser-ney smoothing .
we use an nmt-small model from the opennmt framework for the neural translation .	specifically , we employ the seq2seq model with attention implemented in opennmt .
we use approximate randomization for significance testing .	for assessing significance , we apply the approximate randomization test .
we use glove vectors with 100 dimensions trained on wikipedia and gigaword as word embeddings .	with regard to inputs , we use 50-d glove word embeddings pretrianed on wikipedia and gigaword and 5-d postion embedding .
dimensionality reduction makes the global distributional pattern of a word available in a profile .	a dimensionality reduction creates a space representing the syntactic categories of unambiguous words .
semantic role labeling ( srl ) is the process of assigning semantic roles to strings of words in a sentence according to their relationship to the semantic predicates expressed in the sentence .	semantic role labeling ( srl ) is the task of identifying semantic arguments of predicates in text .
word sense disambiguation ( wsd ) is formally defined as the task of computationally identifying senses of a word in a context .	word sense disambiguation ( wsd ) is the task of assigning sense tags to ambiguous lexical items ( lis ) in a text .
we use srilm toolkit to train a trigram language model with modified kneser-ney smoothing on the target side of training corpus .	a 4-gram language model is trained on the xinhua portion of the gigaword corpus with the srilm toolkit .
in natural language , a word often assumes different meanings , and the task of determining the correct meaning , or sense , of a word in different contexts is known as word sense disambiguation ( wsd ) .	word sense disambiguation ( wsd ) is the task of identifying the correct meaning of a word in context .
the framenet database provides an inventory of semantic frames together with a list of lexical units associated with these frames .	framenet is a semantic resource which provides over 1200 semantic frames that comprise words with similar semantic behaviour .
we obtained both phrase structures and dependency relations for every sentence using the stanford parser .	we compute the syntactic features only for pairs of event mentions from the same sentence , using the stanford dependency parser .
we employ the sentiment analyzer in stanford corenlp to do so .	we use stanford corenlp for pos tagging and lemmatization .
importantly , word embeddings have been effectively used for several nlp tasks .	word embeddings are critical for high-performance neural networks in nlp tasks .
we have used opennmt and marian nmt toolkit to train and test the nmt system .	we used marian toolkit 13 to build competitive nmt systems based on the transformer architecture .
elaborate qualitative and quantitative experimental analyses show the effectiveness of our models .	the qualitative and quantitative experimental analyses demonstrate the efficacy of our models .
therefore , we employ negative sampling and adam to optimize the overall objective function .	we use a binary cross-entropy loss function , and the adam optimizer .
in this paper , we present a novel self-training strategy , which uses information retrieval ( ir ) to collect a cluster of related documents .	in this paper , we apply a novel self-training process on an existing state-of-the-art baseline system .
lin and pantel use a standard monolingual corpus to generate paraphrases , based on dependancy graphs and distributional similarity .	the new approach follows the methodology of lin and pantel for dynamically determining paraphrases in a corpus by measuring the similarity of paths between nodes in syntactic de-pendency trees .
therefore , dependency parsing is a potential “ sweet spot ” that deserves investigation .	dependency parsing is a way of structurally analyzing a sentence from the viewpoint of modification .
in this paper , we present a reinforcement learning framework for inducing mappings from text to actions .	in this paper , we present a reinforcement learning approach for mapping natural language instructions to sequences of executable actions .
we use the 100-dimensional pre-trained word embeddings trained by word2vec 2 and the 100-dimensional randomly initialized pos tag embeddings .	with english gigaword corpus , we use the skip-gram model as implemented in word2vec 3 to induce embeddings .
we used a phrase-based smt model as implemented in the moses toolkit .	in all submitted systems , we use the phrase-based moses decoder .
t盲ckstr枚m et al explore the use of mixed type and token annotations in which a tagger is learned by projecting information via parallel text .	a different approach to cross-lingual pos tagging is proposed by t盲ckstr枚m et al who couple token and type constraints in order to guide learning .
we use srilm toolkits to train two 4-gram language models on the filtered english blog authorship corpus and the xinhua portion of gigaword corpus , respectively .	we use sri language model toolkit to train a 5-gram model with modified kneser-ney smoothing on the target-side training corpus .
math-w-15-1-1-45 itself is efficient in the length of the string .	a language math-w-3-1-2-135 is some subset of math-w-3-1-2-140 .
shallow semantic representations , bearing a more compact information , could prevent the sparseness of deep structural approaches .	shallow semantic representations can prevent the weakness of cosine similarity based models .
kim and hovy proposed to map the semantic frames of framenet into opinion holder and target for only adjectives and verbs .	kim and hovy map the semantic frames of framenet into opinion holder and target for adjectives and verbs to identify these components .
in this paper , we propose a bigram based supervised method for extractive document summarization .	in this paper , we leverage the ilp method as a core component in our summarization system .
we use srilm toolkit to train a trigram language model with modified kneser-ney smoothing on the target side of training corpus .	we trained a 4-gram language model on this data with kneser-ney discounting using srilm .
noraset et al showed that rnn language models can be used to learn to generate definitions for common english words .	noraset et al propose the task of generating definitions based on word embeddings for interpretability purposes .
the algorithm is specified by means of deduction rules , following , and can be implemented using standard tabular techniques .	the algorithm is formulated using the framework of parsing as deduction , extended with weights .
klebanov et al used concreteness as a feature with baseline features and optimal weighting technique .	klebanov et al approach was based on optimal weighting to obtain optimal f-score which lead to comparatively higher recall .
feature weights are tuned using minimum error rate training on the 455 provided references .	the weights of the different feature functions were optimised by means of minimum error rate training .
to the active dual supervision setting , we use the reconstruction error to evaluate the value of feature and example labels .	then by making use of the reconstruction error criterion in matrix factorization , we propose a unified scheme to evaluate the value of feature and example labels .
in this paper , we propose a new method for deriving a kernel from a probabilistic model which is specifically tailored to reranking tasks .	in this paper we propose a method for defining kernels in terms of a probabilistic model of parsing .
as for chinese discourse parser , we build a pipeline system following the annotation procedure of chinese discourse treebank .	and we build a chinese discourse parser following the annotation procedure of chinese discourse treebank .
semantic parsing is the task of converting natural language utterances into formal representations of their meaning .	semantic parsing is the task of mapping a natural language ( nl ) sentence into a completely formal meaning representation ( mr ) or logical form .
sentiment analysis is the task in natural language processing ( nlp ) that deals with classifying opinions according to the polarity of the sentiment they express .	the sentiment analysis is a field of study that investigates feelings present in texts .
information extraction ( ie ) is the task of generating structured information , often in the form of subject-predicate-object relation triples , from unstructured information such as natural language text .	information extraction ( ie ) is the process of identifying events or actions of interest and their participating entities from a text .
table 1 shows the performance for the test data measured by case sensitive bleu .	table 1 summarizes test set performance in bleu , nist and ter .
topkara , topkara , and atallah and topkara et al used machine translation evaluation metrics bleu and nist , automatically measuring how close a stego sentence is to the original .	topkara et al and topkara et al used machine translation evaluation metrics bleu and nist , automatically measuring how close a stego sentence is to the original .
in order to get better translation results , we generate n-best hypotheses with an ensemble model and then train a re-ranker using k-best mira on the validation set .	therefore , we generate 50-best hypothesis from the ensemble system and then tune the model weights with batch-mira on the development set to maximize the bleu score .
the idea of distinguishing between general and domain-specific examples is due to daum茅 and marcu , who used a maximum-entropy model with latent variables to capture the degree of specificity .	daum茅 iii and marcu use an empirical bayes model to estimate a latent variable model grouping instances into domain-specific or common across both domains .
we used srilm to build a 4-gram language model with kneser-ney discounting .	we use a fourgram language model with modified kneser-ney smoothing as implemented in the srilm toolkit .
we have presented multirc , a reading comprehension dataset in which questions require reasoning over multiple sentences to be answered .	in this work , we propose a multi-sentence qa challenge in which questions can be answered only using information from multiple sentences .

to discover all taxonomic relations , i . e . if a pair of terms is not in the training set , it may become a negative example in the learning process , and will be classified as a non-taxonomic relation .

moreover , if a pair of terms is not contained in the training set , there is high possibility that it will become a negative example in the learning process , and will likely be recognized as a non-taxonomic relation .

each turn is represented as a 300-dimensional vector using the pretrained word2vec embedding model that is trained on google news .

each token is represented by its embedding obtained from a pretrained word embedding model trained on part of google news dataset .

for the neural models , we use 100-dimensional glove embeddings , pre-trained on wikipedia and gigaword .

for the classification task , we use pre-trained glove embedding vectors as lexical features .

the feature weights for the log-linear combination of the features are tuned using minimum error rate training on the devset in terms of bleu .

the optimisation of the feature weights of the model is done with minimum error rate training against the bleu evaluation metric .

during the last decade , statistical machine translation systems have evolved from the original word-based approach into phrase-based translation systems .

in the last decade , statistical machine translation has been advanced by expanding the basic unit of translation from word to phrase and grammar .

we also use mini-batch adagrad for optimization and apply dropout .

we use adagrad for deciding the feature update step .

in this paper we have described affecthor , the system which we submitted to the semeval-2018 affects in tweets .

in this paper we describe our submission to semeval-2018 task 1 : affects in tweets .

goldwater et al showed that modeling dependencies between adjacent words dramatically improves word segmentation accuracy .

goldwater et al used hierarchical dirichlet processes to induce contextual word models .

we also report the results using bleu and ter metrics .

we evaluate the translation quality using the case-sensitive bleu-4 metric .

we selected conditional random fields as the baseline model .

to this end , we use conditional random fields .

we tune weights by minimizing bleu loss on the dev set through mert and report bleu scores on the test set .

we tune model weights using minimum error rate training on the wmt 2008 test data .

it is implemented with another rnn with lstm cells with an attention mechanism and a softmax layer .

this system is a basic encoderdecoder with an attention mechanism .

we use a set of 318 english function words from the scikit-learn package .

we use the linearsvc classifier as implemented in scikit-learn package 17 with the default parameters .

experimental results indicate that this combination outperforms the unsupervised boosting method .

experimental results also show that all boosting methods outperform their corresponding methods without boosting .

pitler and nenkova use the entity grid to capture the coherence of a text for readability assessment .

the entity grid is applied to readability assessment by pitler and nenkova .

we used the pharaoh decoder for both the minimum error rate training and test dataset decoding .

we used minimum error rate training to tune the feature weights for maximum bleu on the development set .

we use skip-gram with negative sampling for obtaining the word embeddings .

we derive 100-dimensional word vectors using word2vec skip-gram model trained over the domain corpus .

we use stanford corenlp for feature generation .

we use stanford corenlp for pos tagging and lemmatization .

sentiment analysis in twitter is a particularly challenging task , because of the informal and “ creative ” writing style , with improper use of grammar , figurative language , misspellings and slang .

as sentiment analysis in twitter is a very recent subject , it is certain that more research and improvements are needed .

the exponential log-linear model weights of both the smt and re-scoring stages of our system were set by tuning the system on development data using the mert procedure by means of the publicly available zmert toolkit 1 .

the exponential log-linear model weights of our system are set by tuning the system on development data using the mert procedure by means of the publicly available zmert toolkit 1 .

we present a novel framework for word alignment that incorporates synonym knowledge collected from monolingual linguistic resources .

we proposed a novel framework that incorporates synonyms from monolingual linguistic resources in a word alignment generative model .

relation extraction is the task of automatically detecting occurrences of expressed relations between entities in a text and structuring the detected information in a tabularized form .

relation extraction is the task of finding relationships between two entities from text .

for decoding , we used moses with the default options .

we used the moses toolkit with its default settings .

we used the stanford parser to extract dependency features for each quote and response .

for ptb pos tags , we tagged the text with the stanford parser .

fomicheva and specia investigate bias in monolingual evaluation of mt and conclude reference bias to be a serious issue , with human annotators strongly biased by the reference translation provided .

following this intuition , fomicheva and specia carry out an investigation into bias in monolingual evaluation of mt and conclude that in a monolingual setting , human assessors of mt are strongly biased by the reference translation .

since our method is applicable to various monotone submodular objective functions and can find almost optimal solutions .

thus we obtain a fast greedy method for compressive summarization , which works with various monotone submodular objective functions and enjoys an approximation guarantee .

in this task , we use the 300-dimensional 840b glove word embeddings .

we use the glove vectors of 300 dimension to represent the input words .

in the dataset is accompanied by a single judgement .

each caption in the dataset is accompanied by a single judgement .

verbnet is a very large lexicon of verbs in english that extends levin with explicitly stated syntactic and semantic information .

verbnet is a verb lexicon with syntactic and semantic information for english verbs , referring to levin verb classes to construct the lexical entries .

for a fair comparison to our model , we used word2vec , that pretrain word embeddings at a token level .

for our experiments reported here , we obtained word vectors using the word2vec tool and the text8 corpus .

a 4-gram language model is trained on the monolingual data by srilm toolkit .

we use a fourgram language model with modified kneser-ney smoothing as implemented in the srilm toolkit .

the idea of distant supervision has widely used in the task of relation extraction .

distant supervision has been successfully used for the problem of relation extraction .

we consider natural language generation from the abstract meaning representation .

the latter category is exemplified by abstract meaning representation .

we estimated 5-gram language models using the sri toolkit with modified kneser-ney smoothing .

we also use a 4-gram language model trained using srilm with kneser-ney smoothing .

burstein et al employ this idea for evaluating coherence in student essays .

burstein et al , used it for an educational purpose , and used it to predict the readability of essays .

mccarthy et al provided a partial solution by describing a method to predict the predominant sense , or the most frequent sense , of a word in a corpus .

mccarthy et al propose a method for automatically identifying the predominant sense in a given domain .

coreference resolution is the next step on the way towards discourse understanding .

coreference resolution is the problem of partitioning a sequence of noun phrases ( or mentions ) , as they occur in a natural language text , into a set of referential entities .

previous applications of recursive neural networks to supervised relation extraction are based on constituency-based parsers .

recursive neural network and convolutional neural network have proven powerful in relation classification .

the task we study is part of the news evaluation campaign conducted in 2009 .

this task is part of the news evaluation campaign conducted in 2009 .

these models were implemented using the package scikit-learn .

the two baseline methods were implemented using scikit-learn in python .

in this paper , we demonstrate that significant gains can instead be achieved by using a more constrained , linguistically motivated grammar .

here , we extend the first approach , and show that better lexical generalization provides significant performance gains .

that achieves this by learning a situated model of meaning from an unlabeled video corpus .

this paired corpus is used to train a situated model of meaning that significantly improves video retrieval performance .

collobert et al propose a multi-task learning framework with dnn for various nlp tasks , including part-of-speech tagging , chunking , named entity recognition , and semantic role labelling .

for example , collobert et al effectively used a multilayer neural network for chunking , part-ofspeech tagging , ner and semantic role labelling .

mcdonald and pereira presented a graph-based parser that can generate graphs in which a word may depend on multiple heads , and evaluated it on the danish treebank .

mcdonald and pereira presented a graph-based parser that can generate dependency graphs in which a word may depend on multiple heads .

we use the stanford rule-based system for coreference resolution .

we included the stanford coreference resolution system in our model for this reason .

conditional random fields are popular models for many nlp tasks .

conditional random field is one of the most effective approaches used in ner tasks .

conditional random fields are undirected graphical models trained to maximize a conditional probability .

crfs are undirected graphical models trained to maximize a conditional probability .

coherence is the property of a good human-authored text that makes it easier to read and understand than a randomly-ordered collection of sentences .

coherence is a central aspect in natural language processing of multi-sentence texts .

to train our model we use markov chain monte carlo sampling .

we use gibbs sampling , a markov chain monte carlo method , to sample from the posterior .

the srilm toolkit was used for training the language models using kneser-ney smoothing .

the language models are 4-grams with modified kneser-ney smoothing which have been trained with the srilm toolkit .

sarcasm is a pervasive phenomenon in social media , permitting the concise communication of meaning , affect and attitude .

sarcasm is defined as ‘ the use of irony to mock or convey contempt ’ 1 .

we used the sri language modeling toolkit to train a fivegram model with modified kneser-ney smoothing .

we used srilm -sri language modeling toolkit to train several character models .

that , we believe , brings us a step further in understanding the benefits of multi-task learning .

this , we believe , is an important step toward understanding how mtl works .

we train a 4-gram language model on the xinhua portion of the gigaword corpus using the sri language toolkit with modified kneser-ney smoothing .

for both languages , we used the srilm toolkit to train a 5-gram language model using all monolingual data provided .

a task that is similar to ours is the task of keywords to question generation that has been addressed recently in zheng et al .

a task that is similar to ours is the task of keywords-to-question generation that has been addressed recently in zheng et al .

we estimated 5-gram language models using the sri toolkit with modified kneser-ney smoothing .

we used 5-gram models , estimated using the sri language modeling toolkit with modified kneser-ney smoothing .

we use an nmt-small model from the opennmt framework for the neural translation .

specifically , we employ the seq2seq model with attention implemented in opennmt .

we use approximate randomization for significance testing .

for assessing significance , we apply the approximate randomization test .

we use glove vectors with 100 dimensions trained on wikipedia and gigaword as word embeddings .

with regard to inputs , we use 50-d glove word embeddings pretrianed on wikipedia and gigaword and 5-d postion embedding .

dimensionality reduction makes the global distributional pattern of a word available in a profile .

a dimensionality reduction creates a space representing the syntactic categories of unambiguous words .

semantic role labeling ( srl ) is the process of assigning semantic roles to strings of words in a sentence according to their relationship to the semantic predicates expressed in the sentence .

semantic role labeling ( srl ) is the task of identifying semantic arguments of predicates in text .

word sense disambiguation ( wsd ) is formally defined as the task of computationally identifying senses of a word in a context .

word sense disambiguation ( wsd ) is the task of assigning sense tags to ambiguous lexical items ( lis ) in a text .

we use srilm toolkit to train a trigram language model with modified kneser-ney smoothing on the target side of training corpus .

a 4-gram language model is trained on the xinhua portion of the gigaword corpus with the srilm toolkit .

in natural language , a word often assumes different meanings , and the task of determining the correct meaning , or sense , of a word in different contexts is known as word sense disambiguation ( wsd ) .

word sense disambiguation ( wsd ) is the task of identifying the correct meaning of a word in context .

the framenet database provides an inventory of semantic frames together with a list of lexical units associated with these frames .

framenet is a semantic resource which provides over 1200 semantic frames that comprise words with similar semantic behaviour .

we obtained both phrase structures and dependency relations for every sentence using the stanford parser .

we compute the syntactic features only for pairs of event mentions from the same sentence , using the stanford dependency parser .

we employ the sentiment analyzer in stanford corenlp to do so .

we use stanford corenlp for pos tagging and lemmatization .

importantly , word embeddings have been effectively used for several nlp tasks .

word embeddings are critical for high-performance neural networks in nlp tasks .

we have used opennmt and marian nmt toolkit to train and test the nmt system .

we used marian toolkit 13 to build competitive nmt systems based on the transformer architecture .

elaborate qualitative and quantitative experimental analyses show the effectiveness of our models .

the qualitative and quantitative experimental analyses demonstrate the efficacy of our models .

therefore , we employ negative sampling and adam to optimize the overall objective function .

we use a binary cross-entropy loss function , and the adam optimizer .

in this paper , we present a novel self-training strategy , which uses information retrieval ( ir ) to collect a cluster of related documents .

in this paper , we apply a novel self-training process on an existing state-of-the-art baseline system .

lin and pantel use a standard monolingual corpus to generate paraphrases , based on dependancy graphs and distributional similarity .

the new approach follows the methodology of lin and pantel for dynamically determining paraphrases in a corpus by measuring the similarity of paths between nodes in syntactic de-pendency trees .

therefore , dependency parsing is a potential “ sweet spot ” that deserves investigation .

dependency parsing is a way of structurally analyzing a sentence from the viewpoint of modification .

in this paper , we present a reinforcement learning framework for inducing mappings from text to actions .

in this paper , we present a reinforcement learning approach for mapping natural language instructions to sequences of executable actions .

we use the 100-dimensional pre-trained word embeddings trained by word2vec 2 and the 100-dimensional randomly initialized pos tag embeddings .

with english gigaword corpus , we use the skip-gram model as implemented in word2vec 3 to induce embeddings .

we used a phrase-based smt model as implemented in the moses toolkit .

in all submitted systems , we use the phrase-based moses decoder .

t盲ckstr枚m et al explore the use of mixed type and token annotations in which a tagger is learned by projecting information via parallel text .

a different approach to cross-lingual pos tagging is proposed by t盲ckstr枚m et al who couple token and type constraints in order to guide learning .

we use srilm toolkits to train two 4-gram language models on the filtered english blog authorship corpus and the xinhua portion of gigaword corpus , respectively .

we use sri language model toolkit to train a 5-gram model with modified kneser-ney smoothing on the target-side training corpus .

math-w-15-1-1-45 itself is efficient in the length of the string .

a language math-w-3-1-2-135 is some subset of math-w-3-1-2-140 .

shallow semantic representations , bearing a more compact information , could prevent the sparseness of deep structural approaches .

shallow semantic representations can prevent the weakness of cosine similarity based models .

kim and hovy proposed to map the semantic frames of framenet into opinion holder and target for only adjectives and verbs .

kim and hovy map the semantic frames of framenet into opinion holder and target for adjectives and verbs to identify these components .

in this paper , we propose a bigram based supervised method for extractive document summarization .

in this paper , we leverage the ilp method as a core component in our summarization system .

we use srilm toolkit to train a trigram language model with modified kneser-ney smoothing on the target side of training corpus .

we trained a 4-gram language model on this data with kneser-ney discounting using srilm .

noraset et al showed that rnn language models can be used to learn to generate definitions for common english words .

noraset et al propose the task of generating definitions based on word embeddings for interpretability purposes .

the algorithm is specified by means of deduction rules , following , and can be implemented using standard tabular techniques .

the algorithm is formulated using the framework of parsing as deduction , extended with weights .

klebanov et al used concreteness as a feature with baseline features and optimal weighting technique .

klebanov et al approach was based on optimal weighting to obtain optimal f-score which lead to comparatively higher recall .

feature weights are tuned using minimum error rate training on the 455 provided references .

the weights of the different feature functions were optimised by means of minimum error rate training .

to the active dual supervision setting , we use the reconstruction error to evaluate the value of feature and example labels .

then by making use of the reconstruction error criterion in matrix factorization , we propose a unified scheme to evaluate the value of feature and example labels .

in this paper , we propose a new method for deriving a kernel from a probabilistic model which is specifically tailored to reranking tasks .

in this paper we propose a method for defining kernels in terms of a probabilistic model of parsing .

as for chinese discourse parser , we build a pipeline system following the annotation procedure of chinese discourse treebank .

and we build a chinese discourse parser following the annotation procedure of chinese discourse treebank .

semantic parsing is the task of converting natural language utterances into formal representations of their meaning .

semantic parsing is the task of mapping a natural language ( nl ) sentence into a completely formal meaning representation ( mr ) or logical form .

sentiment analysis is the task in natural language processing ( nlp ) that deals with classifying opinions according to the polarity of the sentiment they express .

the sentiment analysis is a field of study that investigates feelings present in texts .

information extraction ( ie ) is the task of generating structured information , often in the form of subject-predicate-object relation triples , from unstructured information such as natural language text .

information extraction ( ie ) is the process of identifying events or actions of interest and their participating entities from a text .

table 1 shows the performance for the test data measured by case sensitive bleu .

table 1 summarizes test set performance in bleu , nist and ter .

topkara , topkara , and atallah and topkara et al used machine translation evaluation metrics bleu and nist , automatically measuring how close a stego sentence is to the original .

topkara et al and topkara et al used machine translation evaluation metrics bleu and nist , automatically measuring how close a stego sentence is to the original .

in order to get better translation results , we generate n-best hypotheses with an ensemble model and then train a re-ranker using k-best mira on the validation set .

therefore , we generate 50-best hypothesis from the ensemble system and then tune the model weights with batch-mira on the development set to maximize the bleu score .

the idea of distinguishing between general and domain-specific examples is due to daum茅 and marcu , who used a maximum-entropy model with latent variables to capture the degree of specificity .

daum茅 iii and marcu use an empirical bayes model to estimate a latent variable model grouping instances into domain-specific or common across both domains .

we used srilm to build a 4-gram language model with kneser-ney discounting .

we use a fourgram language model with modified kneser-ney smoothing as implemented in the srilm toolkit .

we have presented multirc , a reading comprehension dataset in which questions require reasoning over multiple sentences to be answered .

in this work , we propose a multi-sentence qa challenge in which questions can be answered only using information from multiple sentences .